Skip to content

GitLab

Commit c2eff804 authored by Sugon_ldc's avatar Sugon_ldc
Browse files

add new project

parents
Pipeline #1648 failed with stages
in 0 seconds
Show whitespace changes
Inline Side-by-side
Showing with 3746 additions and 0 deletions
faster_whisper/feature_extractor.py 0 → 100644
View file @ c2eff804
import torch
# Adapted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/whisper/feature_extraction_whisper.py # noqa: E501
class FeatureExtractor:
def __init__(
self,
device: str = "auto",
feature_size=80,
sampling_rate=16000,
hop_length=160,
chunk_length=30,
n_fft=400,
):
if device == "auto":
self.device = "cuda" if torch.cuda.is_available() else "cpu"
else:
self.device = device
self.n_fft = n_fft
self.hop_length = hop_length
self.chunk_length = chunk_length
self.n_samples = chunk_length * sampling_rate
self.nb_max_frames = self.n_samples // hop_length
self.time_per_frame = hop_length / sampling_rate
self.sampling_rate = sampling_rate
self.mel_filters = self.get_mel_filters(
sampling_rate, n_fft, n_mels=feature_size
)
@staticmethod
def get_mel_filters(sr, n_fft, n_mels=128):
"""
Implementation of librosa.filters.mel in Pytorch
"""
# Initialize the weights
n_mels = int(n_mels)
# Center freqs of each FFT bin
fftfreqs = torch.fft.rfftfreq(n=n_fft, d=1.0 / sr)
# 'Center freqs' of mel bands - uniformly spaced between limits
min_mel = 0.0
max_mel = 45.245640471924965
mels = torch.linspace(min_mel, max_mel, n_mels + 2)
# Fill in the linear scale
f_min = 0.0
f_sp = 200.0 / 3
freqs = f_min + f_sp * mels
# And now the nonlinear scale
min_log_hz = 1000.0 # beginning of log region (Hz)
min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)
logstep = torch.log(torch.tensor(6.4)) / 27.0 # step size for log region
# If we have vector data, vectorize
log_t = mels >= min_log_mel
freqs[log_t] = min_log_hz * torch.exp(logstep * (mels[log_t] - min_log_mel))
mel_f = freqs
fdiff = torch.diff(mel_f)
ramps = mel_f.view(-1, 1) - fftfreqs.view(1, -1)
lower = -ramps[:-2] / fdiff[:-1].unsqueeze(1)
upper = ramps[2:] / fdiff[1:].unsqueeze(1)
# Intersect them with each other and zero, vectorized across all i
weights = torch.maximum(torch.zeros_like(lower), torch.minimum(lower, upper))
# Slaney-style mel is scaled to be approx constant energy per channel
enorm = 2.0 / (mel_f[2 : n_mels + 2] - mel_f[:n_mels])
weights *= enorm.unsqueeze(1)
return weights
def __call__(self, waveform, padding=True, chunk_length=None, to_cpu=False):
"""
Compute the log-Mel spectrogram of the provided audio.
"""
if chunk_length is not None:
self.n_samples = chunk_length * self.sampling_rate
self.nb_max_frames = self.n_samples // self.hop_length
if waveform.dtype is not torch.float32:
waveform = waveform.to(torch.float32)
waveform = (
waveform.to(self.device)
if self.device == "cuda" and not waveform.is_cuda
else waveform
)
if padding:
waveform = torch.nn.functional.pad(waveform, (0, self.n_samples))
window = torch.hann_window(self.n_fft).to(waveform.device)
stft = torch.stft(
waveform, self.n_fft, self.hop_length, window=window, return_complex=True
)
magnitudes = stft[..., :-1].abs() ** 2
mel_spec = self.mel_filters.to(waveform.device) @ magnitudes
log_spec = torch.clamp(mel_spec, min=1e-10).log10()
log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
log_spec = (log_spec + 4.0) / 4.0
# When the model is running on multiple GPUs, the output should be moved
# to the CPU since we don't know which GPU will handle the next job.
return log_spec.cpu() if to_cpu else log_spec
faster_whisper/tokenizer.py 0 → 100644
View file @ c2eff804
import string
from functools import cached_property
from typing import List, Optional, Tuple
import tokenizers
class Tokenizer:
"""Simple wrapper around a tokenizers.Tokenizer."""
def __init__(
self,
tokenizer: tokenizers.Tokenizer,
multilingual: bool,
task: Optional[str] = None,
language: Optional[str] = None,
):
self.tokenizer = tokenizer
if multilingual:
if task not in _TASKS:
raise ValueError(
"'%s' is not a valid task (accepted tasks: %s)"
% (task, ", ".join(_TASKS))
)
if language not in _LANGUAGE_CODES:
raise ValueError(
"'%s' is not a valid language code (accepted language codes: %s)"
% (language, ", ".join(_LANGUAGE_CODES))
)
self.task = self.tokenizer.token_to_id("<|%s|>" % task)
self.language = self.tokenizer.token_to_id("<|%s|>" % language)
self.language_code = language
else:
self.task = None
self.language = None
self.language_code = "en"
@cached_property
def transcribe(self) -> int:
return self.tokenizer.token_to_id("<|transcribe|>")
@cached_property
def translate(self) -> int:
return self.tokenizer.token_to_id("<|translate|>")
@cached_property
def sot(self) -> int:
return self.tokenizer.token_to_id("<|startoftranscript|>")
@cached_property
def sot_lm(self) -> int:
return self.tokenizer.token_to_id("<|startoflm|>")
@cached_property
def sot_prev(self) -> int:
return self.tokenizer.token_to_id("<|startofprev|>")
@cached_property
def eot(self) -> int:
return self.tokenizer.token_to_id("<|endoftext|>")
@cached_property
def no_timestamps(self) -> int:
return self.tokenizer.token_to_id("<|notimestamps|>")
@property
def timestamp_begin(self) -> int:
return self.no_timestamps + 1
@property
def sot_sequence(self) -> List[int]:
sequence = [self.sot]
if self.language is not None:
sequence.append(self.language)
if self.task is not None:
sequence.append(self.task)
return sequence
def encode(self, text: str) -> List[int]:
return self.tokenizer.encode(text, add_special_tokens=False).ids
def decode(self, tokens: List[int]) -> str:
text_tokens = [token for token in tokens if token < self.eot]
return self.tokenizer.decode(text_tokens)
def decode_with_timestamps(self, tokens: List[int]) -> str:
outputs = [[]]
for token in tokens:
if token >= self.timestamp_begin:
timestamp = f"<|{(token - self.timestamp_begin) * 0.02:.2f}|>"
outputs.append(timestamp)
outputs.append([])
else:
outputs[-1].append(token)
return "".join(
[s if isinstance(s, str) else self.tokenizer.decode(s) for s in outputs]
)
@cached_property
def non_speech_tokens(self) -> Tuple[int]:
"""
Returns the list of tokens to suppress in order to avoid any speaker tags or non-speech
annotations, to prevent sampling texts that are not actually spoken in the audio, e.g.
- ♪♪♪
- ( SPEAKING FOREIGN LANGUAGE )
- [DAVID] Hey there,
keeping basic punctuations like commas, periods, question marks, exclamation points, etc.
"""
symbols = list('"#()*+/:;<=>@[\\]^_`{|}~「」『』')
symbols += (
"<< >> <<< >>> -- --- -( -[ (' (\" (( )) ((( ))) [[ ]] {{ }} ♪♪ ♪♪♪".split()
)
# symbols that may be a single token or multiple tokens depending on the tokenizer.
# In case they're multiple tokens, suppress the first token, which is safe because:
# These are between U+2640 and U+267F miscellaneous symbols that are okay to suppress
# in generations, and in the 3-byte UTF-8 representation they share the first two bytes.
miscellaneous = set("♩♪♫♬♭♮♯")
assert all(0x2640 <= ord(c) <= 0x267F for c in miscellaneous)
# allow hyphens "-" and single quotes "'" between words, but not at the beginning of a word
result = {self.encode(" -")[0], self.encode(" '")[0]}
for symbol in symbols + list(miscellaneous):
for tokens in [
self.encode(symbol),
self.encode(" " + symbol),
]:
if len(tokens) == 1 or symbol in miscellaneous:
result.add(tokens[0])
return tuple(sorted(result))
def split_to_word_tokens(
self, tokens: List[int]
) -> Tuple[List[str], List[List[int]]]:
if self.language_code in {"zh", "ja", "th", "lo", "my", "yue"}:
# These languages don't typically use spaces, so it is difficult to split words
# without morpheme analysis. Here, we instead split words at any
# position where the tokens are decoded as valid unicode points
return self.split_tokens_on_unicode(tokens)
return self.split_tokens_on_spaces(tokens)
def split_tokens_on_unicode(
self, tokens: List[int]
) -> Tuple[List[str], List[List[int]]]:
decoded_full = self.decode_with_timestamps(tokens)
replacement_char = "\ufffd"
words = []
word_tokens = []
current_tokens = []
unicode_offset = 0
for token in tokens:
current_tokens.append(token)
decoded = self.decode_with_timestamps(current_tokens)
try:
replacement_char_index = decoded.index(replacement_char)
replacement_char_index += unicode_offset
except ValueError:
replacement_char_index = None
if replacement_char_index is None or (
replacement_char_index < len(decoded_full)
and decoded_full[replacement_char_index] == replacement_char
):
words.append(decoded)
word_tokens.append(current_tokens)
current_tokens = []
unicode_offset += len(decoded)
return words, word_tokens
def split_tokens_on_spaces(
self, tokens: List[int]
) -> Tuple[List[str], List[List[int]]]:
subwords, subword_tokens_list = self.split_tokens_on_unicode(tokens)
words = []
word_tokens = []
for subword, subword_tokens in zip(subwords, subword_tokens_list):
special = subword_tokens[0] >= self.eot
with_space = subword.startswith(" ")
punctuation = subword.strip() in string.punctuation
if special or with_space or punctuation or len(words) == 0:
words.append(subword)
word_tokens.append(subword_tokens)
else:
words[-1] = words[-1] + subword
word_tokens[-1].extend(subword_tokens)
return words, word_tokens
_TASKS = (
"transcribe",
"translate",
)
_LANGUAGE_CODES = (
"af",
"am",
"ar",
"as",
"az",
"ba",
"be",
"bg",
"bn",
"bo",
"br",
"bs",
"ca",
"cs",
"cy",
"da",
"de",
"el",
"en",
"es",
"et",
"eu",
"fa",
"fi",
"fo",
"fr",
"gl",
"gu",
"ha",
"haw",
"he",
"hi",
"hr",
"ht",
"hu",
"hy",
"id",
"is",
"it",
"ja",
"jw",
"ka",
"kk",
"km",
"kn",
"ko",
"la",
"lb",
"ln",
"lo",
"lt",
"lv",
"mg",
"mi",
"mk",
"ml",
"mn",
"mr",
"ms",
"mt",
"my",
"ne",
"nl",
"nn",
"no",
"oc",
"pa",
"pl",
"ps",
"pt",
"ro",
"ru",
"sa",
"sd",
"si",
"sk",
"sl",
"sn",
"so",
"sq",
"sr",
"su",
"sv",
"sw",
"ta",
"te",
"tg",
"th",
"tk",
"tl",
"tr",
"tt",
"uk",
"ur",
"uz",
"vi",
"yi",
"yo",
"zh",
"yue",
)
faster_whisper/transcribe.py 0 → 100644
View file @ c2eff804
import itertools
import json
import logging
import os
import random
import zlib
from collections import Counter, defaultdict
from inspect import signature
from typing import BinaryIO, Iterable, List, NamedTuple, Optional, Tuple, Union
import ctranslate2
import numpy as np
import tokenizers
import torch
from pyannote.audio import Model
from tqdm import tqdm
from faster_whisper.audio import decode_audio, pad_or_trim
from faster_whisper.feature_extractor import FeatureExtractor
from faster_whisper.tokenizer import _LANGUAGE_CODES, Tokenizer
from faster_whisper.utils import (
download_model,
format_timestamp,
get_assets_path,
get_end,
get_logger,
)
from faster_whisper.vad import (
SpeechTimestampsMap,
VadOptions,
VoiceActivitySegmentation,
collect_chunks,
get_speech_timestamps,
merge_chunks,
)
class Word(NamedTuple):
start: float
end: float
word: str
probability: float
class Segment(NamedTuple):
id: int
seek: int
start: float
end: float
text: str
tokens: List[int]
avg_logprob: float
compression_ratio: float
no_speech_prob: float
words: Optional[List[Word]]
temperature: Optional[float] = 1.0
# Added additional parameters for multilingual videos and fixes below
class TranscriptionOptions(NamedTuple):
beam_size: int
best_of: int
patience: float
length_penalty: float
repetition_penalty: float
no_repeat_ngram_size: int
log_prob_threshold: Optional[float]
log_prob_low_threshold: Optional[float]
no_speech_threshold: Optional[float]
compression_ratio_threshold: Optional[float]
condition_on_previous_text: bool
prompt_reset_on_temperature: float
temperatures: List[float]
initial_prompt: Optional[Union[str, Iterable[int]]]
prefix: Optional[str]
suppress_blank: bool
suppress_tokens: Optional[List[int]]
without_timestamps: bool
max_initial_timestamp: float
word_timestamps: bool
prepend_punctuations: str
append_punctuations: str
multilingual: bool
output_language: Optional[str]
max_new_tokens: Optional[int]
clip_timestamps: Union[str, List[float]]
hallucination_silence_threshold: Optional[float]
hotwords: Optional[str]
class TranscriptionInfo(NamedTuple):
language: str
language_probability: float
duration: float
duration_after_vad: float
all_language_probs: Optional[List[Tuple[str, float]]]
transcription_options: TranscriptionOptions
vad_options: VadOptions
# The code below is originally from HF pipeline and is used in whisper-x
# (https://github.com/m-bain/whisperX) and adapted for faster_whisper
class BatchedInferencePipeline:
"""
Huggingface Pipeline wrapper for WhisperModel.
