v1.0

model.properties

+# 模型编码
+modelCode=1077
+# 模型名称
+modelName=f5-tts_pytorch
+# 模型描述
+modelDescription=F5-TTS能根据文本内容自动生成带有情感的语音，无论是愤怒、喜悦还是悲伤，都能精准把握情感变化，难辨真伪。
+# 应用场景
+appScenario=推理,语音合成,金融,电商,教育,制造,医疗,能源
+# 框架类型
+frameType=pytorch

openai/whisper-large-v3-turbo/README.md

+---
+language:
+- en
+- zh
+- de
+- es
+- ru
+- ko
+- fr
+- ja
+- pt
+- tr
+- pl
+- ca
+- nl
+- ar
+- sv
+- it
+- id
+- hi
+- fi
+- vi
+- he
+- uk
+- el
+- ms
+- cs
+- ro
+- da
+- hu
+- ta
+- 'no'
+- th
+- ur
+- hr
+- bg
+- lt
+- la
+- mi
+- ml
+- cy
+- sk
+- te
+- fa
+- lv
+- bn
+- sr
+- az
+- sl
+- kn
+- et
+- mk
+- br
+- eu
+- is
+- hy
+- ne
+- mn
+- bs
+- kk
+- sq
+- sw
+- gl
+- mr
+- pa
+- si
+- km
+- sn
+- yo
+- so
+- af
+- oc
+- ka
+- be
+- tg
+- sd
+- gu
+- am
+- yi
+- lo
+- uz
+- fo
+- ht
+- ps
+- tk
+- nn
+- mt
+- sa
+- lb
+- my
+- bo
+- tl
+- mg
+- as
+- tt
+- haw
+- ln
+- ha
+- ba
+- jw
+- su
+license: mit
+tags:
+- audio
+- automatic-speech-recognition
+widget:
+- example_title: Librispeech sample 1
+  src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
+- example_title: Librispeech sample 2
+  src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
+pipeline_tag: automatic-speech-recognition
+base_model:
+- openai/whisper-large-v3
+library_name: transformers
+---
+
+# Whisper
+
+Whisper is a state-of-the-art model for automatic speech recognition (ASR) and speech translation, proposed in the paper 
+[Robust Speech Recognition via Large-Scale Weak Supervision](https://huggingface.co/papers/2212.04356) by Alec Radford 
+et al. from OpenAI. Trained on >5M hours of labeled data, Whisper demonstrates a strong ability to generalise to many 
+datasets and domains in a zero-shot setting.
+
+Whisper large-v3-turbo is a finetuned version of a pruned [Whisper large-v3](https://huggingface.co/openai/whisper-large-v3). In other words, it's the exact same model, except that the number of decoding layers have reduced from 32 to 4.
+As a result, the model is way faster, at the expense of a minor quality degradation. You can find more details about it [in this GitHub discussion](https://github.com/openai/whisper/discussions/2363).
+
+**Disclaimer**: Content for this model card has partly been written by the 🤗 Hugging Face team, and partly copied and 
+pasted from the original model card.
+
+## Usage
+
+Whisper large-v3-turbo is supported in Hugging Face 🤗 Transformers. To run the model, first install the Transformers 
+library. For this example, we'll also install 🤗 Datasets to load toy audio dataset from the Hugging Face Hub, and 
+🤗 Accelerate to reduce the model loading time:
+
+```bash
+pip install --upgrade pip
+pip install --upgrade transformers datasets[audio] accelerate
+```
+
+The model can be used with the [`pipeline`](https://huggingface.co/docs/transformers/main_classes/pipelines#transformers.AutomaticSpeechRecognitionPipeline)
+class to transcribe audios of arbitrary length:
+
+```python
+import torch
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
+from datasets import load_dataset
+
+
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+
+model_id = "openai/whisper-large-v3-turbo"
+
+model = AutoModelForSpeechSeq2Seq.from_pretrained(
+    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
+)
+model.to(device)
+
+processor = AutoProcessor.from_pretrained(model_id)
+
+pipe = pipeline(
+    "automatic-speech-recognition",
+    model=model,
+    tokenizer=processor.tokenizer,
+    feature_extractor=processor.feature_extractor,
+    torch_dtype=torch_dtype,
+    device=device,
+)
+
+dataset = load_dataset("distil-whisper/librispeech_long", "clean", split="validation")
+sample = dataset[0]["audio"]
+
+result = pipe(sample)
+print(result["text"])
+```
+
+To transcribe a local audio file, simply pass the path to your audio file when you call the pipeline:
+
+```python
+result = pipe("audio.mp3")
+```
+
+Multiple audio files can be transcribed in parallel by specifying them as a list and setting the `batch_size` parameter:
+
+```python
+result = pipe(["audio_1.mp3", "audio_2.mp3"], batch_size=2)
+```
+
+Transformers is compatible with all Whisper decoding strategies, such as temperature fallback and condition on previous 
+tokens. The following example demonstrates how to enable these heuristics:
+
+```python
+generate_kwargs = {
+    "max_new_tokens": 448,
+    "num_beams": 1,
+    "condition_on_prev_tokens": False,
+    "compression_ratio_threshold": 1.35,  # zlib compression ratio threshold (in token space)
+    "temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
+    "logprob_threshold": -1.0,
+    "no_speech_threshold": 0.6,
+    "return_timestamps": True,
+}
+
+result = pipe(sample, generate_kwargs=generate_kwargs)
+```
+
+Whisper predicts the language of the source audio automatically. If the source audio language is known *a-priori*, it 
+can be passed as an argument to the pipeline:
+
+```python
+result = pipe(sample, generate_kwargs={"language": "english"})
+```
+
+By default, Whisper performs the task of *speech transcription*, where the source audio language is the same as the target
+text language. To perform *speech translation*, where the target text is in English, set the task to `"translate"`:
+
+```python
+result = pipe(sample, generate_kwargs={"task": "translate"})
+```
+
+Finally, the model can be made to predict timestamps. For sentence-level timestamps, pass the `return_timestamps` argument:
+
+```python
+result = pipe(sample, return_timestamps=True)
+print(result["chunks"])
+```
+
+And for word-level timestamps:
+
+```python
+result = pipe(sample, return_timestamps="word")
+print(result["chunks"])
+```
+
+The above arguments can be used in isolation or in combination. For example, to perform the task of speech transcription 
+where the source audio is in French, and we want to return sentence-level timestamps, the following can be used:
+
+```python
+result = pipe(sample, return_timestamps=True, generate_kwargs={"language": "french", "task": "translate"})
+print(result["chunks"])
+```
+
+<details>
+
+<summary> For more control over the generation parameters, use the model + processor API directly: </summary>
+
+```python
+import torch
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
+from datasets import Audio, load_dataset
+
+
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+
+model_id = "openai/whisper-large-v3-turbo"
+
+model = AutoModelForSpeechSeq2Seq.from_pretrained(
+    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
+)
+model.to(device)
+
+processor = AutoProcessor.from_pretrained(model_id)
+
+dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+dataset = dataset.cast_column("audio", Audio(processor.feature_extractor.sampling_rate))
+sample = dataset[0]["audio"]
+
+inputs = processor(
+    sample["array"],
+    sampling_rate=sample["sampling_rate"],
+    return_tensors="pt",
+    truncation=False,
+    padding="longest",
+    return_attention_mask=True,
+)
+inputs = inputs.to(device, dtype=torch_dtype)
+
+gen_kwargs = {
+    "max_new_tokens": 448,
+    "num_beams": 1,
+    "condition_on_prev_tokens": False,
+    "compression_ratio_threshold": 1.35,  # zlib compression ratio threshold (in token space)
+    "temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
+    "logprob_threshold": -1.0,
+    "no_speech_threshold": 0.6,
+    "return_timestamps": True,
+}
+
+pred_ids = model.generate(**inputs, **gen_kwargs)
+pred_text = processor.batch_decode(pred_ids, skip_special_tokens=True, decode_with_timestamps=False)
+
+print(pred_text)
+```
+
+</details>
+
+## Additional Speed & Memory Improvements
+
+You can apply additional speed and memory improvements to Whisper to further reduce the inference speed and VRAM 
+requirements.
+
+### Chunked Long-Form
+
+Whisper has a receptive field of 30-seconds. To transcribe audios longer than this, one of two long-form algorithms are
+required:
+1. **Sequential:** uses a "sliding window" for buffered inference, transcribing 30-second slices one after the other
+2. **Chunked:** splits long audio files into shorter ones (with a small overlap between segments), transcribes each segment independently, and stitches the resulting transcriptions at the boundaries
+
+The sequential long-form algorithm should be used in either of the following scenarios:
+1. Transcription accuracy is the most important factor, and speed is less of a consideration
+2. You are transcribing **batches** of long audio files, in which case the latency of sequential is comparable to chunked, while being up to 0.5% WER more accurate
+
+Conversely, the chunked algorithm should be used when:
+1. Transcription speed is the most important factor
+2. You are transcribing a **single** long audio file
+
+By default, Transformers uses the sequential algorithm. To enable the chunked algorithm, pass the `chunk_length_s` 
+parameter to the `pipeline`. For large-v3, a chunk length of 30-seconds is optimal. To activate batching over long 
+audio files, pass the argument `batch_size`:
+
+```python
+import torch
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
+from datasets import load_dataset
+
+
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+
+model_id = "openai/whisper-large-v3-turbo"
+
+model = AutoModelForSpeechSeq2Seq.from_pretrained(
+    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
+)
+model.to(device)
+
+processor = AutoProcessor.from_pretrained(model_id)
+
+pipe = pipeline(
+    "automatic-speech-recognition",
+    model=model,
+    tokenizer=processor.tokenizer,
+    feature_extractor=processor.feature_extractor,
+    chunk_length_s=30,
+    batch_size=16,  # batch size for inference - set based on your device
+    torch_dtype=torch_dtype,
+    device=device,
+)
+
+dataset = load_dataset("distil-whisper/librispeech_long", "clean", split="validation")
+sample = dataset[0]["audio"]
+
+result = pipe(sample)
+print(result["text"])
+```
+
+#### Torch compile
+
+The Whisper forward pass is compatible with [`torch.compile`](https://pytorch.org/docs/stable/generated/torch.compile.html)
+for 4.5x speed-ups.
+
+**Note:** `torch.compile` is currently not compatible with the Chunked long-form algorithm or Flash Attention 2 ⚠️
+
+```python
+import torch
+from torch.nn.attention import SDPBackend, sdpa_kernel
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
+from datasets import load_dataset
+from tqdm import tqdm
+
+torch.set_float32_matmul_precision("high")
+
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+
+model_id = "openai/whisper-large-v3-turbo"
+
+model = AutoModelForSpeechSeq2Seq.from_pretrained(
+    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