Copyright (c) 2022, Max Bain
All rights reserved.
Modified by Mobius Labs GmbH
"""
def __init__(
self,
model,
use_vad_model: bool = True,
options: Optional[NamedTuple] = None,
tokenizer=None,
chunk_length: int = 30,
vad_device: Union[int, str, "torch.device"] = "auto",
vad_onset: float = 0.500,
vad_offset: float = 0.363,
language: Optional[str] = None,
):
self.model: WhisperModel = model
self.tokenizer = tokenizer
self.options = options
self.preset_language = language
self.use_vad_model = use_vad_model
self.vad_onset = vad_onset
self.vad_offset = vad_offset
self.vad_model_path = os.path.join(get_assets_path(), "pyannote_vad_model.bin")
if self.use_vad_model:
self.vad_device = self.get_device(vad_device)
self.vad_model = self.load_vad_model(
vad_onset=self.vad_onset, vad_offset=self.vad_offset
)
else:
self.vad_model = None
self.chunk_length = chunk_length # VAD merging size
self.last_speech_timestamp = 0.0
def get_device(self, device: Union[int, str, "torch.device"]):
"""
Converts the input device into a torch.device object.
The input can be an integer, a string, or a `torch.device` object.
The function handles a special case where the input device is "auto".
When "auto" is specified, the device will default to the
device of the model (self.model.device). If the model's device is also "auto",
it selects "cuda" if a CUDA-capable device is available; otherwise, it selects "cpu".
"""
if isinstance(device, torch.device):
return device
elif isinstance(device, str):
if device == "auto" and self.model.device == "auto":
device = "cuda" if torch.cuda.is_available() else "cpu"
elif device == "auto":
device = self.model.device
return torch.device(device)
elif device < 0:
return torch.device("cpu")
else:
return torch.device(f"cuda:{device}")
def forward(self, features, segments_metadata, **forward_params):
encoder_output, outputs = self.model.generate_segment_batched(
features, self.tokenizer, forward_params
)
segmented_outputs = []
segment_sizes = []
for segment_metadata, output in zip(segments_metadata, outputs):
duration = segment_metadata["end_time"] - segment_metadata["start_time"]
segment_size = int(duration * self.model.frames_per_second)
segment_sizes.append(segment_size)
(
subsegments,
seek,
single_timestamp_ending,
) = self.model._split_segments_by_timestamps(
tokenizer=self.tokenizer,
tokens=output["tokens"],
time_offset=segment_metadata["start_time"],
segment_size=segment_size,
segment_duration=duration,
seek=0,
)
segmented_outputs.append(
[
dict(
text=self.tokenizer.decode(subsegment["tokens"]),
avg_logprob=output["avg_logprob"],
no_speech_prob=output["no_speech_prob"],
tokens=subsegment["tokens"],
start=subsegment["start"],
end=subsegment["end"],
compression_ratio=get_compression_ratio(
self.tokenizer.decode(subsegment["tokens"])
),
)
for subsegment in subsegments
]
)
if forward_params["word_timestamps"]:
self.last_speech_timestamp = self.model.add_word_timestamps(
segmented_outputs,
self.tokenizer,
encoder_output,
segment_sizes,
forward_params["prepend_punctuations"],
forward_params["append_punctuations"],
self.last_speech_timestamp,
)
return segmented_outputs
def get_language_and_tokenizer(
self, audio, task: Optional[str] = None, language: Optional[str] = None
):
all_language_probs = None
language_probability = 1.0
if self.tokenizer is None:
if not language:
(
language,
language_probability,
all_language_probs,
) = self.model.detect_language(audio)
task = task or "transcribe"
self.tokenizer = Tokenizer(
self.model.hf_tokenizer,
self.model.model.is_multilingual,
task=task,
language=language,
)
else:
if task is not None:
self.tokenizer.task = self.tokenizer.tokenizer.token_to_id(
f"<|{task}|>"
)
if language is not None:
self.tokenizer.language = self.tokenizer.tokenizer.token_to_id(
f"<|{language}|>"
)
self.tokenizer.language_code = language
return language, language_probability, task, all_language_probs
@staticmethod
def audio_split(audio, segments, sampling_rate):
"""Returns splitted audio chunks as iterator"""
audio_segments = []
segments_metadata = []
for seg in segments:
f1 = int(seg["start"] * sampling_rate)
f2 = int(seg["end"] * sampling_rate)
seg_metadata = {
"start_time": seg["start"],
"end_time": seg["end"],
"stitched_seg": seg["segments"],
}
audio_segments.append(audio[f1:f2])
segments_metadata.append(seg_metadata)
return audio_segments, segments_metadata
def load_vad_model(self, vad_onset=0.500, vad_offset=0.363):
vad_model = Model.from_pretrained(self.vad_model_path)
hyperparameters = {
"onset": vad_onset,
"offset": vad_offset,
"min_duration_on": 0.1,
"min_duration_off": 0.1,
}
vad_pipeline = VoiceActivitySegmentation(
segmentation=vad_model, device=torch.device(self.vad_device)
)
vad_pipeline.instantiate(hyperparameters)
return vad_pipeline
def transcribe(
self,
audio: Union[str, torch.Tensor, np.ndarray],
vad_segments: Optional[List[dict]] = None,
batch_size: int = 16,
language: Optional[str] = None,
task: str = None,
log_progress: bool = False,
beam_size: int = 5,
best_of: int = 5,
patience: float = 1,
length_penalty: float = 1,
repetition_penalty: float = 1,
no_repeat_ngram_size: int = 0,
temperature: Union[float, List[float], Tuple[float, ...]] = [
0.0,
0.2,
0.4,
0.6,
0.8,
1.0,
],
compression_ratio_threshold: Optional[float] = 2.4,
log_prob_threshold: Optional[float] = -1.0,
log_prob_low_threshold: Optional[float] = None,
no_speech_threshold: Optional[float] = 0.6,
initial_prompt: Optional[Union[str, Iterable[int]]] = None,
prefix: Optional[str] = None,
suppress_blank: bool = True,
suppress_tokens: Optional[List[int]] = [-1],
prepend_punctuations: str = "\"'“¿([{-",
append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
max_new_tokens: Optional[int] = None,
hotwords: Optional[str] = None,
word_timestamps: bool = False,
without_timestamps: bool = True,
) -> Tuple[Iterable[Segment], TranscriptionInfo]:
"""transcribe audio in chunks in batched fashion and return with language info.
Arguments:
audio: audio file as numpy array/path for batched transcription.
vad_segments: Optionally provide list of dictionaries each containing "start", "end",
and "segments" keys.
"start" and "end" keys specify the start and end of the voiced region within
30 sec boundary. An additional key "segments" contains all the start
and end of voiced regions within that 30sec boundary as a list of tuples.
If no vad_segments specified, it uses internal vad model automatically segment them.
batch_size: the maximum number of parallel requests to model for decoding.
language: The language spoken in the audio.
task: either "transcribe" or "translate".
log_progress: whether to show progress bar or not.
beam_size: Beam size to use for decoding.
best_of: Number of candidates when sampling with non-zero temperature.
patience: Beam search patience factor.
length_penalty: Exponential length penalty constant.
repetition_penalty: Penalty applied to the score of previously generated tokens
(set > 1 to penalize).
no_repeat_ngram_size: Prevent repetitions of ngrams with this size (set 0 to disable).
temperature: Temperature for sampling. It can be a tuple of temperatures,
which will be successively used upon failures according to either
`compression_ratio_threshold` or `log_prob_threshold`.
compression_ratio_threshold: If the gzip compression ratio is above this value,
treat as failed.
log_prob_threshold: If the average log probability over sampled tokens is
below this value, treat as failed.
log_prob_low_threshold: This parameter alone is sufficient to skip an output text,
whereas log_prob_threshold also looks for appropriate no_speech_threshold value.
This value should be less than log_prob_threshold.
no_speech_threshold: If the no_speech probability is higher than this value AND
the average log probability over sampled tokens is below `log_prob_threshold`,
consider the segment as silent.
initial_prompt: Optional text string or iterable of token ids to provide as a
prompt for the first window.
prefix: Optional text to provide as a prefix for the first window.
suppress_blank: Suppress blank outputs at the beginning of the sampling.
suppress_tokens: List of token IDs to suppress. -1 will suppress a default set
of symbols as defined in `tokenizer.non_speech_tokens()`.
prepend_punctuations: If word_timestamps is True, merge these punctuation symbols
with the next word
append_punctuations: If word_timestamps is True, merge these punctuation symbols
with the previous word
max_new_tokens: Maximum number of new tokens to generate per-chunk. If not set,
the maximum will be set by the default max_length.
hotwords:
Hotwords/hint phrases to the model. Has no effect if prefix is not None.
word_timestamps: Extract word-level timestamps using the cross-attention pattern
and dynamic time warping, and include the timestamps for each word in each segment.
Set as False.
without_timestamps: Only sample text tokens.
Static params: (Fixed for batched version)
max_initial_timestamp: The initial timestamp cannot be later than this, set at 0.0.
multilingual: If True, perform transcription on multilingual videos. Set as False.
output_language: Valid only if multilingual is set to True.
Specifies the string representing the output language. One of
'en' (English) or 'hybrid' (code-switched transcription). set as None.
condition_on_previous_text: If True, the previous output of the model is provided
as a prompt for the next window; disabling may make the text inconsistent across
windows, but the model becomes less prone to getting stuck in a failure loop,
such as repetition looping or timestamps going out of sync. Set as False
prompt_reset_on_temperature: Resets prompt if temperature is above this value.
Arg has effect only if condition_on_previous_text is True. Set at 0.5
#TODO: support "hallucination_silence_threshold" when "word_timestamps=True"
hallucination_silence_threshold: Optional[float]
When word_timestamps is True, skip silent periods longer than this threshold
(in seconds) when a possible hallucination is detected. set as None.
clip_timestamps:
Comma-separated list start,end,start,end,... timestamps (in seconds) of clips to
process. The last end timestamp defaults to the end of the file. Set as "0".
unused:
language_detection_threshold: If the maximum probability of the language tokens is
higher than this value, the language is detected.
language_detection_segments: Number of segments to consider for the language detection.
vad_filter: Enable the voice activity detection (VAD) to filter out parts of the audio
without speech. This step is using the Silero VAD model
https://github.com/snakers4/silero-vad.
vad_parameters: Dictionary of Silero VAD parameters or VadOptions class (see available
parameters and default values in the class `VadOptions`).
chunk_length: The length of audio segments. If it is not None, it will overwrite the
default chunk_length of the FeatureExtractor.
Returns:
A tuple with:
- a generator over transcribed batched segments.
- an instance of TranscriptionInfo.
"""
sampling_rate = self.model.feature_extractor.sampling_rate
if isinstance(audio, np.ndarray):
audio = torch.from_numpy(audio)
elif not isinstance(audio, torch.Tensor):
audio = decode_audio(audio, sampling_rate=sampling_rate)
duration = audio.shape[0] / sampling_rate
# if no segment split is provided, use vad_model and generate segments
if not vad_segments:
# run the audio if it is less than 30 sec even without vad_segments
if self.use_vad_model:
vad_segments = self.vad_model(
{
"waveform": audio.unsqueeze(0),
"sample_rate": 16000,
}
)
vad_segments = merge_chunks(
vad_segments,
self.chunk_length,
onset=self.vad_onset,
offset=self.vad_offset,
)
elif duration < self.chunk_length:
vad_segments = [
{"start": 0.0, "end": duration, "segments": [(0.0, duration)]}
]
else:
raise RuntimeError(
"No vad segments found. Set 'use_vad_model' to True while loading the model"
)
if self.model.model.is_multilingual:
language = language or self.preset_language
elif language != "en":
if language is not None:
self.model.logger.warning(
f"English-only model is used, but {language} language is"
"chosen, setting language to 'en'."
)
language = "en"
(
language,
language_probability,
task,
all_language_probs,
) = self.get_language_and_tokenizer(audio, task, language)
duration_after_vad = sum(
segment["end"] - segment["start"] for segment in vad_segments
)
# batched options: see the difference with default options in WhisperModel
batched_options = TranscriptionOptions(
beam_size=beam_size,
best_of=best_of,
patience=patience,
length_penalty=length_penalty,
repetition_penalty=repetition_penalty,
no_repeat_ngram_size=no_repeat_ngram_size,
log_prob_threshold=log_prob_threshold,
log_prob_low_threshold=log_prob_low_threshold,
no_speech_threshold=no_speech_threshold,
compression_ratio_threshold=compression_ratio_threshold,
temperatures=(
temperature if isinstance(temperature, (list, tuple)) else [temperature]
),
initial_prompt=initial_prompt,
prefix=prefix,
suppress_blank=suppress_blank,
suppress_tokens=get_suppressed_tokens(self.tokenizer, suppress_tokens),
prepend_punctuations=prepend_punctuations,
append_punctuations=append_punctuations,
max_new_tokens=max_new_tokens,
hotwords=hotwords,
word_timestamps=word_timestamps,
hallucination_silence_threshold=None,
condition_on_previous_text=False,
clip_timestamps="0",
prompt_reset_on_temperature=0.5,
multilingual=False,
output_language=None,
without_timestamps=without_timestamps,
max_initial_timestamp=0.0,
)
info = TranscriptionInfo(
language=language,
language_probability=language_probability,
duration=duration,
duration_after_vad=duration_after_vad,
transcription_options=batched_options,
vad_options=None,
all_language_probs=all_language_probs,
)
audio_segments, segments_metadata = self.audio_split(
audio, vad_segments, sampling_rate
)
to_cpu = (
self.model.model.device == "cuda" and len(self.model.model.device_index) > 1
)
audio_segments = torch.nested.nested_tensor(audio_segments).to_padded_tensor(
padding=0
)
features = torch.stack(
[
self.model.feature_extractor(audio_segment, to_cpu=to_cpu)[
..., : self.model.feature_extractor.nb_max_frames
]
for audio_segment in audio_segments
]
)
segments = self._batched_segments_generator(
features,
segments_metadata,
batch_size,
batched_options,
log_progress,
)
return segments, info
def _batched_segments_generator(
self, features, segments_metadata, batch_size, options, log_progress
):
pbar = tqdm(total=len(features), disable=not log_progress, position=0)
seg_idx = 0
for i in range(0, len(features), batch_size):
results = self.forward(
features[i : i + batch_size],
segments_metadata[i : i + batch_size],
**options._asdict(),
)
for result in results:
for segment in result:
seg_idx += 1
yield Segment(
seek=int(result[-1]["end"] * self.model.frames_per_second),
id=seg_idx,
text=segment["text"],
start=round(segment["start"], 3),
end=round(segment["end"], 3),
words=(
None
if not options.word_timestamps
else [Word(**word) for word in segment["words"]]
),
tokens=segment["tokens"],
avg_logprob=segment["avg_logprob"],
no_speech_prob=segment["no_speech_prob"],
compression_ratio=segment["compression_ratio"],
)
pbar.update(1)
pbar.close()
# revert the tokenizer if multilingual inference is enabled
if self.preset_language is None:
self.tokenizer = None
self.last_speech_timestamp = 0.0
class WhisperModel:
def __init__(
self,
model_size_or_path: str,
device: str = "auto",
device_index: Union[int, List[int]] = 0,
compute_type: str = "default",
cpu_threads: int = 16,
num_workers: int = 1,
download_root: Optional[str] = None,
local_files_only: bool = False,
files: dict = None,
**model_kwargs,
):
"""Initializes the Whisper model.
Args:
model_size_or_path: Size of the model to use (tiny, tiny.en, base, base.en,
small, small.en, distil-small.en, medium, medium.en, distil-medium.en, large-v1,
large-v2, large-v3, large, distil-large-v2 or distil-large-v3), a path to a
converted model directory, or a CTranslate2-converted Whisper model ID from the HF Hub.
When a size or a model ID is configured, the converted model is downloaded
from the Hugging Face Hub.
device: Device to use for computation ("cpu", "cuda", "auto").
device_index: Device ID to use.
The model can also be loaded on multiple GPUs by passing a list of IDs
(e.g. [0, 1, 2, 3]). In that case, multiple transcriptions can run in parallel
when transcribe() is called from multiple Python threads (see also num_workers).
compute_type: Type to use for computation.
See https://opennmt.net/CTranslate2/quantization.html.
cpu_threads: Number of threads to use when running on CPU (4 by default).
A non zero value overrides the OMP_NUM_THREADS environment variable.
num_workers: When transcribe() is called from multiple Python threads,
having multiple workers enables true parallelism when running the model
(concurrent calls to self.model.generate() will run in parallel).
This can improve the global throughput at the cost of increased memory usage.
download_root: Directory where the models should be saved. If not set, the models
are saved in the standard Hugging Face cache directory.
local_files_only: If True, avoid downloading the file and return the path to the
local cached file if it exists.
files: Load model files from the memory. This argument is a dictionary mapping file names
to file contents as file-like or bytes objects. If this is set, model_path acts as an
identifier for this model.
"""
self.logger = get_logger()
tokenizer_bytes, preprocessor_bytes = None, None
if files:
model_path = model_size_or_path
tokenizer_bytes = files.pop("tokenizer.json", None)
preprocessor_bytes = files.pop("preprocessor_config.json", None)
elif os.path.isdir(model_size_or_path):
model_path = model_size_or_path
else:
model_path = download_model(
model_size_or_path,
local_files_only=local_files_only,
cache_dir=download_root,
)
self.device = device
# set the random seed to make sure consistency across runs
ctranslate2.set_random_seed(42)
self.model = ctranslate2.models.Whisper(
model_path,
device=self.device,
device_index=device_index,
compute_type=compute_type,
intra_threads=cpu_threads,
inter_threads=num_workers,
files=files,
**model_kwargs,
)
tokenizer_file = os.path.join(model_path, "tokenizer.json")
if tokenizer_bytes:
self.hf_tokenizer = tokenizers.Tokenizer.from_buffer(tokenizer_bytes)
elif os.path.isfile(tokenizer_file):
self.hf_tokenizer = tokenizers.Tokenizer.from_file(tokenizer_file)
else:
self.hf_tokenizer = tokenizers.Tokenizer.from_pretrained(
"openai/whisper-tiny" + ("" if self.model.is_multilingual else ".en")
)
self.feat_kwargs = self._get_feature_kwargs(model_path, preprocessor_bytes)
self.feature_extractor = FeatureExtractor(
**self.feat_kwargs, device=self.device
)
self.input_stride = 2
self.num_samples_per_token = (
self.feature_extractor.hop_length * self.input_stride
)
self.frames_per_second = (
self.feature_extractor.sampling_rate // self.feature_extractor.hop_length
)
self.tokens_per_second = (
self.feature_extractor.sampling_rate // self.num_samples_per_token
)
self.time_precision = 0.02
self.max_length = 448
@property
def supported_languages(self) -> List[str]:
"""The languages supported by the model."""
return list(_LANGUAGE_CODES) if self.model.is_multilingual else ["en"]
def _get_feature_kwargs(self, model_path, preprocessor_bytes=None) -> dict:
config = {}
try:
config_path = os.path.join(model_path, "preprocessor_config.json")
if preprocessor_bytes:
config = json.loads(preprocessor_bytes)
elif os.path.isfile(config_path):
with open(config_path, "r", encoding="utf-8") as file:
config = json.load(file)
else:
return config
valid_keys = signature(FeatureExtractor.__init__).parameters.keys()
return {k: v for k, v in config.items() if k in valid_keys}
except json.JSONDecodeError as e:
self.logger.warning("Could not load preprocessor config: %s", e)
return config
def transcribe(
self,
audio: Union[str, BinaryIO, torch.Tensor, np.ndarray],
language: Optional[str] = None,
task: str = "transcribe",
beam_size: int = 5,
best_of: int = 5,
patience: float = 1,
length_penalty: float = 1,
repetition_penalty: float = 1,
no_repeat_ngram_size: int = 0,
temperature: Union[float, List[float], Tuple[float, ...]] = [
0.0,
0.2,
0.4,
0.6,
0.8,
1.0,
],
compression_ratio_threshold: Optional[float] = 2.4,
log_prob_threshold: Optional[float] = -1.0,
log_prob_low_threshold: Optional[float] = None,
no_speech_threshold: Optional[float] = 0.6,
condition_on_previous_text: bool = True,
prompt_reset_on_temperature: float = 0.5,
initial_prompt: Optional[Union[str, Iterable[int]]] = None,
prefix: Optional[str] = None,
suppress_blank: bool = True,
suppress_tokens: Optional[List[int]] = [-1],
without_timestamps: bool = False,
max_initial_timestamp: float = 1.0,
word_timestamps: bool = False,
prepend_punctuations: str = "\"'“¿([{-",
append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
multilingual: bool = False,
output_language: Optional[str] = None,
vad_filter: bool = False,
vad_parameters: Optional[Union[dict, VadOptions]] = None,
max_new_tokens: Optional[int] = None,
chunk_length: Optional[int] = None,
clip_timestamps: Union[str, List[float]] = "0",
hallucination_silence_threshold: Optional[float] = None,
hotwords: Optional[str] = None,
language_detection_threshold: Optional[float] = None,
language_detection_segments: int = 1,
) -> Tuple[Iterable[Segment], TranscriptionInfo]:
"""Transcribes an input file.
Arguments:
audio: Path to the input file (or a file-like object), or the audio waveform.
language: The language spoken in the audio. It should be a language code such
as "en" or "fr". If not set, the language will be detected in the first 30 seconds
of audio.
task: Task to execute (transcribe or translate).
beam_size: Beam size to use for decoding.
best_of: Number of candidates when sampling with non-zero temperature.
patience: Beam search patience factor.
length_penalty: Exponential length penalty constant.
repetition_penalty: Penalty applied to the score of previously generated tokens
(set > 1 to penalize).
no_repeat_ngram_size: Prevent repetitions of ngrams with this size (set 0 to disable).
temperature: Temperature for sampling. It can be a tuple of temperatures,
which will be successively used upon failures according to either
`compression_ratio_threshold` or `log_prob_threshold`.
compression_ratio_threshold: If the gzip compression ratio is above this value,
treat as failed.
log_prob_threshold: If the average log probability over sampled tokens is
below this value, treat as failed.
log_prob_low_threshold: This parameter alone is sufficient to skip an output text,
wheras log_prob_threshold also looks for appropriate no_speech_threshold value.
This value should be less than log_prob_threshold.
no_speech_threshold: If the no_speech probability is higher than this value AND
the average log probability over sampled tokens is below `log_prob_threshold`,
consider the segment as silent.
condition_on_previous_text: If True, the previous output of the model is provided
as a prompt for the next window; disabling may make the text inconsistent across
windows, but the model becomes less prone to getting stuck in a failure loop,
such as repetition looping or timestamps going out of sync.
prompt_reset_on_temperature: Resets prompt if temperature is above this value.
Arg has effect only if condition_on_previous_text is True.
initial_prompt: Optional text string or iterable of token ids to provide as a
prompt for the first window.
prefix: Optional text to provide as a prefix for the first window.
suppress_blank: Suppress blank outputs at the beginning of the sampling.
suppress_tokens: List of token IDs to suppress. -1 will suppress a default set
of symbols as defined in `tokenizer.non_speech_tokens()`.
without_timestamps: Only sample text tokens.
max_initial_timestamp: The initial timestamp cannot be later than this.
word_timestamps: Extract word-level timestamps using the cross-attention pattern
and dynamic time warping, and include the timestamps for each word in each segment.
prepend_punctuations: If word_timestamps is True, merge these punctuation symbols
with the next word
append_punctuations: If word_timestamps is True, merge these punctuation symbols
with the previous word
multilingual: If True, perform transcription on multilingual videos
and return the transcript based
on the 'output_language' flag.
output_language: Valid only if multilingual is set to True.
Specifies the string representing the output language. One of
'en' (English) or 'hybrid' (code-switched transcription).
vad_filter: Enable the voice activity detection (VAD) to filter out parts of the audio
without speech. This step is using the Silero VAD model
https://github.com/snakers4/silero-vad.
vad_parameters: Dictionary of Silero VAD parameters or VadOptions class (see available
parameters and default values in the class `VadOptions`).
max_new_tokens: Maximum number of new tokens to generate per-chunk. If not set,
the maximum will be set by the default max_length.
chunk_length: The length of audio segments. If it is not None, it will overwrite the
default chunk_length of the FeatureExtractor.
clip_timestamps:
Comma-separated list start,end,start,end,... timestamps (in seconds) of clips to
process. The last end timestamp defaults to the end of the file.
vad_filter will be ignored if clip_timestamps is used.
hallucination_silence_threshold:
When word_timestamps is True, skip silent periods longer than this threshold
(in seconds) when a possible hallucination is detected
hotwords:
Hotwords/hint phrases to provide the model with. Has no effect if prefix is not None.
language_detection_threshold: If the maximum probability of the language tokens is higher
than this value, the language is detected.
language_detection_segments: Number of segments to consider for the language detection.
Returns:
A tuple with:
- a generator over transcribed segments
- an instance of TranscriptionInfo
"""
sampling_rate = self.feature_extractor.sampling_rate
if isinstance(audio, np.ndarray):
audio = torch.from_numpy(audio)
elif not isinstance(audio, torch.Tensor):
audio = decode_audio(audio, sampling_rate=sampling_rate)
duration = audio.shape[0] / sampling_rate
duration_after_vad = duration
self.logger.info(
"Processing audio with duration %s", format_timestamp(duration)
)
if vad_filter and clip_timestamps == "0":
if vad_parameters is None:
vad_parameters = VadOptions()
elif isinstance(vad_parameters, dict):
vad_parameters = VadOptions(**vad_parameters)
speech_chunks = get_speech_timestamps(audio, vad_parameters)
audio = collect_chunks(audio, speech_chunks)
duration_after_vad = audio.shape[0] / sampling_rate
self.logger.info(
"VAD filter removed %s of audio",
format_timestamp(duration - duration_after_vad),
)
if self.logger.isEnabledFor(logging.DEBUG):
self.logger.debug(
"VAD filter kept the following audio segments: %s",
", ".join(
"[%s -> %s]"
% (
format_timestamp(chunk["start"] / sampling_rate),
format_timestamp(chunk["end"] / sampling_rate),
)
for chunk in speech_chunks
),
)
else:
speech_chunks = None
to_cpu = self.model.device == "cuda" and len(self.model.device_index) > 1
features = self.feature_extractor(
audio, chunk_length=chunk_length, to_cpu=to_cpu
)
encoder_output = None
all_language_probs = None
# setting output_language for multilingual videos
if multilingual:
if output_language is None:
output_language = "en"
elif output_language not in ["en", "hybrid"]:
raise ValueError("Output language needs to be one of 'en'/'hybrid'.")