+).to(device)
+
+# Enable static cache and compile the forward pass
+model.generation_config.cache_implementation = "static"
+model.generation_config.max_new_tokens = 256
+model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
+
+processor = AutoProcessor.from_pretrained(model_id)
+
+pipe = pipeline(
+    "automatic-speech-recognition",
+    model=model,
+    tokenizer=processor.tokenizer,
+    feature_extractor=processor.feature_extractor,
+    torch_dtype=torch_dtype,
+    device=device,
+)
+
+dataset = load_dataset("distil-whisper/librispeech_long", "clean", split="validation")
+sample = dataset[0]["audio"]
+
+# 2 warmup steps
+for _ in tqdm(range(2), desc="Warm-up step"):
+    with sdpa_kernel(SDPBackend.MATH):
+        result = pipe(sample.copy(), generate_kwargs={"min_new_tokens": 256, "max_new_tokens": 256})
+
+# fast run
+with sdpa_kernel(SDPBackend.MATH):
+    result = pipe(sample.copy())
+
+print(result["text"])
+```
+
+#### Flash Attention 2
+
+We recommend using [Flash-Attention 2](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#flashattention-2) if your GPU supports it and you are not using [torch.compile](#torch-compile). 
+To do so, first install [Flash Attention](https://github.com/Dao-AILab/flash-attention):
+
+```
+pip install flash-attn --no-build-isolation
+```
+
+Then pass `attn_implementation="flash_attention_2"` to `from_pretrained`:
+
+```python
+model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, attn_implementation="flash_attention_2")
+```
+
+#### Torch Scale-Product-Attention (SDPA)
+
+If your GPU does not support Flash Attention, we recommend making use of PyTorch [scaled dot-product attention (SDPA)](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html). 
+This attention implementation is activated **by default** for PyTorch versions 2.1.1 or greater. To check 
+whether you have a compatible PyTorch version, run the following Python code snippet:
+
+```python
+from transformers.utils import is_torch_sdpa_available
+
+print(is_torch_sdpa_available())
+```
+
+If the above returns `True`, you have a valid version of PyTorch installed and SDPA is activated by default. If it 
+returns `False`, you need to upgrade your PyTorch version according to the [official instructions](https://pytorch.org/get-started/locally/)
+
+Once a valid PyTorch version is installed, SDPA is activated by default. It can also be set explicitly by specifying 
+`attn_implementation="sdpa"` as follows:
+
+```python
+model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, attn_implementation="sdpa")
+```
+
+For more information about how to use the SDPA refer to the [Transformers SDPA documentation](https://huggingface.co/docs/transformers/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention).
+
+
+## Model details
+
+Whisper is a Transformer based encoder-decoder model, also referred to as a _sequence-to-sequence_ model. There are two
+flavours of Whisper model: English-only and multilingual. The English-only models were trained on the task of English 
+speech recognition. The multilingual models were trained simultaneously on multilingual speech recognition and speech 
+translation. For speech recognition, the model predicts transcriptions in the *same* language as the audio. For speech 
+translation, the model predicts transcriptions to a *different* language to the audio.
+
+Whisper checkpoints come in five configurations of varying model sizes. The smallest four are available as English-only 
+and multilingual. The largest checkpoints are multilingual only. All ten of the pre-trained checkpoints 
+are available on the [Hugging Face Hub](https://huggingface.co/models?search=openai/whisper). The 
+checkpoints are summarised in the following table with links to the models on the Hub:
+
+| Size     | Parameters | English-only                                         | Multilingual                                        |
+|----------|------------|------------------------------------------------------|-----------------------------------------------------|
+| tiny     | 39 M       | [✓](https://huggingface.co/openai/whisper-tiny.en)   | [✓](https://huggingface.co/openai/whisper-tiny)     |
+| base     | 74 M       | [✓](https://huggingface.co/openai/whisper-base.en)   | [✓](https://huggingface.co/openai/whisper-base)     |
+| small    | 244 M      | [✓](https://huggingface.co/openai/whisper-small.en)  | [✓](https://huggingface.co/openai/whisper-small)    |
+| medium   | 769 M      | [✓](https://huggingface.co/openai/whisper-medium.en) | [✓](https://huggingface.co/openai/whisper-medium)   |
+| large    | 1550 M     | x                                                    | [✓](https://huggingface.co/openai/whisper-large)    |
+| large-v2 | 1550 M     | x                                                    | [✓](https://huggingface.co/openai/whisper-large-v2) |
+| large-v3 | 1550 M     | x                                                    | [✓](https://huggingface.co/openai/whisper-large-v3) |
+| large-v3-turbo | 809 M     | x                                                    | [✓](https://huggingface.co/openai/whisper-large-v3-turbo) |
+
+
+## Fine-Tuning
+
+The pre-trained Whisper model demonstrates a strong ability to generalise to different datasets and domains. However, 
+its predictive capabilities can be improved further for certain languages and tasks through *fine-tuning*. The blog 
+post [Fine-Tune Whisper with 🤗 Transformers](https://huggingface.co/blog/fine-tune-whisper) provides a step-by-step 
+guide to fine-tuning the Whisper model with as little as 5 hours of labelled data.
+
+### Evaluated Use
+
+The primary intended users of these models are AI researchers studying robustness, generalization, capabilities, biases, and constraints of the current model. However, Whisper is also potentially quite useful as an ASR solution for developers, especially for English speech recognition. We recognize that once models are released, it is impossible to restrict access to only “intended” uses or to draw reasonable guidelines around what is or is not research.
+
+The models are primarily trained and evaluated on ASR and speech translation to English tasks. They show strong ASR results in ~10 languages. They may exhibit additional capabilities, particularly if fine-tuned on certain tasks like voice activity detection, speaker classification, or speaker diarization but have not been robustly evaluated in these areas. We strongly recommend that users perform robust evaluations of the models in a particular context and domain before deploying them.
+
+In particular, we caution against using Whisper models to transcribe recordings of individuals taken without their consent or purporting to use these models for any kind of subjective classification. We recommend against use in high-risk domains like decision-making contexts, where flaws in accuracy can lead to pronounced flaws in outcomes. The models are intended to transcribe and translate speech, use of the model for classification is not only not evaluated but also not appropriate, particularly to infer human attributes.
+
+
+## Training Data
+
+No information provided.
+
+## Performance and Limitations
+
+Our studies show that, over many existing ASR systems, the models exhibit improved robustness to accents, background noise, technical language, as well as zero shot translation from multiple languages into English; and that accuracy on speech recognition and translation is near the state-of-the-art level. 
+
+However, because the models are trained in a weakly supervised manner using large-scale noisy data, the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination). We hypothesize that this happens because, given their general knowledge of language, the models combine trying to predict the next word in audio with trying to transcribe the audio itself.
+
+Our models perform unevenly across languages, and we observe lower accuracy on low-resource and/or low-discoverability languages or languages where we have less training data. The models also exhibit disparate performance on different accents and dialects of particular languages, which may include higher word error rate across speakers of different genders, races, ages, or other demographic criteria. Our full evaluation results are presented in [the paper accompanying this release](https://cdn.openai.com/papers/whisper.pdf). 
+
+In addition, the sequence-to-sequence architecture of the model makes it prone to generating repetitive texts, which can be mitigated to some degree by beam search and temperature scheduling but not perfectly. Further analysis on these limitations are provided in [the paper](https://cdn.openai.com/papers/whisper.pdf). It is likely that this behavior and hallucinations may be worse on lower-resource and/or lower-discoverability languages.
+
+
+## Broader Implications
+
+We anticipate that Whisper models’ transcription capabilities may be used for improving accessibility tools. While Whisper models cannot be used for real-time transcription out of the box – their speed and size suggest that others may be able to build applications on top of them that allow for near-real-time speech recognition and translation. The real value of beneficial applications built on top of Whisper models suggests that the disparate performance of these models may have real economic implications.
+
+There are also potential dual use concerns that come with releasing Whisper. While we hope the technology will be used primarily for beneficial purposes, making ASR technology more accessible could enable more actors to build capable surveillance technologies or scale up existing surveillance efforts, as the speed and accuracy allow for affordable automatic transcription and translation of large volumes of audio communication. Moreover, these models may have some capabilities to recognize specific individuals out of the box, which in turn presents safety concerns related both to dual use and disparate performance. In practice, we expect that the cost of transcription is not the limiting factor of scaling up surveillance projects.
+
+
+### BibTeX entry and citation info
+```bibtex
+@misc{radford2022whisper,
+  doi = {10.48550/ARXIV.2212.04356},
+  url = {https://arxiv.org/abs/2212.04356},
+  author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
+  title = {Robust Speech Recognition via Large-Scale Weak Supervision},
+  publisher = {arXiv},
+  year = {2022},
+  copyright = {arXiv.org perpetual, non-exclusive license}
+}
+```
\ No newline at end of file