# detecting the language if not provided
if language is None:
if not self.model.is_multilingual:
language = "en"
language_probability = 1
else:
if (
language_detection_segments is None
or language_detection_segments < 1
):
language_detection_segments = 1
start_timestamp = (
float(clip_timestamps.split(",")[0])
if isinstance(clip_timestamps, str)
else clip_timestamps[0]
)
content_frames = (
features.shape[-1] - self.feature_extractor.nb_max_frames
)
seek = (
int(start_timestamp * self.frames_per_second)
if start_timestamp * self.frames_per_second < content_frames
else 0
)
end_frames = min(
seek
+ self.feature_extractor.nb_max_frames
* language_detection_segments,
content_frames,
)
detected_language_info = {}
while seek <= end_frames:
segment = features[
:, seek : seek + self.feature_extractor.nb_max_frames
]
encoder_output = self.encode(segment)
# results is a list of tuple[str, float] with language names and
# probabilities.
results = self.model.detect_language(encoder_output)[0]
# Parse language names to strip out markers
all_language_probs = [
(token[2:-2], prob) for (token, prob) in results
]
# Get top language token and probability
language, language_probability = all_language_probs[0]
if (
language_detection_threshold is None
or language_probability > language_detection_threshold
):
break
detected_language_info.setdefault(language, []).append(
language_probability
)
seek += segment.shape[-1]
else:
# If no language detected for all segments, the majority vote of the highest
# projected languages for all segments is used to determine the language.
language = max(
detected_language_info,
key=lambda lang: len(detected_language_info[lang]),
)
language_probability = max(detected_language_info[language])
self.logger.info(
"Detected language '%s' with probability %.2f",
language,
language_probability,
)
else:
if not self.model.is_multilingual and language != "en":
self.logger.warning(
"The current model is English-only but the language parameter is set to '%s'; "
"using 'en' instead." % language
)
language = "en"
language_probability = 1
tokenizer = Tokenizer(
self.hf_tokenizer,
self.model.is_multilingual,
task=task,
language=language,
)
options = TranscriptionOptions(
beam_size=beam_size,
best_of=best_of,
patience=patience,
length_penalty=length_penalty,
repetition_penalty=repetition_penalty,
no_repeat_ngram_size=no_repeat_ngram_size,
log_prob_threshold=log_prob_threshold,
log_prob_low_threshold=log_prob_low_threshold,
no_speech_threshold=no_speech_threshold,
compression_ratio_threshold=compression_ratio_threshold,
condition_on_previous_text=condition_on_previous_text,
prompt_reset_on_temperature=prompt_reset_on_temperature,
temperatures=(
temperature if isinstance(temperature, (list, tuple)) else [temperature]
),
initial_prompt=initial_prompt,
prefix=prefix,
suppress_blank=suppress_blank,
suppress_tokens=(
get_suppressed_tokens(tokenizer, suppress_tokens)
if suppress_tokens
else suppress_tokens
),
without_timestamps=without_timestamps,
max_initial_timestamp=max_initial_timestamp,
word_timestamps=word_timestamps,
prepend_punctuations=prepend_punctuations,
append_punctuations=append_punctuations,
multilingual=multilingual,
output_language=output_language,
max_new_tokens=max_new_tokens,
clip_timestamps=clip_timestamps,
hallucination_silence_threshold=hallucination_silence_threshold,
hotwords=hotwords,
)
segments = self.generate_segments(features, tokenizer, options, encoder_output)
if speech_chunks:
segments = restore_speech_timestamps(segments, speech_chunks, sampling_rate)
info = TranscriptionInfo(
language=language,
language_probability=language_probability,
duration=duration,
duration_after_vad=duration_after_vad,
transcription_options=options,
vad_options=vad_parameters,
all_language_probs=all_language_probs,
)
return segments, info
def _split_segments_by_timestamps(
self,
tokenizer: Tokenizer,
tokens: List[int],
time_offset: float,
segment_size: int,
segment_duration: float,
seek: int,
) -> List[List[int]]:
current_segments = []
single_timestamp_ending = (
len(tokens) >= 2 and tokens[-2] < tokenizer.timestamp_begin <= tokens[-1]
)
consecutive_timestamps = [
i
for i in range(len(tokens))
if i > 0
and tokens[i] >= tokenizer.timestamp_begin
and tokens[i - 1] >= tokenizer.timestamp_begin
]
if len(consecutive_timestamps) > 0:
slices = list(consecutive_timestamps)
if single_timestamp_ending:
slices.append(len(tokens))
last_slice = 0
for current_slice in slices:
sliced_tokens = tokens[last_slice:current_slice]
start_timestamp_position = sliced_tokens[0] - tokenizer.timestamp_begin
end_timestamp_position = sliced_tokens[-1] - tokenizer.timestamp_begin
start_time = (
time_offset + start_timestamp_position * self.time_precision
)
end_time = time_offset + end_timestamp_position * self.time_precision
current_segments.append(
dict(
seek=seek,
start=start_time,
end=end_time,
tokens=sliced_tokens,
)
)
last_slice = current_slice
if single_timestamp_ending:
# single timestamp at the end means no speech after the last timestamp.
seek += segment_size
else:
# otherwise, ignore the unfinished segment and seek to the last timestamp
last_timestamp_position = (
tokens[last_slice - 1] - tokenizer.timestamp_begin
)
seek += last_timestamp_position * self.input_stride
else:
duration = segment_duration
timestamps = [
token for token in tokens if token >= tokenizer.timestamp_begin
]
if len(timestamps) > 0 and timestamps[-1] != tokenizer.timestamp_begin:
last_timestamp_position = timestamps[-1] - tokenizer.timestamp_begin
duration = last_timestamp_position * self.time_precision
current_segments.append(
dict(
seek=seek,
start=time_offset,
end=time_offset + duration,
tokens=tokens,
)
)
seek += segment_size
return current_segments, seek, single_timestamp_ending
def generate_segments(
self,
features: torch.Tensor,
tokenizer: Tokenizer,
options: TranscriptionOptions,
encoder_output: Optional[ctranslate2.StorageView] = None,
) -> Iterable[Segment]:
content_frames = features.shape[-1] - self.feature_extractor.nb_max_frames
content_duration = float(content_frames * self.feature_extractor.time_per_frame)
if isinstance(options.clip_timestamps, str):
options = options._replace(
clip_timestamps=[
float(ts)
for ts in (
options.clip_timestamps.split(",")
if options.clip_timestamps
else []
)
]
)
seek_points: List[int] = [
round(ts * self.frames_per_second) for ts in options.clip_timestamps
]
if len(seek_points) == 0:
seek_points.append(0)
if len(seek_points) % 2 == 1:
seek_points.append(content_frames)
seek_clips: List[Tuple[int, int]] = list(
zip(seek_points[::2], seek_points[1::2])
)
punctuation = "\"'“¿([{-\"'.。,，!！?？:：”)]}、"
idx = 0
clip_idx = 0
seek = seek_clips[clip_idx][0]
all_tokens = []
prompt_reset_since = 0
if options.initial_prompt is not None:
if isinstance(options.initial_prompt, str):
initial_prompt = " " + options.initial_prompt.strip()
initial_prompt_tokens = tokenizer.encode(initial_prompt)
all_tokens.extend(initial_prompt_tokens)
else:
all_tokens.extend(options.initial_prompt)
last_speech_timestamp = 0.0
# NOTE: This loop is obscurely flattened to make the diff readable.
# A later commit should turn this into a simpler nested loop.
# for seek_clip_start, seek_clip_end in seek_clips:
# while seek < seek_clip_end
while clip_idx < len(seek_clips):
seek_clip_start, seek_clip_end = seek_clips[clip_idx]
if seek_clip_end > content_frames:
seek_clip_end = content_frames
if seek < seek_clip_start:
seek = seek_clip_start
if seek >= seek_clip_end:
clip_idx += 1
if clip_idx < len(seek_clips):
seek = seek_clips[clip_idx][0]
continue
time_offset = seek * self.feature_extractor.time_per_frame
window_end_time = float(
(seek + self.feature_extractor.nb_max_frames)
* self.feature_extractor.time_per_frame
)
segment_size = min(
self.feature_extractor.nb_max_frames,
content_frames - seek,
seek_clip_end - seek,
)
segment = features[:, seek : seek + segment_size]
segment_duration = segment_size * self.feature_extractor.time_per_frame
segment = pad_or_trim(segment, self.feature_extractor.nb_max_frames)
if self.logger.isEnabledFor(logging.DEBUG):
self.logger.debug(
"Processing segment at %s", format_timestamp(time_offset)
)
previous_tokens = all_tokens[prompt_reset_since:]
if encoder_output is None:
encoder_output = self.encode(segment)
# Perform language detection at every segment to update task based on output language,
# if the language is english, task is transcribe,
# else the task is translate to english (default)
# or transcribe if 'output_language' is 'hybrid'.
if options.multilingual:
results = self.model.detect_language(encoder_output)
language_token, language_probability = results[0][0]
language = language_token[2:-2]
if options.output_language == "en" and language != "en":
task = "translate"
else:
task = "transcribe"
# Update tokenizer based on task and language
tokenizer.task = tokenizer.tokenizer.token_to_id(f"<|{task}|>")
tokenizer.language = tokenizer.tokenizer.token_to_id(language_token)
tokenizer.language_code = language
# Update prompt based on task and language
prompt = self.get_prompt(
tokenizer,
previous_tokens,
without_timestamps=options.without_timestamps,
prefix=options.prefix if seek == 0 else None,
hotwords=options.hotwords,
)
if seek > 0 or encoder_output is None:
encoder_output = self.encode(segment)
(
result,
avg_logprob,
temperature,
compression_ratio,
) = self.generate_with_fallback(encoder_output, prompt, tokenizer, options)
if options.no_speech_threshold is not None:
# no voice activity check
should_skip = result.no_speech_prob > options.no_speech_threshold
if (
options.log_prob_threshold is not None
and avg_logprob > options.log_prob_threshold
):
# don't skip if the logprob is high enough, despite the no_speech_prob
should_skip = False
if should_skip:
self.logger.debug(
"No speech threshold is met (%f > %f)",
result.no_speech_prob,
options.no_speech_threshold,
)
# Skip if the logprob is very low (below the threshold value),
# despite no_speech_prob being low (ex: Too ambiguous outputs)
if options.log_prob_low_threshold:
if avg_logprob < options.log_prob_low_threshold:
should_skip = True
self.logger.debug(
"log prob low threshold is met (%f > %f)",
avg_logprob,
options.log_prob_low_threshold,
)
if should_skip:
# fast-forward to the next segment boundary
seek += segment_size
continue
tokens = result.sequences_ids[0]
previous_seek = seek
# anomalous words are very long/short/improbable
def word_anomaly_score(word: dict) -> float:
probability = word.get("probability", 0.0)
duration = word["end"] - word["start"]
score = 0.0
if probability < 0.15:
score += 1.0
if duration < 0.133:
score += (0.133 - duration) * 15
if duration > 2.0:
score += duration - 2.0
return score
def is_segment_anomaly(segment: Optional[dict]) -> bool:
if segment is None or not segment["words"]:
return False
words = [w for w in segment["words"] if w["word"] not in punctuation]
words = words[:8]
score = sum(word_anomaly_score(w) for w in words)
return score >= 3 or score + 0.01 >= len(words)
def next_words_segment(segments: List[dict]) -> Optional[dict]:
return next((s for s in segments if s["words"]), None)
(
current_segments,
seek,
single_timestamp_ending,
) = self._split_segments_by_timestamps(
tokenizer=tokenizer,
tokens=tokens,
time_offset=time_offset,
segment_size=segment_size,
segment_duration=segment_duration,
seek=seek,
)
if options.word_timestamps:
self.add_word_timestamps(
[current_segments],
tokenizer,
encoder_output,
segment_size,
options.prepend_punctuations,
options.append_punctuations,
last_speech_timestamp=last_speech_timestamp,
)
if not single_timestamp_ending:
last_word_end = get_end(current_segments)
if last_word_end is not None and last_word_end > time_offset:
seek = round(last_word_end * self.frames_per_second)
# skip silence before possible hallucinations
if options.hallucination_silence_threshold is not None:
threshold = options.hallucination_silence_threshold
# if first segment might be a hallucination, skip leading silence
first_segment = next_words_segment(current_segments)
if first_segment is not None and is_segment_anomaly(first_segment):
gap = first_segment["start"] - time_offset
if gap > threshold:
seek = previous_seek + round(gap * self.frames_per_second)
continue
# skip silence before any possible hallucination that is surrounded
# by silence or more hallucinations
hal_last_end = last_speech_timestamp
for si in range(len(current_segments)):
segment = current_segments[si]
if not segment["words"]:
continue
if is_segment_anomaly(segment):
next_segment = next_words_segment(
current_segments[si + 1 :]
)
if next_segment is not None:
hal_next_start = next_segment["words"][0]["start"]
else:
hal_next_start = time_offset + segment_duration
silence_before = (
segment["start"] - hal_last_end > threshold
or segment["start"] < threshold
or segment["start"] - time_offset < 2.0
)
silence_after = (
hal_next_start - segment["end"] > threshold
or is_segment_anomaly(next_segment)
or window_end_time - segment["end"] < 2.0
)
if silence_before and silence_after:
seek = round(
max(time_offset + 1, segment["start"])
* self.frames_per_second
)
if content_duration - segment["end"] < threshold:
seek = content_frames
current_segments[si:] = []
break
hal_last_end = segment["end"]
last_word_end = get_end(current_segments)
if last_word_end is not None:
last_speech_timestamp = last_word_end
for segment in current_segments:
tokens = segment["tokens"]
text = tokenizer.decode(tokens)
if segment["start"] == segment["end"] or not text.strip():
continue
all_tokens.extend(tokens)
idx += 1
yield Segment(
id=idx,
seek=seek,
start=segment["start"],
end=segment["end"],
text=text,
tokens=tokens,
temperature=temperature,
avg_logprob=avg_logprob,
compression_ratio=compression_ratio,
no_speech_prob=result.no_speech_prob,
words=(
[Word(**word) for word in segment["words"]]
if options.word_timestamps
else None
),
)
if (
not options.condition_on_previous_text
or temperature > options.prompt_reset_on_temperature
):
if options.condition_on_previous_text:
self.logger.debug(
"Reset prompt. prompt_reset_on_temperature threshold is met %f > %f",
temperature,
options.prompt_reset_on_temperature,
)
prompt_reset_since = len(all_tokens)
def encode(self, features: torch.Tensor) -> ctranslate2.StorageView:
# When the model is running on multiple GPUs, the encoder output should be moved
# to the CPU since we don't know which GPU will handle the next job.
to_cpu = self.model.device == "cuda" and len(self.model.device_index) > 1
if features.ndim == 2:
features = features.unsqueeze(0)
features = get_ctranslate2_storage(features)
return self.model.encode(features, to_cpu=to_cpu)
def generate_with_fallback(
self,
encoder_output: ctranslate2.StorageView,
prompt: List[int],
tokenizer: Tokenizer,
options: TranscriptionOptions,
) -> Tuple[ctranslate2.models.WhisperGenerationResult, float, float, float]:
decode_result = None
all_results = []
below_cr_threshold_results = []
max_initial_timestamp_index = int(
round(options.max_initial_timestamp / self.time_precision)
)
if options.max_new_tokens is not None:
max_length = len(prompt) + options.max_new_tokens
else:
max_length = self.max_length
if max_length > self.max_length:
raise ValueError(
f"The length of the prompt is {len(prompt)}, and the `max_new_tokens` "
f"{max_length - len(prompt)}. Thus, the combined length of the prompt "
f"and `max_new_tokens` is: {max_length}. This exceeds the "
f"`max_length` of the Whisper model: {self.max_length}. "
"You should either reduce the length of your prompt, or "
"reduce the value of `max_new_tokens`, "
f"so that their combined length is less that {self.max_length}."
)
for temperature in options.temperatures:
if temperature > 0:
kwargs = {
"beam_size": 1,
"num_hypotheses": options.best_of,
"sampling_topk": 0,
"sampling_temperature": temperature,
}
else:
kwargs = {
"beam_size": options.beam_size,
"patience": options.patience,
}
result = self.model.generate(
encoder_output,
[prompt],
length_penalty=options.length_penalty,
repetition_penalty=options.repetition_penalty,
no_repeat_ngram_size=options.no_repeat_ngram_size,
max_length=max_length,
return_scores=True,
return_no_speech_prob=True,
suppress_blank=options.suppress_blank,
suppress_tokens=options.suppress_tokens,
max_initial_timestamp_index=max_initial_timestamp_index,
**kwargs,
)[0]
tokens = result.sequences_ids[0]
# Recover the average log prob from the returned score.
seq_len = len(tokens)
cum_logprob = result.scores[0] * (seq_len**options.length_penalty)
avg_logprob = cum_logprob / (seq_len + 1)
text = tokenizer.decode(tokens).strip()
compression_ratio = get_compression_ratio(text)
decode_result = (
result,
avg_logprob,
temperature,
compression_ratio,
)
all_results.append(decode_result)
needs_fallback = False
if options.compression_ratio_threshold is not None:
if compression_ratio > options.compression_ratio_threshold:
needs_fallback = True # too repetitive
self.logger.debug(
"Compression ratio threshold is not met with temperature %.1f (%f > %f)",
temperature,
compression_ratio,
options.compression_ratio_threshold,
)
else:
below_cr_threshold_results.append(decode_result)
if (
options.log_prob_threshold is not None
and avg_logprob < options.log_prob_threshold
):
needs_fallback = True # average log probability is too low
self.logger.debug(
"Log probability threshold is not met with temperature %.1f (%f < %f)",
temperature,
avg_logprob,
options.log_prob_threshold,
)
if (
options.no_speech_threshold is not None
and result.no_speech_prob > options.no_speech_threshold
and options.log_prob_threshold is not None
and avg_logprob < options.log_prob_threshold
):
needs_fallback = False # silence
if not needs_fallback:
break
else:
# all failed, select the result with the highest average log probability
decode_result = max(
below_cr_threshold_results or all_results, key=lambda x: x[1]
)
# to pass final temperature for prompt_reset_on_temperature
decode_result = (
decode_result[0],
decode_result[1],
temperature,
decode_result[3],
)
return decode_result
def get_prompt(
self,
tokenizer: Tokenizer,
previous_tokens: List[int],
without_timestamps: bool = False,
prefix: Optional[str] = None,
hotwords: Optional[str] = None,
) -> List[int]:
prompt = []
if previous_tokens or (hotwords and not prefix):
prompt.append(tokenizer.sot_prev)
if hotwords and not prefix:
hotwords_tokens = tokenizer.encode(" " + hotwords.strip())
if len(hotwords_tokens) >= self.max_length // 2:
hotwords_tokens = hotwords_tokens[: self.max_length // 2 - 1]
prompt.extend(hotwords_tokens)
if previous_tokens:
prompt.extend(previous_tokens[-(self.max_length // 2 - 1) :])
prompt.extend(tokenizer.sot_sequence)
if without_timestamps:
prompt.append(tokenizer.no_timestamps)
if prefix:
prefix_tokens = tokenizer.encode(" " + prefix.strip())
if len(prefix_tokens) >= self.max_length // 2:
prefix_tokens = prefix_tokens[: self.max_length // 2 - 1]
if not without_timestamps:
prompt.append(tokenizer.timestamp_begin)
prompt.extend(prefix_tokens)
return prompt
def add_word_timestamps(
self,
segments: List[dict],
tokenizer: Tokenizer,
encoder_output: ctranslate2.StorageView,
num_frames: int,
prepend_punctuations: str,
append_punctuations: str,
last_speech_timestamp: float,
) -> float:
if len(segments) == 0:
return
text_tokens = []
text_tokens_per_segment = []
for segment in segments:
segment_tokens = [
[token for token in subsegment["tokens"] if token < tokenizer.eot]
for subsegment in segment
]
text_tokens.append(list(itertools.chain.from_iterable(segment_tokens)))
text_tokens_per_segment.append(segment_tokens)
alignments = self.find_alignment(
tokenizer, text_tokens, encoder_output, num_frames
)
median_max_durations = []
for alignment in alignments:
word_durations = np.array(
[word["end"] - word["start"] for word in alignment]
)
word_durations = word_durations[word_durations.nonzero()]
median_duration = (
np.median(word_durations) if len(word_durations) > 0 else 0.0
)
median_duration = min(0.7, float(median_duration))
max_duration = median_duration * 2
# hack: truncate long words at sentence boundaries.
# a better segmentation algorithm based on VAD should be able to replace this.
if len(word_durations) > 0:
sentence_end_marks = ".。!！?？"
# ensure words at sentence boundaries
# are not longer than twice the median word duration.
for i in range(1, len(alignment)):
if alignment[i]["end"] - alignment[i]["start"] > max_duration:
if alignment[i]["word"] in sentence_end_marks:
alignment[i]["end"] = alignment[i]["start"] + max_duration
elif alignment[i - 1]["word"] in sentence_end_marks:
alignment[i]["start"] = alignment[i]["end"] - max_duration
merge_punctuations(alignment, prepend_punctuations, append_punctuations)
median_max_durations.append((median_duration, max_duration))
for segment_idx, segment in enumerate(segments):
word_index = 0
time_offset = segment[0]["start"]
median_duration, max_duration = median_max_durations[segment_idx]
for subsegment_idx, subsegment in enumerate(segment):
saved_tokens = 0
words = []
while word_index < len(alignments[segment_idx]) and saved_tokens < len(
text_tokens_per_segment[segment_idx][subsegment_idx]
):
timing = alignments[segment_idx][word_index]
if timing["word"]:
words.append(
dict(
word=timing["word"],
start=round(time_offset + timing["start"], 2),
end=round(time_offset + timing["end"], 2),
probability=timing["probability"],
)
)
saved_tokens += len(timing["tokens"])
word_index += 1
# hack: truncate long words at segment boundaries.
# a better segmentation algorithm based on VAD should be able to replace this.
if len(words) > 0:
# ensure the first and second word after a pause is not longer than
# twice the median word duration.
if words[0][
"end"
] - last_speech_timestamp > median_duration * 4 and (
words[0]["end"] - words[0]["start"] > max_duration
or (
len(words) > 1
and words[1]["end"] - words[0]["start"] > max_duration * 2
)
):
if (
len(words) > 1
and words[1]["end"] - words[1]["start"] > max_duration
):
boundary = max(
words[1]["end"] / 2, words[1]["end"] - max_duration
)
words[0]["end"] = words[1]["start"] = boundary
words[0]["start"] = max(0, words[0]["end"] - max_duration)
# prefer the segment-level start timestamp if the first word is too long.
if (
subsegment["start"] < words[0]["end"]
and subsegment["start"] - 0.5 > words[0]["start"]
):
words[0]["start"] = max(
0,
min(words[0]["end"] - median_duration, subsegment["start"]),
)
else:
subsegment["start"] = words[0]["start"]
# prefer the segment-level end timestamp if the last word is too long.
if (
subsegment["end"] > words[-1]["start"]
and subsegment["end"] + 0.5 < words[-1]["end"]
):
words[-1]["end"] = max(
words[-1]["start"] + median_duration, subsegment["end"]
)
else:
subsegment["end"] = words[-1]["end"]
last_speech_timestamp = subsegment["end"]
segments[segment_idx][subsegment_idx]["words"] = words
return last_speech_timestamp
def find_alignment(
self,
tokenizer: Tokenizer,
text_tokens: List[int],
encoder_output: ctranslate2.StorageView,
num_frames: int,
median_filter_width: int = 7,
) -> List[dict]:
if len(text_tokens) == 0:
return []
results = self.model.align(
encoder_output,
tokenizer.sot_sequence,
text_tokens,
num_frames,
median_filter_width=median_filter_width,
)
return_list = []
for result, text_token in zip(results, text_tokens):
text_token_probs = result.text_token_probs
alignments = result.alignments
text_indices = np.array([pair[0] for pair in alignments])
time_indices = np.array([pair[1] for pair in alignments])
words, word_tokens = tokenizer.split_to_word_tokens(
text_token + [tokenizer.eot]
)
if len(word_tokens) <= 1:
# return on eot only
# >>> np.pad([], (1, 0))
# array([0.])
# This results in crashes when we lookup jump_times with float, like
# IndexError: arrays used as indices must be of integer (or boolean) type
return []
word_boundaries = np.pad(
np.cumsum([len(t) for t in word_tokens[:-1]]), (1, 0)
)
if len(word_boundaries) <= 1:
return []
jumps = np.pad(np.diff(text_indices), (1, 0), constant_values=1).astype(
bool
)
jump_times = time_indices[jumps] / self.tokens_per_second
start_times = jump_times[word_boundaries[:-1]]
end_times = jump_times[word_boundaries[1:]]
word_probabilities = [
np.mean(text_token_probs[i:j])
for i, j in zip(word_boundaries[:-1], word_boundaries[1:])
]
return_list.append(
[
dict(
word=word,
tokens=tokens,
start=start,
end=end,
probability=probability,
)
for word, tokens, start, end, probability in zip(
words, word_tokens, start_times, end_times, word_probabilities
)
]
)
return return_list
def generate_segment_batched(
self,
features: torch.Tensor,
tokenizer: Tokenizer,
options: dict,
):
batch_size = features.shape[0]
all_tokens = []
prompt_reset_since = 0
if options["initial_prompt"] is not None:
initial_prompt = " " + options["initial_prompt"].strip()
initial_prompt_tokens = tokenizer.encode(initial_prompt)
all_tokens.extend(initial_prompt_tokens)
previous_tokens = all_tokens[prompt_reset_since:]
prompt = self.get_prompt(
tokenizer,
previous_tokens,
without_timestamps=options["without_timestamps"],
prefix=options["prefix"],
)
encoder_output = self.encode(features)
result = self.model.generate(
encoder_output,
[prompt] * batch_size,
beam_size=options["beam_size"],
patience=options["patience"],
length_penalty=options["length_penalty"],
max_length=self.max_length,
suppress_blank=options["suppress_blank"],
suppress_tokens=options["suppress_tokens"],
return_scores=True,
return_no_speech_prob=True,
)
output = []
for res in result:
output.append({})
# return scores
seq_len = len(res.sequences_ids[0])
cum_logprob = res.scores[0] * (seq_len ** options["length_penalty"])
output[-1]["avg_logprob"] = cum_logprob / (seq_len + 1)
# return no speech prob
output[-1]["no_speech_prob"] = res.no_speech_prob
output[-1]["tokens"] = res.sequences_ids[0]
return encoder_output, output
def detect_language(self, audio: torch.Tensor):
to_cpu = self.model.device == "cuda" and len(self.model.device_index) > 1
segment = self.feature_extractor(audio, padding=True, to_cpu=to_cpu)[
:, : self.feature_extractor.nb_max_frames
]
encoder_output = self.encode(segment)
results = self.model.detect_language(encoder_output)
language_token, language_probability = results[0][0]
language = language_token[2:-2]
self.logger.info(
f"Detected language: {language} ({language_probability:.2f}) in first 30s of audio..."