pyproject.toml

+[build-system]
+requires = ["setuptools >= 61.0", "setuptools-scm>=8.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "f5-tts"
+dynamic = ["version"]
+description = "F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching"
+readme = "README.md"
+license = {text = "MIT License"}
+classifiers = [
+    "License :: OSI Approved :: MIT License",
+    "Operating System :: OS Independent",
+    "Programming Language :: Python :: 3",
+]
+dependencies = [
+    "accelerate>=0.33.0",
+    "bitsandbytes>0.37.0",
+    "cached_path",
+    "click",
+    "datasets",
+    "ema_pytorch>=0.5.2",
+    "gradio>=3.45.2",
+    "jieba",
+    "librosa",
+    "matplotlib",
+    "numpy<=1.26.4",
+    "pydub",
+    "pypinyin",
+    "safetensors",
+    "soundfile",
+    "tomli",
+    "torch>=2.0.0",
+    "torchaudio>=2.0.0",
+    "torchdiffeq",
+    "tqdm>=4.65.0",
+    "transformers",
+    "transformers_stream_generator",
+    "vocos",
+    "wandb",
+    "x_transformers>=1.31.14",
+]
+
+[project.optional-dependencies]
+eval = [
+    "faster_whisper==0.10.1",
+    "funasr",
+    "jiwer",
+    "modelscope",
+    "zhconv",
+    "zhon",
+]
+
+[project.urls]
+Homepage = "https://github.com/SWivid/F5-TTS"
+
+[project.scripts]
+"f5-tts_infer-cli" = "f5_tts.infer.infer_cli:main"
+"f5-tts_infer-gradio" = "f5_tts.infer.infer_gradio:main"
+"f5-tts_finetune-cli" = "f5_tts.train.finetune_cli:main"
+"f5-tts_finetune-gradio" = "f5_tts.train.finetune_gradio:main"

ruff.toml

+line-length = 120
+target-version = "py310"
+
+[lint]
+# Only ignore variables with names starting with "_".
+dummy-variable-rgx = "^_.*$"
+
+[lint.isort]
+force-single-line = true
+lines-after-imports = 2