)
all_language_probs = [(token[2:-2], prob) for (token, prob) in results[0]]
return language, language_probability, all_language_probs
def detect_language_multi_segment(
self, audio: Union[str, BinaryIO, torch.Tensor], params: Optional[dict] = None
):
"""
Detect language based on N highly-confident segments of a language.
"""
# The threshold is used to decide if the audio is silence or not.
# The default is 0.02 (2.0%) i.e, if more than 2.0% of the audio is silent,
# the audio is considered as silence.
if not params:
params = {
"multilingual": False,
"speech_percentage_threshold": 0.02,
"language_detection_segments": 4,
"vad_filter": True,
"vad_min_silence_duration": 2500,
"language_threshold": 0.7,
}
if params.get("multilingual", False):
logging.warning(
"lang_id is not supported for multilingual audios, detecting the major language."
)
speech_percentage_threshold = params.get("speech_percentage_threshold", 0.02)
language_threshold = params.get("language_threshold", 0.7)
num_detection_segments = params.get("language_detection_segments", 4)
vad_filter_enabled = params.get("vad_filter", True)
vad_params = dict(
min_silence_duration_ms=params.get("vad_min_silence_duration", 2500)
)
if vad_filter_enabled:
vad_params = VadOptions(**vad_params)
# decode audio if it is not decoded already
sampling_rate = self.feature_extractor.sampling_rate
if not isinstance(audio, torch.Tensor):
audio: torch.Tensor = decode_audio(audio, sampling_rate=sampling_rate)
# calculate duration of audio as number of seconds
# audio.shape[0] is the number of samples in the audio
# sampling_rate is the number of samples per second
# if we divide the number of samples by the number of samples per second,
# we get the duration in seconds
duration = audio.shape[0] / sampling_rate
# Check if vad is enabled, and collect voiced segments
if vad_filter_enabled:
# get chunks of audio that contain speech
speech_chunks = get_speech_timestamps(audio, vad_params)
# merge chunks of audio that contain speech into a single array
audio = collect_chunks(audio, speech_chunks)
# calculate new duration of audio without silence
duration_vad = audio.shape[0] / sampling_rate
logging.debug(
f"Lang ID: VAD filter removed {duration - duration_vad} sec of audio"
)
# if the audio after VAD is less than 2% of the original audio, consider it as silence
if duration_vad / duration < speech_percentage_threshold:
return {"language_code": None, "language_confidence": 1.0}
# update duration to be the duration after VAD
duration = duration_vad
# if the duration of the audio is less than 1 second, consider it as silence
if duration < 1.0:
return {"language_code": None, "language_confidence": 1.0}
# number of feature frames in 30 seconds of audio is 3000
nb_max_frames = self.feature_extractor.nb_max_frames
# extract features from audio with padding (default)
to_cpu = self.model.device == "cuda" and len(self.model.device_index) > 1
features = self.feature_extractor(audio, to_cpu=to_cpu)
# number of segments in the audio
num_segments = features.shape[-1] // nb_max_frames
# more number of segments than possible with the duration of file
if num_detection_segments > num_segments:
logging.warning(
f"Lang ID: Can not have more segments, setting {num_segments} segments."
)
num_detection_segments = num_segments
# create a list of indices to randomly select segments from
indices = list(range(num_detection_segments))
# fix seed to get deterministic results
random.seed(0)
random.shuffle(indices)
detected_languages = []
all_language_probabilities = defaultdict(list)
confident_language_probabilities = defaultdict(list)
num_confident_segments_per_language = defaultdict(int)
# Iterate over the randomly selected indices of the segments.
#
# For each segment, extract features and detect language.
#
# If the language is confident, add it to the list of confident segments for that language.
#
# If the number of confident segments for a language
# is greater than or equal to the number of detection segments,
# return the language and the average probability of the language.
#
# If we are unable to get sufficient number of confident predcitions,
# return the most frequently detected language with maximum probability.
#
# We need to get sufficient number of confident predictions per language, not in total.
for i in indices:
segment_features = features[:, i * nb_max_frames : (i + 1) * nb_max_frames]
try:
encoder_output = self.encode(segment_features)
results = self.model.detect_language(encoder_output)[0]
except ValueError as e: # or RuntimeError
logging.error(f"Inference error:{e}")
# results is the list of classes (languages) and their probabilities (descending),
# for eg: [('<|de|>', 0.482177734375),('<|en|>', 0.283447265625),...]
# take top language token and probability
# and parse language token to strip out markers
# for eg: '<|de|>' -> 'de'
language_token = results[0][0]
language = language_token[2:-2]
language_probability = results[0][1]
detected_languages.append(language)
all_language_probabilities[language].append(language_probability)
# only consider if the language prediction is confident
if language_probability > language_threshold:
num_confident_segments_per_language[language] += 1
# Add language and probability to the list of languages when it is confident
confident_language_probabilities[language].append(language_probability)
# return the language when sufficient number of confident segments is achieved
if (
num_confident_segments_per_language[language]
>= num_detection_segments
):
# Considering the average probability of only confident segments
mean = sum(confident_language_probabilities[language]) / len(
confident_language_probabilities[language]
)
return {
"language_code": language,
"language_confidence": mean,
}
# if we are unable to get sufficient number of confident predictions,
# return the most frequently detected language.
# if there is a tie, return the one with maximum average probability.
counter = Counter(detected_languages)
# Define the key function to select frequent language with attached probabilities
def key_func(language):
# Calculate the frequency of the language
frequency = counter[language]
# Calculate the average probability of the language
prob_avg = sum(all_language_probabilities[language]) / len(
all_language_probabilities[language]
)
return frequency, prob_avg
if detected_languages:
# Use the key function to find the language with maximum frequency and probability
max_language = max(detected_languages, key=key_func)
max_probability = sum(all_language_probabilities[max_language]) / len(
all_language_probabilities[max_language]
)
# Do additional checks for silence for non-confident case
# calculate RMS amplitude and DC offset
dc_offset = audio.mean()
audio_minus_dc_offset = audio - dc_offset
is_silent = (
torch.all(audio.abs() < 0.01)
or torch.sqrt(torch.mean(audio_minus_dc_offset**2)) < 0.01
)
if is_silent:
return {"language_code": None, "language_confidence": 1.0}
return {
"language_code": max_language,
"language_confidence": max_probability,
}
# Language is not detected for any segment and none of prev conditions met
return {"language_code": None, "language_confidence": 1.0}
def restore_speech_timestamps(
segments: Iterable[Segment],
speech_chunks: List[dict],
sampling_rate: int,
) -> Iterable[Segment]:
ts_map = SpeechTimestampsMap(speech_chunks, sampling_rate)
for segment in segments:
if segment.words:
words = []
for word in segment.words:
# Ensure the word start and end times are resolved to the same chunk.
middle = (word.start + word.end) / 2
chunk_index = ts_map.get_chunk_index(middle)
word = word._replace(
start=ts_map.get_original_time(word.start, chunk_index),
end=ts_map.get_original_time(word.end, chunk_index),
)
words.append(word)
segment = segment._replace(
start=words[0].start,
end=words[-1].end,
words=words,
)
else:
segment = segment._replace(
start=ts_map.get_original_time(segment.start),
end=ts_map.get_original_time(segment.end),
)
yield segment
def get_ctranslate2_storage(segment: torch.Tensor) -> ctranslate2.StorageView:
segment = segment.contiguous()
segment = ctranslate2.StorageView.from_array(
segment if segment.is_cuda else segment.numpy()
) # torch cpu tensors don't implement __array_interface__
# https://github.com/pytorch/pytorch/issues/51156
return segment
def get_compression_ratio(text: str) -> float:
text_bytes = text.encode("utf-8")
return len(text_bytes) / len(zlib.compress(text_bytes))
def get_suppressed_tokens(
tokenizer: Tokenizer,
suppress_tokens: Tuple[int],
) -> Optional[List[int]]:
if -1 in suppress_tokens:
suppress_tokens = [t for t in suppress_tokens if t >= 0]
suppress_tokens.extend(tokenizer.non_speech_tokens)
elif suppress_tokens is None or len(suppress_tokens) == 0:
suppress_tokens = [] # interpret empty string as an empty list
else:
assert isinstance(suppress_tokens, list), "suppress_tokens must be a list"
suppress_tokens.extend(
[
tokenizer.transcribe,
tokenizer.translate,
tokenizer.sot,
tokenizer.sot_prev,
tokenizer.sot_lm,
]
)
return tuple(sorted(set(suppress_tokens)))
def merge_punctuations(alignment: List[dict], prepended: str, appended: str) -> None:
# merge prepended punctuations
i = len(alignment) - 2
j = len(alignment) - 1
while i >= 0:
previous = alignment[i]
following = alignment[j]
if previous["word"].startswith(" ") and previous["word"].strip() in prepended:
# prepend it to the following word
following["word"] = previous["word"] + following["word"]
if "tokens" in alignment[0].keys():
following["tokens"] = previous["tokens"] + following["tokens"]
previous["tokens"] = []
previous["word"] = ""
else:
j = i
i -= 1
# merge appended punctuations
i = 0
j = 1
while j < len(alignment):
previous = alignment[i]
following = alignment[j]
if not previous["word"].endswith(" ") and following["word"] in appended:
# append it to the previous word
previous["word"] = previous["word"] + following["word"]
if "tokens" in alignment[0].keys():
previous["tokens"] = previous["tokens"] + following["tokens"]
following["tokens"] = []
following["word"] = ""
else:
i = j
j += 1
faster_whisper/utils.py 0 → 100644
View file @ c2eff804
import logging
import os
import re
from typing import List, Optional
import huggingface_hub
import requests
from tqdm.auto import tqdm
_MODELS = {
"tiny.en": "Systran/faster-whisper-tiny.en",
"tiny": "Systran/faster-whisper-tiny",
"base.en": "Systran/faster-whisper-base.en",
"base": "Systran/faster-whisper-base",
"small.en": "Systran/faster-whisper-small.en",
"small": "Systran/faster-whisper-small",
"medium.en": "Systran/faster-whisper-medium.en",
"medium": "Systran/faster-whisper-medium",
"large-v1": "Systran/faster-whisper-large-v1",
"large-v2": "Systran/faster-whisper-large-v2",
"large-v3": "Systran/faster-whisper-large-v3",
"large": "Systran/faster-whisper-large-v3",
"distil-large-v2": "Systran/faster-distil-whisper-large-v2",
"distil-medium.en": "Systran/faster-distil-whisper-medium.en",
"distil-small.en": "Systran/faster-distil-whisper-small.en",
"distil-large-v3": "Systran/faster-distil-whisper-large-v3",
}
def available_models() -> List[str]:
"""Returns the names of available models."""
return list(_MODELS.keys())
def get_assets_path():
"""Returns the path to the assets directory."""
return os.path.join(os.path.dirname(os.path.abspath(__file__)), "assets")
def get_logger():
"""Returns the module logger."""
return logging.getLogger("faster_whisper")
def download_model(
size_or_id: str,
output_dir: Optional[str] = None,
local_files_only: bool = False,
cache_dir: Optional[str] = None,
):
"""Downloads a CTranslate2 Whisper model from the Hugging Face Hub.
Args:
size_or_id: Size of the model to download from https://huggingface.co/Systran
(tiny, tiny.en, base, base.en, small, small.en, distil-small.en, medium, medium.en,
distil-medium.en, large-v1, large-v2, large-v3, large, distil-large-v2,
distil-large-v3), or a CTranslate2-converted model ID from the Hugging Face Hub
(e.g. Systran/faster-whisper-large-v3).
output_dir: Directory where the model should be saved. If not set, the model is saved in
the cache directory.
local_files_only: If True, avoid downloading the file and return the path to the local
cached file if it exists.
cache_dir: Path to the folder where cached files are stored.
Returns:
The path to the downloaded model.
Raises:
ValueError: if the model size is invalid.
"""
if re.match(r".*/.*", size_or_id):
repo_id = size_or_id
else:
repo_id = _MODELS.get(size_or_id)
if repo_id is None:
raise ValueError(
"Invalid model size '%s', expected one of: %s"
% (size_or_id, ", ".join(_MODELS.keys()))
)
allow_patterns = [
"config.json",
"preprocessor_config.json",
"model.bin",
"tokenizer.json",
"vocabulary.*",
]
kwargs = {
"local_files_only": local_files_only,
"allow_patterns": allow_patterns,
"tqdm_class": disabled_tqdm,
}
if output_dir is not None:
kwargs["local_dir"] = output_dir
kwargs["local_dir_use_symlinks"] = False
if cache_dir is not None:
kwargs["cache_dir"] = cache_dir
try:
return huggingface_hub.snapshot_download(repo_id, **kwargs)
except (
huggingface_hub.utils.HfHubHTTPError,
requests.exceptions.ConnectionError,
) as exception:
logger = get_logger()
logger.warning(
"An error occured while synchronizing the model %s from the Hugging Face Hub:\n%s",
repo_id,
exception,
)
logger.warning(
"Trying to load the model directly from the local cache, if it exists."
)
kwargs["local_files_only"] = True
return huggingface_hub.snapshot_download(repo_id, **kwargs)
def format_timestamp(
seconds: float,
always_include_hours: bool = False,
decimal_marker: str = ".",
) -> str:
assert seconds >= 0, "non-negative timestamp expected"
milliseconds = round(seconds * 1000.0)
hours = milliseconds // 3_600_000
milliseconds -= hours * 3_600_000
minutes = milliseconds // 60_000
milliseconds -= minutes * 60_000
seconds = milliseconds // 1_000
milliseconds -= seconds * 1_000
hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
return (
f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
)
class disabled_tqdm(tqdm):
def __init__(self, *args, **kwargs):
kwargs["disable"] = True
super().__init__(*args, **kwargs)
def get_end(segments: List[dict]) -> Optional[float]:
return next(
(w["end"] for s in reversed(segments) for w in reversed(s["words"])),
segments[-1]["end"] if segments else None,
)
faster_whisper/vad.py 0 → 100644
View file @ c2eff804
import bisect
import functools
import os
from abc import ABC
from collections.abc import Callable
from typing import List, NamedTuple, Optional, Union
import numpy as np
import torch
from pyannote.audio.core.io import AudioFile
from pyannote.audio.pipelines import VoiceActivityDetection
from pyannote.audio.pipelines.utils import PipelineModel
from pyannote.core import Annotation, Segment, SlidingWindowFeature
from faster_whisper.utils import get_assets_path
# The code below is adapted from https://github.com/snakers4/silero-vad.
class VadOptions(NamedTuple):
"""VAD options.
Attributes:
threshold: Speech threshold. Silero VAD outputs speech probabilities for each audio chunk,
probabilities ABOVE this value are considered as SPEECH. It is better to tune this
parameter for each dataset separately, but "lazy" 0.5 is pretty good for most datasets.
min_speech_duration_ms: Final speech chunks shorter min_speech_duration_ms are thrown out.
max_speech_duration_s: Maximum duration of speech chunks in seconds. Chunks longer
than max_speech_duration_s will be split at the timestamp of the last silence that
lasts more than 100ms (if any), to prevent aggressive cutting. Otherwise, they will be
split aggressively just before max_speech_duration_s.
min_silence_duration_ms: In the end of each speech chunk wait for min_silence_duration_ms
before separating it
speech_pad_ms: Final speech chunks are padded by speech_pad_ms each side
"""
threshold: float = 0.5
min_speech_duration_ms: int = 250
max_speech_duration_s: float = float("inf")
min_silence_duration_ms: int = 2000
speech_pad_ms: int = 400
def get_speech_timestamps(
audio: torch.Tensor,
vad_options: Optional[VadOptions] = None,
**kwargs,
) -> List[dict]:
"""This method is used for splitting long audios into speech chunks using silero VAD.
Args:
audio: One dimensional float array.
vad_options: Options for VAD processing.
kwargs: VAD options passed as keyword arguments for backward compatibility.
Returns:
List of dicts containing begin and end samples of each speech chunk.
"""
if vad_options is None:
vad_options = VadOptions(**kwargs)
threshold = vad_options.threshold
min_speech_duration_ms = vad_options.min_speech_duration_ms
max_speech_duration_s = vad_options.max_speech_duration_s
min_silence_duration_ms = vad_options.min_silence_duration_ms
window_size_samples = 512
speech_pad_ms = vad_options.speech_pad_ms
sampling_rate = 16000
min_speech_samples = sampling_rate * min_speech_duration_ms / 1000
speech_pad_samples = sampling_rate * speech_pad_ms / 1000
max_speech_samples = (
sampling_rate * max_speech_duration_s
- window_size_samples
- 2 * speech_pad_samples
)
min_silence_samples = sampling_rate * min_silence_duration_ms / 1000
min_silence_samples_at_max_speech = sampling_rate * 98 / 1000
audio_length_samples = len(audio)
model = get_vad_model()
state, context = model.get_initial_states(batch_size=1)
speech_probs = []
for current_start_sample in range(0, audio_length_samples, window_size_samples):
chunk = audio[current_start_sample : current_start_sample + window_size_samples]
if len(chunk) < window_size_samples:
chunk = np.pad(chunk, (0, int(window_size_samples - len(chunk))))
speech_prob, state, context = model(chunk, state, context, sampling_rate)
speech_probs.append(speech_prob)
triggered = False
speeches = []
current_speech = {}
neg_threshold = threshold - 0.15
# to save potential segment end (and tolerate some silence)
temp_end = 0
# to save potential segment limits in case of maximum segment size reached
prev_end = next_start = 0
for i, speech_prob in enumerate(speech_probs):
if (speech_prob >= threshold) and temp_end:
temp_end = 0
if next_start < prev_end:
next_start = window_size_samples * i
if (speech_prob >= threshold) and not triggered:
triggered = True
current_speech["start"] = window_size_samples * i
continue
if (
triggered
and (window_size_samples * i) - current_speech["start"] > max_speech_samples
):
if prev_end:
current_speech["end"] = prev_end
speeches.append(current_speech)
current_speech = {}
# previously reached silence (< neg_thres) and is still not speech (< thres)
if next_start < prev_end:
triggered = False
else:
current_speech["start"] = next_start
prev_end = next_start = temp_end = 0
else:
current_speech["end"] = window_size_samples * i
speeches.append(current_speech)
current_speech = {}
prev_end = next_start = temp_end = 0
triggered = False
continue
if (speech_prob < neg_threshold) and triggered:
if not temp_end:
temp_end = window_size_samples * i
# condition to avoid cutting in very short silence
if (window_size_samples * i) - temp_end > min_silence_samples_at_max_speech:
prev_end = temp_end
if (window_size_samples * i) - temp_end < min_silence_samples:
continue
else:
current_speech["end"] = temp_end
if (
current_speech["end"] - current_speech["start"]
) > min_speech_samples:
speeches.append(current_speech)
current_speech = {}
prev_end = next_start = temp_end = 0
triggered = False
continue
if (
current_speech
and (audio_length_samples - current_speech["start"]) > min_speech_samples
):
current_speech["end"] = audio_length_samples
speeches.append(current_speech)
for i, speech in enumerate(speeches):
if i == 0:
speech["start"] = int(max(0, speech["start"] - speech_pad_samples))
if i != len(speeches) - 1:
silence_duration = speeches[i + 1]["start"] - speech["end"]
if silence_duration < 2 * speech_pad_samples:
speech["end"] += int(silence_duration // 2)
speeches[i + 1]["start"] = int(
max(0, speeches[i + 1]["start"] - silence_duration // 2)
)
else:
speech["end"] = int(
min(audio_length_samples, speech["end"] + speech_pad_samples)
)
speeches[i + 1]["start"] = int(
max(0, speeches[i + 1]["start"] - speech_pad_samples)
)
else:
speech["end"] = int(
min(audio_length_samples, speech["end"] + speech_pad_samples)
)
return speeches
def collect_chunks(audio: torch.Tensor, chunks: List[dict]) -> torch.Tensor:
"""Collects and concatenates audio chunks."""
if not chunks:
return torch.tensor([], dtype=torch.float32)
return torch.cat([audio[chunk["start"] : chunk["end"]] for chunk in chunks])
class SpeechTimestampsMap:
"""Helper class to restore original speech timestamps."""
def __init__(self, chunks: List[dict], sampling_rate: int, time_precision: int = 2):
self.sampling_rate = sampling_rate
self.time_precision = time_precision
self.chunk_end_sample = []
self.total_silence_before = []
previous_end = 0
silent_samples = 0
for chunk in chunks:
silent_samples += chunk["start"] - previous_end
previous_end = chunk["end"]
self.chunk_end_sample.append(chunk["end"] - silent_samples)
self.total_silence_before.append(silent_samples / sampling_rate)
def get_original_time(
self,
time: float,
chunk_index: Optional[int] = None,
) -> float:
if chunk_index is None:
chunk_index = self.get_chunk_index(time)
total_silence_before = self.total_silence_before[chunk_index]
return round(total_silence_before + time, self.time_precision)
def get_chunk_index(self, time: float) -> int:
sample = int(time * self.sampling_rate)
return min(
bisect.bisect(self.chunk_end_sample, sample),
len(self.chunk_end_sample) - 1,
)
@functools.lru_cache
def get_vad_model():
"""Returns the VAD model instance."""
path = os.path.join(get_assets_path(), "silero_vad.onnx")
return SileroVADModel(path)
class SileroVADModel:
def __init__(self, path):
try:
import onnxruntime
except ImportError as e:
raise RuntimeError(
"Applying the VAD filter requires the onnxruntime package"
) from e
opts = onnxruntime.SessionOptions()
opts.inter_op_num_threads = 1
opts.intra_op_num_threads = 1
opts.log_severity_level = 4
self.session = onnxruntime.InferenceSession(
path,
providers=["CPUExecutionProvider"],
sess_options=opts,
)
def get_initial_states(self, batch_size: int):
state = np.zeros((2, batch_size, 128), dtype=np.float32)
context = np.zeros((batch_size, 64), dtype=np.float32)
return state, context
def __call__(self, x, state, context, sr: int):
if len(x.shape) == 1:
x = np.expand_dims(x, 0)
if len(x.shape) > 2:
raise ValueError(
f"Too many dimensions for input audio chunk {len(x.shape)}"
)
if sr / x.shape[1] > 31.25:
raise ValueError("Input audio chunk is too short")
x = np.concatenate([context, x], axis=1)
ort_inputs = {
"input": x,
"state": state,
"sr": np.array(sr, dtype="int64"),
}
out, state = self.session.run(None, ort_inputs)
context = x[..., -64:]
return out, state, context
# BSD 2-Clause License
# Copyright (c) 2024, Max Bain
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
# 2. Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
# The code below is copied from whisper-x (https://github.com/m-bain/whisperX)
# and adapted for faster_whisper.
class SegmentX:
def __init__(self, start, end, speaker=None):
self.start = start
self.end = end
self.speaker = speaker
class VoiceActivitySegmentation(VoiceActivityDetection, ABC):
"""Pipeline wrapper class for Voice Activity Segmentation based on VAD scores."""
def __init__(
self,
segmentation: PipelineModel = "pyannote/segmentation",
device: Optional[Union[str, torch.device]] = None,
fscore: bool = False,
use_auth_token: Optional[str] = None,
**inference_kwargs,
):
"""Initialize the pipeline with the model name and the optional device.
Args:
dict parameters of VoiceActivityDetection class from pyannote:
segmentation (PipelineModel): Loaded model name.
device (torch.device or None): Device to perform the segmentation.
fscore (bool): Flag indicating whether to compute F-score during inference.
use_auth_token (str or None): Optional authentication token for model access.
inference_kwargs (dict): Additional arguments from VoiceActivityDetection pipeline.
"""
super().__init__(
segmentation=segmentation,
device=device,
fscore=fscore,
use_auth_token=use_auth_token,
**inference_kwargs,
)
def apply(
self, file: AudioFile, hook: Optional[Callable] = None
) -> SlidingWindowFeature:
"""Apply voice activity detection on the audio file.
Args:
file (AudioFile): Processed file.
hook (callable): Hook called with signature: hook("step_name", step_artefact, file=file)
Returns:
segmentations (SlidingWindowFeature): Voice activity segmentation.
"""
# setup hook (e.g. for debugging purposes)
hook = self.setup_hook(file, hook=hook)
# apply segmentation model if needed
# output shape is (num_chunks, num_frames, 1)
if self.training:
if self.CACHED_SEGMENTATION in file:
segmentations = file[self.CACHED_SEGMENTATION]
else:
segmentations = self._segmentation(file)
file[self.CACHED_SEGMENTATION] = segmentations
else:
segmentations: SlidingWindowFeature = self._segmentation(file)
return segmentations
class BinarizeVadScores:
"""Binarize detection scores using hysteresis thresholding.
Reference:
Gregory Gelly and Jean-Luc Gauvain. "Minimum Word Error Training of
RNN-based Voice Activity Detection", InterSpeech 2015.
Modified by Max Bain to include WhisperX's min-cut operation
https://arxiv.org/abs/2303.00747
"""
def __init__(
self,
onset: float = 0.5,
offset: Optional[float] = None,
min_duration_on: float = 0.0,
min_duration_off: float = 0.0,
pad_onset: float = 0.0,
pad_offset: float = 0.0,
max_duration: float = float("inf"),
):
"""Initializes the parameters for Binarizing the VAD scores.
Args:
onset (float, optional):
Onset threshold. Defaults to 0.5.
offset (float, optional):
Offset threshold. Defaults to `onset`.
min_duration_on (float, optional):
Remove active regions shorter than that many seconds. Defaults to 0s.
min_duration_off (float, optional):
Fill inactive regions shorter than that many seconds. Defaults to 0s.
pad_onset (float, optional):
Extend active regions by moving their start time by that many seconds.
Defaults to 0s.
pad_offset (float, optional):
Extend active regions by moving their end time by that many seconds.
Defaults to 0s.
max_duration (float):
The maximum length of an active segment.
"""
super().__init__()
self.onset = onset
self.offset = offset or onset
self.pad_onset = pad_onset
self.pad_offset = pad_offset
self.min_duration_on = min_duration_on
self.min_duration_off = min_duration_off
self.max_duration = max_duration
def __get_active_regions(self, scores: SlidingWindowFeature) -> Annotation:
"""Extract active regions from VAD scores.
Args:
scores (SlidingWindowFeature): Detection scores.