src/f5_tts/api.py

+import random
+import sys
+from importlib.resources import files
+
+import soundfile as sf
+import torch
+import tqdm
+from cached_path import cached_path
+
+from f5_tts.infer.utils_infer import (
+    hop_length,
+    infer_process,
+    load_model,
+    load_vocoder,
+    preprocess_ref_audio_text,
+    remove_silence_for_generated_wav,
+    save_spectrogram,
+    target_sample_rate,
+)
+from f5_tts.model import DiT, UNetT
+from f5_tts.model.utils import seed_everything
+
+
+class F5TTS:
+    def __init__(
+        self,
+        model_type="F5-TTS",
+        ckpt_file="",
+        vocab_file="",
+        ode_method="euler",
+        use_ema=True,
+        vocoder_name="vocos",
+        local_path=None,
+        device=None,
+    ):
+        # Initialize parameters
+        self.final_wave = None
+        self.target_sample_rate = target_sample_rate
+        self.hop_length = hop_length
+        self.seed = -1
+        self.mel_spec_type = vocoder_name
+
+        # Set device
+        self.device = device or (
+            "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+        )
+
+        # Load models
+        self.load_vocoder_model(vocoder_name, local_path)
+        self.load_ema_model(model_type, ckpt_file, vocoder_name, vocab_file, ode_method, use_ema)
+
+    def load_vocoder_model(self, vocoder_name, local_path):
+        self.vocoder = load_vocoder(vocoder_name, local_path is not None, local_path, self.device)
+
+    def load_ema_model(self, model_type, ckpt_file, mel_spec_type, vocab_file, ode_method, use_ema):
+        if model_type == "F5-TTS":
+            if not ckpt_file:
+                if mel_spec_type == "vocos":
+                    ckpt_file = str(cached_path("hf://SWivid/F5-TTS/F5TTS_Base/model_1200000.safetensors"))
+                elif mel_spec_type == "bigvgan":
+                    ckpt_file = str(cached_path("hf://SWivid/F5-TTS/F5TTS_Base_bigvgan/model_1250000.pt"))
+            model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
+            model_cls = DiT
+        elif model_type == "E2-TTS":
+            if not ckpt_file:
+                ckpt_file = str(cached_path("hf://SWivid/E2-TTS/E2TTS_Base/model_1200000.safetensors"))
+            model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
+            model_cls = UNetT
+        else:
+            raise ValueError(f"Unknown model type: {model_type}")
+
+        self.ema_model = load_model(
+            model_cls, model_cfg, ckpt_file, mel_spec_type, vocab_file, ode_method, use_ema, self.device
+        )
+
+    def export_wav(self, wav, file_wave, remove_silence=False):
+        sf.write(file_wave, wav, self.target_sample_rate)
+
+        if remove_silence:
+            remove_silence_for_generated_wav(file_wave)
+
+    def export_spectrogram(self, spect, file_spect):
+        save_spectrogram(spect, file_spect)
+
+    def infer(
+        self,
+        ref_file,
+        ref_text,
+        gen_text,
+        show_info=print,
+        progress=tqdm,
+        target_rms=0.1,
+        cross_fade_duration=0.15,
+        sway_sampling_coef=-1,
+        cfg_strength=2,
+        nfe_step=32,
+        speed=1.0,
+        fix_duration=None,
+        remove_silence=False,
+        file_wave=None,
+        file_spect=None,
+        seed=-1,
+    ):
+        if seed == -1:
+            seed = random.randint(0, sys.maxsize)
+        seed_everything(seed)
+        self.seed = seed
+
+        ref_file, ref_text = preprocess_ref_audio_text(ref_file, ref_text, device=self.device)
+
+        wav, sr, spect = infer_process(
+            ref_file,
+            ref_text,
+            gen_text,
+            self.ema_model,
+            self.vocoder,
+            self.mel_spec_type,
+            show_info=show_info,
+            progress=progress,
+            target_rms=target_rms,
+            cross_fade_duration=cross_fade_duration,
+            nfe_step=nfe_step,
+            cfg_strength=cfg_strength,
+            sway_sampling_coef=sway_sampling_coef,
+            speed=speed,
+            fix_duration=fix_duration,
+            device=self.device,
+        )
+
+        if file_wave is not None:
+            self.export_wav(wav, file_wave, remove_silence)
+
+        if file_spect is not None:
+            self.export_spectrogram(spect, file_spect)
+
+        return wav, sr, spect
+
+
+if __name__ == "__main__":
+    f5tts = F5TTS()
+
+    wav, sr, spect = f5tts.infer(
+        ref_file=str(files("f5_tts").joinpath("infer/examples/basic/basic_ref_en.wav")),
+        ref_text="some call me nature, others call me mother nature.",
+        gen_text="""I don't really care what you call me. I've been a silent spectator, watching species evolve, empires rise and fall. But always remember, I am mighty and enduring. Respect me and I'll nurture you; ignore me and you shall face the consequences.""",
+        file_wave=str(files("f5_tts").joinpath("../../tests/api_out.wav")),
+        file_spect=str(files("f5_tts").joinpath("../../tests/api_out.png")),
+        seed=-1,  # random seed = -1
+    )
+
+    print("seed :", f5tts.seed)

src/f5_tts/eval/README.md

+
+# Evaluation
+
+Install packages for evaluation:
+
+```bash
+pip install -e .[eval]
+```
+
+## Generating Samples for Evaluation
+
+### Prepare Test Datasets
+
+1. *Seed-TTS testset*: Download from [seed-tts-eval](https://github.com/BytedanceSpeech/seed-tts-eval).
+2. *LibriSpeech test-clean*: Download from [OpenSLR](http://www.openslr.org/12/).
+3. Unzip the downloaded datasets and place them in the `data/` directory.
+4. Update the path for *LibriSpeech test-clean* data in `src/f5_tts/eval/eval_infer_batch.py`
+5. Our filtered LibriSpeech-PC 4-10s subset: `data/librispeech_pc_test_clean_cross_sentence.lst`
+
+### Batch Inference for Test Set
+
+To run batch inference for evaluations, execute the following commands:
+
+```bash
+# batch inference for evaluations
+accelerate config  # if not set before
+bash src/f5_tts/eval/eval_infer_batch.sh
+```
+
+## Objective Evaluation on Generated Results
+
+### Download Evaluation Model Checkpoints
+
+1. Chinese ASR Model: [Paraformer-zh](https://huggingface.co/funasr/paraformer-zh)
+2. English ASR Model: [Faster-Whisper](https://huggingface.co/Systran/faster-whisper-large-v3)
+3. WavLM Model: Download from [Google Drive](https://drive.google.com/file/d/1-aE1NfzpRCLxA4GUxX9ITI3F9LlbtEGP/view).
+
+Then update in the following scripts with the paths you put evaluation model ckpts to.
+
+### Objective Evaluation
+
+Update the path with your batch-inferenced results, and carry out WER / SIM evaluations:
+```bash
+# Evaluation for Seed-TTS test set
+python src/f5_tts/eval/eval_seedtts_testset.py
+
+# Evaluation for LibriSpeech-PC test-clean (cross-sentence)
+python src/f5_tts/eval/eval_librispeech_test_clean.py
+```
\ No newline at end of file

src/f5_tts/eval/ecapa_tdnn.py

+# just for speaker similarity evaluation, third-party code
+
+# From https://github.com/microsoft/UniSpeech/blob/main/downstreams/speaker_verification/models/
+# part of the code is borrowed from https://github.com/lawlict/ECAPA-TDNN
+
+import os
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+""" Res2Conv1d + BatchNorm1d + ReLU
+"""
+
+
+class Res2Conv1dReluBn(nn.Module):
+    """
+    in_channels == out_channels == channels
+    """
+
+    def __init__(self, channels, kernel_size=1, stride=1, padding=0, dilation=1, bias=True, scale=4):
+        super().__init__()
+        assert channels % scale == 0, "{} % {} != 0".format(channels, scale)
+        self.scale = scale
+        self.width = channels // scale
+        self.nums = scale if scale == 1 else scale - 1
+
+        self.convs = []
+        self.bns = []
+        for i in range(self.nums):
+            self.convs.append(nn.Conv1d(self.width, self.width, kernel_size, stride, padding, dilation, bias=bias))
+            self.bns.append(nn.BatchNorm1d(self.width))
+        self.convs = nn.ModuleList(self.convs)
+        self.bns = nn.ModuleList(self.bns)
+
+    def forward(self, x):
+        out = []
+        spx = torch.split(x, self.width, 1)
+        for i in range(self.nums):
+            if i == 0:
+                sp = spx[i]
+            else:
+                sp = sp + spx[i]
+            # Order: conv -> relu -> bn
+            sp = self.convs[i](sp)
+            sp = self.bns[i](F.relu(sp))
+            out.append(sp)
+        if self.scale != 1:
+            out.append(spx[self.nums])
+        out = torch.cat(out, dim=1)
+
+        return out
+
+
+""" Conv1d + BatchNorm1d + ReLU
+"""
+
+
+class Conv1dReluBn(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size=1, stride=1, padding=0, dilation=1, bias=True):
+        super().__init__()
+        self.conv = nn.Conv1d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias=bias)
+        self.bn = nn.BatchNorm1d(out_channels)
+
+    def forward(self, x):
+        return self.bn(F.relu(self.conv(x)))
+
+
+""" The SE connection of 1D case.
+"""
+
+
+class SE_Connect(nn.Module):
+    def __init__(self, channels, se_bottleneck_dim=128):
+        super().__init__()
+        self.linear1 = nn.Linear(channels, se_bottleneck_dim)
+        self.linear2 = nn.Linear(se_bottleneck_dim, channels)
+
+    def forward(self, x):
+        out = x.mean(dim=2)
+        out = F.relu(self.linear1(out))
+        out = torch.sigmoid(self.linear2(out))
+        out = x * out.unsqueeze(2)
+
+        return out
+
+
+""" SE-Res2Block of the ECAPA-TDNN architecture.
+"""
+
+# def SE_Res2Block(channels, kernel_size, stride, padding, dilation, scale):
+#     return nn.Sequential(
+#         Conv1dReluBn(channels, 512, kernel_size=1, stride=1, padding=0),
+#         Res2Conv1dReluBn(512, kernel_size, stride, padding, dilation, scale=scale),
+#         Conv1dReluBn(512, channels, kernel_size=1, stride=1, padding=0),
+#         SE_Connect(channels)
+#     )
+
+
+class SE_Res2Block(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation, scale, se_bottleneck_dim):
+        super().__init__()
+        self.Conv1dReluBn1 = Conv1dReluBn(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+        self.Res2Conv1dReluBn = Res2Conv1dReluBn(out_channels, kernel_size, stride, padding, dilation, scale=scale)
+        self.Conv1dReluBn2 = Conv1dReluBn(out_channels, out_channels, kernel_size=1, stride=1, padding=0)
+        self.SE_Connect = SE_Connect(out_channels, se_bottleneck_dim)
+
+        self.shortcut = None
+        if in_channels != out_channels:
+            self.shortcut = nn.Conv1d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=1,
+            )
+
+    def forward(self, x):
+        residual = x
+        if self.shortcut:
+            residual = self.shortcut(x)
+
+        x = self.Conv1dReluBn1(x)
+        x = self.Res2Conv1dReluBn(x)
+        x = self.Conv1dReluBn2(x)
+        x = self.SE_Connect(x)
+
+        return x + residual
+
+
+""" Attentive weighted mean and standard deviation pooling.
+"""
+
+
+class AttentiveStatsPool(nn.Module):
+    def __init__(self, in_dim, attention_channels=128, global_context_att=False):
+        super().__init__()
+        self.global_context_att = global_context_att
+
+        # Use Conv1d with stride == 1 rather than Linear, then we don't need to transpose inputs.
+        if global_context_att:
+            self.linear1 = nn.Conv1d(in_dim * 3, attention_channels, kernel_size=1)  # equals W and b in the paper
+        else:
+            self.linear1 = nn.Conv1d(in_dim, attention_channels, kernel_size=1)  # equals W and b in the paper
+        self.linear2 = nn.Conv1d(attention_channels, in_dim, kernel_size=1)  # equals V and k in the paper
+
+    def forward(self, x):
+        if self.global_context_att:
+            context_mean = torch.mean(x, dim=-1, keepdim=True).expand_as(x)
+            context_std = torch.sqrt(torch.var(x, dim=-1, keepdim=True) + 1e-10).expand_as(x)
+            x_in = torch.cat((x, context_mean, context_std), dim=1)
+        else:
+            x_in = x
+
+        # DON'T use ReLU here! In experiments, I find ReLU hard to converge.
+        alpha = torch.tanh(self.linear1(x_in))
+        # alpha = F.relu(self.linear1(x_in))
+        alpha = torch.softmax(self.linear2(alpha), dim=2)
+        mean = torch.sum(alpha * x, dim=2)
+        residuals = torch.sum(alpha * (x**2), dim=2) - mean**2
+        std = torch.sqrt(residuals.clamp(min=1e-9))
+        return torch.cat([mean, std], dim=1)
+
+
+class ECAPA_TDNN(nn.Module):
+    def __init__(
+        self,
+        feat_dim=80,
+        channels=512,
+        emb_dim=192,
+        global_context_att=False,
+        feat_type="wavlm_large",
+        sr=16000,
+        feature_selection="hidden_states",
+        update_extract=False,
+        config_path=None,
+    ):
+        super().__init__()
+
+        self.feat_type = feat_type
+        self.feature_selection = feature_selection
+        self.update_extract = update_extract
+        self.sr = sr
+
+        torch.hub._validate_not_a_forked_repo = lambda a, b, c: True
+        try:
+            local_s3prl_path = os.path.expanduser("~/.cache/torch/hub/s3prl_s3prl_main")
+            self.feature_extract = torch.hub.load(local_s3prl_path, feat_type, source="local", config_path=config_path)
+        except:  # noqa: E722
+            self.feature_extract = torch.hub.load("s3prl/s3prl", feat_type)
+
+        if len(self.feature_extract.model.encoder.layers) == 24 and hasattr(
+            self.feature_extract.model.encoder.layers[23].self_attn, "fp32_attention"
+        ):
+            self.feature_extract.model.encoder.layers[23].self_attn.fp32_attention = False
+        if len(self.feature_extract.model.encoder.layers) == 24 and hasattr(
+            self.feature_extract.model.encoder.layers[11].self_attn, "fp32_attention"
+        ):
+            self.feature_extract.model.encoder.layers[11].self_attn.fp32_attention = False
+
+        self.feat_num = self.get_feat_num()
+        self.feature_weight = nn.Parameter(torch.zeros(self.feat_num))
+
+        if feat_type != "fbank" and feat_type != "mfcc":
+            freeze_list = ["final_proj", "label_embs_concat", "mask_emb", "project_q", "quantizer"]
+            for name, param in self.feature_extract.named_parameters():
+                for freeze_val in freeze_list:
+                    if freeze_val in name:
+                        param.requires_grad = False
+                        break
+
+        if not self.update_extract:
+            for param in self.feature_extract.parameters():
+                param.requires_grad = False
+
+        self.instance_norm = nn.InstanceNorm1d(feat_dim)
+        # self.channels = [channels] * 4 + [channels * 3]
+        self.channels = [channels] * 4 + [1536]
+
+        self.layer1 = Conv1dReluBn(feat_dim, self.channels[0], kernel_size=5, padding=2)
+        self.layer2 = SE_Res2Block(
+            self.channels[0],
+            self.channels[1],
+            kernel_size=3,
+            stride=1,
+            padding=2,
+            dilation=2,
+            scale=8,
+            se_bottleneck_dim=128,
+        )
+        self.layer3 = SE_Res2Block(
+            self.channels[1],
+            self.channels[2],
+            kernel_size=3,
+            stride=1,
+            padding=3,
+            dilation=3,
+            scale=8,
+            se_bottleneck_dim=128,
+        )
+        self.layer4 = SE_Res2Block(
+            self.channels[2],
+            self.channels[3],
+            kernel_size=3,
+            stride=1,
+            padding=4,
+            dilation=4,
+            scale=8,
+            se_bottleneck_dim=128,
+        )
+
+        # self.conv = nn.Conv1d(self.channels[-1], self.channels[-1], kernel_size=1)
+        cat_channels = channels * 3
+        self.conv = nn.Conv1d(cat_channels, self.channels[-1], kernel_size=1)
+        self.pooling = AttentiveStatsPool(
+            self.channels[-1], attention_channels=128, global_context_att=global_context_att
+        )
+        self.bn = nn.BatchNorm1d(self.channels[-1] * 2)
+        self.linear = nn.Linear(self.channels[-1] * 2, emb_dim)
+
+    def get_feat_num(self):
+        self.feature_extract.eval()
+        wav = [torch.randn(self.sr).to(next(self.feature_extract.parameters()).device)]
+        with torch.no_grad():
+            features = self.feature_extract(wav)
+        select_feature = features[self.feature_selection]
+        if isinstance(select_feature, (list, tuple)):
+            return len(select_feature)
+        else:
+            return 1
+
+    def get_feat(self, x):
+        if self.update_extract:
+            x = self.feature_extract([sample for sample in x])
+        else:
+            with torch.no_grad():
+                if self.feat_type == "fbank" or self.feat_type == "mfcc":
+                    x = self.feature_extract(x) + 1e-6  # B x feat_dim x time_len
+                else:
+                    x = self.feature_extract([sample for sample in x])
+
+        if self.feat_type == "fbank":
+            x = x.log()
+
+        if self.feat_type != "fbank" and self.feat_type != "mfcc":
+            x = x[self.feature_selection]
+            if isinstance(x, (list, tuple)):
+                x = torch.stack(x, dim=0)
+            else:
+                x = x.unsqueeze(0)
+            norm_weights = F.softmax(self.feature_weight, dim=-1).unsqueeze(-1).unsqueeze(-1).unsqueeze(-1)
+            x = (norm_weights * x).sum(dim=0)
+            x = torch.transpose(x, 1, 2) + 1e-6
+
+        x = self.instance_norm(x)
+        return x
+
+    def forward(self, x):
+        x = self.get_feat(x)
+
+        out1 = self.layer1(x)
+        out2 = self.layer2(out1)
+        out3 = self.layer3(out2)
+        out4 = self.layer4(out3)
+
+        out = torch.cat([out2, out3, out4], dim=1)
+        out = F.relu(self.conv(out))
+        out = self.bn(self.pooling(out))
+        out = self.linear(out)
+
+        return out
+
+
+def ECAPA_TDNN_SMALL(
+    feat_dim,
+    emb_dim=256,
+    feat_type="wavlm_large",
+    sr=16000,
+    feature_selection="hidden_states",
+    update_extract=False,
+    config_path=None,
+):
+    return ECAPA_TDNN(
+        feat_dim=feat_dim,
+        channels=512,
+        emb_dim=emb_dim,
+        feat_type=feat_type,
+        sr=sr,
+        feature_selection=feature_selection,
+        update_extract=update_extract,
+        config_path=config_path,
+    )

src/f5_tts/eval/eval_infer_batch.py

+import os
+import sys
+
+sys.path.append(os.getcwd())
+
+import argparse
+import time
+from importlib.resources import files
+
+import torch
+import torchaudio
+from accelerate import Accelerator
+from tqdm import tqdm
+
+from f5_tts.eval.utils_eval import (
+    get_inference_prompt,
+    get_librispeech_test_clean_metainfo,
+    get_seedtts_testset_metainfo,
+)
+from f5_tts.infer.utils_infer import load_checkpoint, load_vocoder
+from f5_tts.model import CFM, DiT, UNetT
+from f5_tts.model.utils import get_tokenizer
+
+accelerator = Accelerator()
+device = f"cuda:{accelerator.process_index}"
+
+
+# --------------------- Dataset Settings -------------------- #
+
+target_sample_rate = 24000
+n_mel_channels = 100
+hop_length = 256
+win_length = 1024
+n_fft = 1024
+target_rms = 0.1
+
+
+tokenizer = "pinyin"
+rel_path = str(files("f5_tts").joinpath("../../"))
+
+
+def main():
+    # ---------------------- infer setting ---------------------- #
+
+    parser = argparse.ArgumentParser(description="batch inference")
+
+    parser.add_argument("-s", "--seed", default=None, type=int)
+    parser.add_argument("-d", "--dataset", default="Emilia_ZH_EN")
+    parser.add_argument("-n", "--expname", required=True)
+    parser.add_argument("-c", "--ckptstep", default=1200000, type=int)
+    parser.add_argument("-m", "--mel_spec_type", default="vocos", type=str, choices=["bigvgan", "vocos"])
+
+    parser.add_argument("-nfe", "--nfestep", default=32, type=int)
+    parser.add_argument("-o", "--odemethod", default="euler")
+    parser.add_argument("-ss", "--swaysampling", default=-1, type=float)
+
+    parser.add_argument("-t", "--testset", required=True)
+
+    args = parser.parse_args()
+
+    seed = args.seed
+    dataset_name = args.dataset
+    exp_name = args.expname
+    ckpt_step = args.ckptstep
+    ckpt_path = rel_path + f"/ckpts/{exp_name}/model_{ckpt_step}.pt"
+    mel_spec_type = args.mel_spec_type
+
+    nfe_step = args.nfestep
+    ode_method = args.odemethod
+    sway_sampling_coef = args.swaysampling
+
+    testset = args.testset
+
+    infer_batch_size = 1  # max frames. 1 for ddp single inference (recommended)
+    cfg_strength = 2.0
+    speed = 1.0
+    use_truth_duration = False
+    no_ref_audio = False
+
+    if exp_name == "F5TTS_Base":
+        model_cls = DiT
+        model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
+
+    elif exp_name == "E2TTS_Base":
+        model_cls = UNetT
+        model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
+
+    if testset == "ls_pc_test_clean":
+        metalst = rel_path + "/data/librispeech_pc_test_clean_cross_sentence.lst"
+        librispeech_test_clean_path = "<SOME_PATH>/LibriSpeech/test-clean"  # test-clean path
+        metainfo = get_librispeech_test_clean_metainfo(metalst, librispeech_test_clean_path)
+
+    elif testset == "seedtts_test_zh":
+        metalst = rel_path + "/data/seedtts_testset/zh/meta.lst"
+        metainfo = get_seedtts_testset_metainfo(metalst)
+
+    elif testset == "seedtts_test_en":
+        metalst = rel_path + "/data/seedtts_testset/en/meta.lst"
+        metainfo = get_seedtts_testset_metainfo(metalst)
+
+    # path to save genereted wavs
+    output_dir = (
+        f"{rel_path}/"
+        f"results/{exp_name}_{ckpt_step}/{testset}/"
+        f"seed{seed}_{ode_method}_nfe{nfe_step}_{mel_spec_type}"
+        f"{f'_ss{sway_sampling_coef}' if sway_sampling_coef else ''}"
+        f"_cfg{cfg_strength}_speed{speed}"
+        f"{'_gt-dur' if use_truth_duration else ''}"
+        f"{'_no-ref-audio' if no_ref_audio else ''}"
+    )
+
+    # -------------------------------------------------#
+
+    use_ema = True
+
+    prompts_all = get_inference_prompt(
+        metainfo,
+        speed=speed,
+        tokenizer=tokenizer,
+        target_sample_rate=target_sample_rate,
+        n_mel_channels=n_mel_channels,
+        hop_length=hop_length,
+        mel_spec_type=mel_spec_type,
+        target_rms=target_rms,
+        use_truth_duration=use_truth_duration,
+        infer_batch_size=infer_batch_size,
+    )
+
+    # Vocoder model
+    local = False
+    if mel_spec_type == "vocos":
+        vocoder_local_path = "../checkpoints/charactr/vocos-mel-24khz"
+    elif mel_spec_type == "bigvgan":
+        vocoder_local_path = "../checkpoints/bigvgan_v2_24khz_100band_256x"
+    vocoder = load_vocoder(vocoder_name=mel_spec_type, is_local=local, local_path=vocoder_local_path)
+
+    # Tokenizer
+    vocab_char_map, vocab_size = get_tokenizer(dataset_name, tokenizer)
+
+    # Model
+    model = CFM(
+        transformer=model_cls(**model_cfg, text_num_embeds=vocab_size, mel_dim=n_mel_channels),
+        mel_spec_kwargs=dict(
+            n_fft=n_fft,
+            hop_length=hop_length,
+            win_length=win_length,
+            n_mel_channels=n_mel_channels,
+            target_sample_rate=target_sample_rate,
+            mel_spec_type=mel_spec_type,
+        ),
+        odeint_kwargs=dict(
+            method=ode_method,
+        ),
+        vocab_char_map=vocab_char_map,
+    ).to(device)
+
+    dtype = torch.float32 if mel_spec_type == "bigvgan" else None
+    model = load_checkpoint(model, ckpt_path, device, dtype=dtype, use_ema=use_ema)
+
+    if not os.path.exists(output_dir) and accelerator.is_main_process:
+        os.makedirs(output_dir)
+
+    # start batch inference
+    accelerator.wait_for_everyone()
+    start = time.time()
+
+    with accelerator.split_between_processes(prompts_all) as prompts:
+        for prompt in tqdm(prompts, disable=not accelerator.is_local_main_process):
+            utts, ref_rms_list, ref_mels, ref_mel_lens, total_mel_lens, final_text_list = prompt
+            ref_mels = ref_mels.to(device)
+            ref_mel_lens = torch.tensor(ref_mel_lens, dtype=torch.long).to(device)
+            total_mel_lens = torch.tensor(total_mel_lens, dtype=torch.long).to(device)
+
+            # Inference
+            with torch.inference_mode():
+                generated, _ = model.sample(
+                    cond=ref_mels,
+                    text=final_text_list,
+                    duration=total_mel_lens,
+                    lens=ref_mel_lens,
+                    steps=nfe_step,
+                    cfg_strength=cfg_strength,
+                    sway_sampling_coef=sway_sampling_coef,
+                    no_ref_audio=no_ref_audio,
+                    seed=seed,
+                )
+                # Final result
+                for i, gen in enumerate(generated):
+                    gen = gen[ref_mel_lens[i] : total_mel_lens[i], :].unsqueeze(0)
+                    gen_mel_spec = gen.permute(0, 2, 1)
+                    if mel_spec_type == "vocos":
+                        generated_wave = vocoder.decode(gen_mel_spec)
+                    elif mel_spec_type == "bigvgan":
+                        generated_wave = vocoder(gen_mel_spec)
+
+                    if ref_rms_list[i] < target_rms:
+                        generated_wave = generated_wave * ref_rms_list[i] / target_rms
+                    torchaudio.save(f"{output_dir}/{utts[i]}.wav", generated_wave.squeeze(0).cpu(), target_sample_rate)
+
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        timediff = time.time() - start
+        print(f"Done batch inference in {timediff / 60 :.2f} minutes.")
+
+
+if __name__ == "__main__":
+    main()

src/f5_tts/eval/eval_infer_batch.sh

+#!/bin/bash
+
+# e.g. F5-TTS, 16 NFE
+accelerate launch src/f5_tts/eval/eval_infer_batch.py -s 0 -n "F5TTS_Base" -t "seedtts_test_zh" -nfe 16
+accelerate launch src/f5_tts/eval/eval_infer_batch.py -s 0 -n "F5TTS_Base" -t "seedtts_test_en" -nfe 16
+accelerate launch src/f5_tts/eval/eval_infer_batch.py -s 0 -n "F5TTS_Base" -t "ls_pc_test_clean" -nfe 16
+
+# e.g. Vanilla E2 TTS, 32 NFE
+accelerate launch src/f5_tts/eval/eval_infer_batch.py -s 0 -n "E2TTS_Base" -t "seedtts_test_zh" -o "midpoint" -ss 0
+accelerate launch src/f5_tts/eval/eval_infer_batch.py -s 0 -n "E2TTS_Base" -t "seedtts_test_en" -o "midpoint" -ss 0
+accelerate launch src/f5_tts/eval/eval_infer_batch.py -s 0 -n "E2TTS_Base" -t "ls_pc_test_clean" -o "midpoint" -ss 0
+
+# etc.

src/f5_tts/eval/eval_librispeech_test_clean.py

+# Evaluate with Librispeech test-clean, ~3s prompt to generate 4-10s audio (the way of valle/voicebox evaluation)
+
+import sys
+import os
+
+sys.path.append(os.getcwd())
+
+import multiprocessing as mp
+from importlib.resources import files
+
+import numpy as np
+
+from f5_tts.eval.utils_eval import (
+    get_librispeech_test,
+    run_asr_wer,
+    run_sim,
+)
+
+rel_path = str(files("f5_tts").joinpath("../../"))
+
+
+eval_task = "wer"  # sim | wer
+lang = "en"
+metalst = rel_path + "/data/librispeech_pc_test_clean_cross_sentence.lst"
+librispeech_test_clean_path = "<SOME_PATH>/LibriSpeech/test-clean"  # test-clean path
+gen_wav_dir = "PATH_TO_GENERATED"  # generated wavs
+
+gpus = [0, 1, 2, 3, 4, 5, 6, 7]
+test_set = get_librispeech_test(metalst, gen_wav_dir, gpus, librispeech_test_clean_path)
+
+## In LibriSpeech, some speakers utilized varying voice characteristics for different characters in the book,
+## leading to a low similarity for the ground truth in some cases.
+# test_set = get_librispeech_test(metalst, gen_wav_dir, gpus, librispeech_test_clean_path, eval_ground_truth = True)  # eval ground truth
+
+local = False
+if local:  # use local custom checkpoint dir
+    asr_ckpt_dir = "../checkpoints/Systran/faster-whisper-large-v3"
+else:
+    asr_ckpt_dir = ""  # auto download to cache dir
+
+wavlm_ckpt_dir = "../checkpoints/UniSpeech/wavlm_large_finetune.pth"
+
+
+# --------------------------- WER ---------------------------
+
+if eval_task == "wer":
+    wers = []
+
+    with mp.Pool(processes=len(gpus)) as pool:
+        args = [(rank, lang, sub_test_set, asr_ckpt_dir) for (rank, sub_test_set) in test_set]
+        results = pool.map(run_asr_wer, args)
+        for wers_ in results:
+            wers.extend(wers_)
+
+    wer = round(np.mean(wers) * 100, 3)
+    print(f"\nTotal {len(wers)} samples")
+    print(f"WER      : {wer}%")
+
+
+# --------------------------- SIM ---------------------------
+
+if eval_task == "sim":
+    sim_list = []
+
+    with mp.Pool(processes=len(gpus)) as pool:
+        args = [(rank, sub_test_set, wavlm_ckpt_dir) for (rank, sub_test_set) in test_set]
+        results = pool.map(run_sim, args)
+        for sim_ in results:
+            sim_list.extend(sim_)
+
+    sim = round(sum(sim_list) / len(sim_list), 3)
+    print(f"\nTotal {len(sim_list)} samples")
+    print(f"SIM      : {sim}")

src/f5_tts/eval/eval_seedtts_testset.py

+# Evaluate with Seed-TTS testset
+
+import sys
+import os
+
+sys.path.append(os.getcwd())
+
+import multiprocessing as mp
+from importlib.resources import files
+
+import numpy as np
+
+from f5_tts.eval.utils_eval import (
+    get_seed_tts_test,
+    run_asr_wer,
+    run_sim,
+)
+
+rel_path = str(files("f5_tts").joinpath("../../"))
+
+
+eval_task = "wer"  # sim | wer
+lang = "zh"  # zh | en
+metalst = rel_path + f"/data/seedtts_testset/{lang}/meta.lst"  # seed-tts testset
+# gen_wav_dir = rel_path + f"/data/seedtts_testset/{lang}/wavs"  # ground truth wavs
+gen_wav_dir = "PATH_TO_GENERATED"  # generated wavs
+
+
+# NOTE. paraformer-zh result will be slightly different according to the number of gpus, cuz batchsize is different
+#       zh 1.254 seems a result of 4 workers wer_seed_tts
+gpus = [0, 1, 2, 3, 4, 5, 6, 7]
+test_set = get_seed_tts_test(metalst, gen_wav_dir, gpus)
+
+local = False
+if local:  # use local custom checkpoint dir
+    if lang == "zh":
+        asr_ckpt_dir = "../checkpoints/funasr"  # paraformer-zh dir under funasr
+    elif lang == "en":
+        asr_ckpt_dir = "../checkpoints/Systran/faster-whisper-large-v3"
+else:
+    asr_ckpt_dir = ""  # auto download to cache dir
+
+wavlm_ckpt_dir = "../checkpoints/UniSpeech/wavlm_large_finetune.pth"
+
+
+# --------------------------- WER ---------------------------
+
+if eval_task == "wer":
+    wers = []
+
+    with mp.Pool(processes=len(gpus)) as pool:
+        args = [(rank, lang, sub_test_set, asr_ckpt_dir) for (rank, sub_test_set) in test_set]
+        results = pool.map(run_asr_wer, args)
+        for wers_ in results:
+            wers.extend(wers_)
+
+    wer = round(np.mean(wers) * 100, 3)
+    print(f"\nTotal {len(wers)} samples")
+    print(f"WER      : {wer}%")
+
+
+# --------------------------- SIM ---------------------------
+
+if eval_task == "sim":
+    sim_list = []
+
+    with mp.Pool(processes=len(gpus)) as pool:
+        args = [(rank, sub_test_set, wavlm_ckpt_dir) for (rank, sub_test_set) in test_set]
+        results = pool.map(run_sim, args)
+        for sim_ in results:
+            sim_list.extend(sim_)
+
+    sim = round(sum(sim_list) / len(sim_list), 3)
+    print(f"\nTotal {len(sim_list)} samples")
+    print(f"SIM      : {sim}")

src/f5_tts/eval/utils_eval.py

+import math
+import os
+import random
+import string
+
+import torch
+import torch.nn.functional as F
+import torchaudio
+from tqdm import tqdm
+
+from f5_tts.eval.ecapa_tdnn import ECAPA_TDNN_SMALL
+from f5_tts.model.modules import MelSpec
+from f5_tts.model.utils import convert_char_to_pinyin
+
+
+# seedtts testset metainfo: utt, prompt_text, prompt_wav, gt_text, gt_wav
+def get_seedtts_testset_metainfo(metalst):
+    f = open(metalst)
+    lines = f.readlines()
+    f.close()
+    metainfo = []
+    for line in lines:
+        if len(line.strip().split("|")) == 5:
+            utt, prompt_text, prompt_wav, gt_text, gt_wav = line.strip().split("|")
+        elif len(line.strip().split("|")) == 4:
+            utt, prompt_text, prompt_wav, gt_text = line.strip().split("|")
+            gt_wav = os.path.join(os.path.dirname(metalst), "wavs", utt + ".wav")
+        if not os.path.isabs(prompt_wav):
+            prompt_wav = os.path.join(os.path.dirname(metalst), prompt_wav)
+        metainfo.append((utt, prompt_text, prompt_wav, gt_text, gt_wav))
+    return metainfo
+
+
+# librispeech test-clean metainfo: gen_utt, ref_txt, ref_wav, gen_txt, gen_wav
+def get_librispeech_test_clean_metainfo(metalst, librispeech_test_clean_path):
+    f = open(metalst)
+    lines = f.readlines()
+    f.close()
+    metainfo = []
+    for line in lines:
+        ref_utt, ref_dur, ref_txt, gen_utt, gen_dur, gen_txt = line.strip().split("\t")
+
+        # ref_txt = ref_txt[0] + ref_txt[1:].lower() + '.'  # if use librispeech test-clean (no-pc)
+        ref_spk_id, ref_chaptr_id, _ = ref_utt.split("-")
+        ref_wav = os.path.join(librispeech_test_clean_path, ref_spk_id, ref_chaptr_id, ref_utt + ".flac")
+
+        # gen_txt = gen_txt[0] + gen_txt[1:].lower() + '.'  # if use librispeech test-clean (no-pc)
+        gen_spk_id, gen_chaptr_id, _ = gen_utt.split("-")
+        gen_wav = os.path.join(librispeech_test_clean_path, gen_spk_id, gen_chaptr_id, gen_utt + ".flac")
+
+        metainfo.append((gen_utt, ref_txt, ref_wav, " " + gen_txt, gen_wav))
+
+    return metainfo
+
+
+# padded to max length mel batch
+def padded_mel_batch(ref_mels):
+    max_mel_length = torch.LongTensor([mel.shape[-1] for mel in ref_mels]).amax()
+    padded_ref_mels = []
+    for mel in ref_mels:
+        padded_ref_mel = F.pad(mel, (0, max_mel_length - mel.shape[-1]), value=0)
+        padded_ref_mels.append(padded_ref_mel)
+    padded_ref_mels = torch.stack(padded_ref_mels)
+    padded_ref_mels = padded_ref_mels.permute(0, 2, 1)
+    return padded_ref_mels
+
+
+# get prompts from metainfo containing: utt, prompt_text, prompt_wav, gt_text, gt_wav
+
+
+def get_inference_prompt(
+    metainfo,
+    speed=1.0,
+    tokenizer="pinyin",
+    polyphone=True,
+    target_sample_rate=24000,
+    n_fft=1024,
+    win_length=1024,
+    n_mel_channels=100,
+    hop_length=256,
+    mel_spec_type="vocos",
+    target_rms=0.1,
+    use_truth_duration=False,
+    infer_batch_size=1,