Returns:
active (Annotation): Active regions.
"""
num_frames, num_classes = scores.data.shape
frames = scores.sliding_window
timestamps = [frames[i].middle for i in range(num_frames)]
# annotation meant to store 'active' regions
active = Annotation()
for k, k_scores in enumerate(scores.data.T):
label = k if scores.labels is None else scores.labels[k]
# initial state
start = timestamps[0]
is_active = k_scores[0] > self.onset
curr_scores = [k_scores[0]]
curr_timestamps = [start]
t = start
# optionally add `strict=False` for python 3.10 or later
for t, y in zip(timestamps[1:], k_scores[1:]):
# currently active
if is_active:
curr_duration = t - start
if curr_duration > self.max_duration:
search_after = len(curr_scores) // 2
# divide segment
min_score_div_idx = search_after + np.argmin(
curr_scores[search_after:]
)
min_score_t = curr_timestamps[min_score_div_idx]
region = Segment(
start - self.pad_onset, min_score_t + self.pad_offset
)
active[region, k] = label
start = curr_timestamps[min_score_div_idx]
curr_scores = curr_scores[min_score_div_idx + 1 :]
curr_timestamps = curr_timestamps[min_score_div_idx + 1 :]
# switching from active to inactive
elif y < self.offset:
region = Segment(start - self.pad_onset, t + self.pad_offset)
active[region, k] = label
start = t
is_active = False
curr_scores = []
curr_timestamps = []
curr_scores.append(y)
curr_timestamps.append(t)
# currently inactive
else:
# switching from inactive to active
if y > self.onset:
start = t
is_active = True
# if active at the end, add final region
if is_active:
region = Segment(start - self.pad_onset, t + self.pad_offset)
active[region, k] = label
return active
def __call__(self, scores: SlidingWindowFeature) -> Annotation:
"""Binarize detection scores.
Args:
scores (SlidingWindowFeature): Detection scores.
Returns:
active (Annotation): Binarized scores.
"""
active = self.__get_active_regions(scores)
# because of padding, some active regions might be overlapping: merge them.
# also: fill same speaker gaps shorter than min_duration_off
if self.pad_offset > 0.0 or self.pad_onset > 0.0 or self.min_duration_off > 0.0:
if self.max_duration < float("inf"):
raise NotImplementedError("This would break current max_duration param")
active = active.support(collar=self.min_duration_off)
# remove tracks shorter than min_duration_on
if self.min_duration_on > 0:
for segment, track in list(active.itertracks()):
if segment.duration < self.min_duration_on:
del active[segment, track]
return active
def merge_chunks(
segments,
chunk_length,
onset: float = 0.5,
offset: Optional[float] = None,
edge_padding: float = 0.1,
):
"""
Merge operation described in whisper-x paper
"""
curr_end = 0
merged_segments = []
seg_idxs = []
speaker_idxs = []
assert chunk_length > 0
binarize = BinarizeVadScores(max_duration=chunk_length, onset=onset, offset=offset)
segments = binarize(segments)
segments_list = []
for speech_turn in segments.get_timeline():
segments_list.append(
SegmentX(
max(0.0, speech_turn.start - edge_padding),
speech_turn.end + edge_padding,
"UNKNOWN",
)
) # 100ms edge padding to account for edge errors
if len(segments_list) == 0:
print("No active speech found in audio")
return []
# Make sur the starting point is the start of the segment.
curr_start = segments_list[0].start
for idx, seg in enumerate(segments_list):
# if any segment start timing is less than previous segment end timing,
# reset the edge padding. Similarly for end timing.
if idx > 0:
if seg.start < segments_list[idx - 1].end:
seg.start += edge_padding
if idx < len(segments_list) - 1:
if seg.end > segments_list[idx + 1].start:
seg.end -= edge_padding
if seg.end - curr_start > chunk_length and curr_end - curr_start > 0:
merged_segments.append(
{
"start": curr_start,
"end": curr_end,
"segments": seg_idxs,
}
)
curr_start = seg.start
seg_idxs = []
speaker_idxs = []
curr_end = seg.end
seg_idxs.append((seg.start, seg.end))
speaker_idxs.append(seg.speaker)
# add final
merged_segments.append(
{
"start": curr_start,
"end": curr_end,
"segments": seg_idxs,
}
)
return merged_segments
faster_whisper/version.py 0 → 100644
View file @ c2eff804
"""Version information."""
__version__ = "1.0.3"
requirements.conversion.txt 0 → 100644
View file @ c2eff804
transformers[torch]>=4.23
requirements.txt 0 → 100644
View file @ c2eff804
ctranslate2>=4.0,<5
huggingface_hub>=0.13
tokenizers>=0.13,<1
onnxruntime>=1.14,<2
pyannote-audio>=3.1.1
torch>=2.1.1
torchaudio>=2.1.2
tqdm
\ No newline at end of file
setup.cfg 0 → 100644
View file @ c2eff804
[flake8]
max-line-length = 100
ignore =
E203,
W503,
[isort]
profile=black
lines_between_types=1
setup.py 0 → 100644
View file @ c2eff804
import os
from setuptools import find_packages, setup
base_dir = os.path.dirname(os.path.abspath(__file__))
def get_long_description():
readme_path = os.path.join(base_dir, "README.md")
with open(readme_path, encoding="utf-8") as readme_file:
return readme_file.read()
def get_project_version():
version_path = os.path.join(base_dir, "faster_whisper", "version.py")
version = {}
with open(version_path, encoding="utf-8") as fp:
exec(fp.read(), version)
return version["__version__"]
def get_requirements(path):
with open(path, encoding="utf-8") as requirements:
return [requirement.strip() for requirement in requirements]
install_requires = get_requirements(os.path.join(base_dir, "requirements.txt"))
conversion_requires = get_requirements(
os.path.join(base_dir, "requirements.conversion.txt")
)
setup(
name="faster-whisper",
version=get_project_version(),
license="MIT",
description="Faster Whisper transcription with CTranslate2",
long_description=get_long_description(),
long_description_content_type="text/markdown",
author="Guillaume Klein",
url="https://github.com/SYSTRAN/faster-whisper",
classifiers=[
"Development Status :: 4 - Beta",
"Intended Audience :: Developers",
"Intended Audience :: Science/Research",
"License :: OSI Approved :: MIT License",
"Programming Language :: Python :: 3",
"Programming Language :: Python :: 3 :: Only",
"Programming Language :: Python :: 3.8",
"Programming Language :: Python :: 3.9",
"Programming Language :: Python :: 3.10",
"Programming Language :: Python :: 3.11",
"Topic :: Scientific/Engineering :: Artificial Intelligence",
],
keywords="openai whisper speech ctranslate2 inference quantization transformer",
python_requires=">=3.8",
install_requires=install_requires,
extras_require={
"conversion": conversion_requires,
"dev": [
"black==23.*",
"flake8==6.*",
"isort==5.*",
"pytest==7.*",
],
},
packages=find_packages(),
include_package_data=True,
)
tests/conftest.py 0 → 100644
View file @ c2eff804
import os
import pytest
@pytest.fixture
def data_dir():
return os.path.join(os.path.dirname(os.path.abspath(__file__)), "data")
@pytest.fixture
def jfk_path(data_dir):
return os.path.join(data_dir, "jfk.flac")
@pytest.fixture
def physcisworks_path(data_dir):
return os.path.join(data_dir, "physicsworks.wav")
tests/test_transcribe.py 0 → 100644
View file @ c2eff804
import os
from faster_whisper import BatchedInferencePipeline, WhisperModel, decode_audio
from faster_whisper.tokenizer import Tokenizer
from faster_whisper.transcribe import get_suppressed_tokens
def test_supported_languages():
model = WhisperModel("tiny.en")
assert model.supported_languages == ["en"]
def test_transcribe(jfk_path):
model = WhisperModel("tiny")
segments, info = model.transcribe(jfk_path, word_timestamps=True)
assert info.all_language_probs is not None
assert info.language == "en"
assert info.language_probability > 0.9
assert info.duration == 11
# Get top language info from all results, which should match the
# already existing metadata
top_lang, top_lang_score = info.all_language_probs[0]
assert info.language == top_lang
assert abs(info.language_probability - top_lang_score) < 1e-16
segments = list(segments)
assert len(segments) == 1
segment = segments[0]
assert segment.text == (
" And so my fellow Americans ask not what your country can do for you, "
"ask what you can do for your country."
)
assert segment.text == "".join(word.word for word in segment.words)
assert segment.start == segment.words[0].start
assert segment.end == segment.words[-1].end
batched_model = BatchedInferencePipeline(model=model, use_vad_model=False)
result, info = batched_model.transcribe(jfk_path, word_timestamps=True)
assert info.language == "en"
assert info.language_probability > 0.7
segments = []
for segment in result:
segments.append(
{"start": segment.start, "end": segment.end, "text": segment.text}
)
assert len(segments) == 1
assert segment.text == (
" And so my fellow Americans ask not what your country can do for you, "
"ask what you can do for your country."
)
def test_batched_transcribe(physcisworks_path):
model = WhisperModel("tiny")
batched_model = BatchedInferencePipeline(model=model)
result, info = batched_model.transcribe(physcisworks_path, batch_size=16)
assert info.language == "en"
assert info.language_probability > 0.7
segments = []
for segment in result:
segments.append(
{"start": segment.start, "end": segment.end, "text": segment.text}
)
# number of near 30 sec segments
assert len(segments) == 8
result, info = batched_model.transcribe(
physcisworks_path,
batch_size=16,
without_timestamps=False,
word_timestamps=True,
)
segments = []
for segment in result:
assert segment.words is not None
segments.append(
{"start": segment.start, "end": segment.end, "text": segment.text}
)
assert len(segments) > 8
def test_prefix_with_timestamps(jfk_path):
model = WhisperModel("tiny")
segments, _ = model.transcribe(jfk_path, prefix="And so my fellow Americans")
segments = list(segments)
assert len(segments) == 1
segment = segments[0]
assert segment.text == (
" And so my fellow Americans ask not what your country can do for you, "
"ask what you can do for your country."
)
assert segment.start == 0
assert 10 < segment.end < 11
def test_vad(jfk_path):
model = WhisperModel("tiny")
segments, info = model.transcribe(
jfk_path,
vad_filter=True,
vad_parameters=dict(min_silence_duration_ms=500, speech_pad_ms=200),
)
segments = list(segments)
assert len(segments) == 1
segment = segments[0]
assert segment.text == (
" And so my fellow Americans ask not what your country can do for you, "
"ask what you can do for your country."
)
assert 0 < segment.start < 1
assert 10 < segment.end < 11
assert info.vad_options.min_silence_duration_ms == 500
assert info.vad_options.speech_pad_ms == 200
def test_stereo_diarization(data_dir):
model = WhisperModel("tiny")
audio_path = os.path.join(data_dir, "stereo_diarization.wav")
left, right = decode_audio(audio_path, split_stereo=True)
segments, _ = model.transcribe(left)
transcription = "".join(segment.text for segment in segments).strip()
assert transcription == (
"He began a confused complaint against the wizard, "
"who had vanished behind the curtain on the left."
)
segments, _ = model.transcribe(right)
transcription = "".join(segment.text for segment in segments).strip()
assert transcription == "The horizon seems extremely distant."
def test_multisegment_lang_id(physcisworks_path):
model = WhisperModel("tiny")
language_info = model.detect_language_multi_segment(physcisworks_path)
assert language_info["language_code"] == "en"
assert language_info["language_confidence"] > 0.8
def test_suppressed_tokens_minus_1():
model = WhisperModel("tiny.en")
tokenizer = Tokenizer(model.hf_tokenizer, False)
tokens = get_suppressed_tokens(tokenizer, [-1])
assert tokens == (
1,
2,
7,
8,
9,
10,
14,
25,
26,
27,
28,
29,
31,
58,
59,
60,
61,
62,
63,
90,
91,
92,
93,
357,
366,
438,
532,
685,
705,
796,
930,
1058,
1220,
1267,
1279,
1303,
1343,
1377,
1391,
1635,
1782,
1875,
2162,
2361,
2488,
3467,
4008,
4211,
4600,
4808,
5299,
5855,
6329,
7203,
9609,
9959,
10563,
10786,
11420,
11709,
11907,
13163,
13697,
13700,
14808,
15306,
16410,
16791,
17992,
19203,
19510,
20724,
22305,
22935,
27007,
30109,
30420,
33409,
34949,
40283,
40493,
40549,
47282,
49146,
50257,
50357,
50358,
50359,
50360,
)
def test_suppressed_tokens_minus_value():
model = WhisperModel("tiny.en")
tokenizer = Tokenizer(model.hf_tokenizer, False)
tokens = get_suppressed_tokens(tokenizer, [13])
assert tokens == (13, 50257, 50357, 50358, 50359, 50360)
tests/test_utils.py 0 → 100644
View file @ c2eff804
import os
from faster_whisper import available_models, download_model
def test_available_models():
models = available_models()
assert isinstance(models, list)
assert "tiny" in models
def test_download_model(tmpdir):
output_dir = str(tmpdir.join("model"))
model_dir = download_model("tiny", output_dir=output_dir)
assert model_dir == output_dir
assert os.path.isdir(model_dir)
assert not os.path.islink(model_dir)
for filename in os.listdir(model_dir):
path = os.path.join(model_dir, filename)
assert not os.path.islink(path)
def test_download_model_in_cache(tmpdir):
cache_dir = str(tmpdir.join("model"))
download_model("tiny", cache_dir=cache_dir)
assert os.path.isdir(cache_dir)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment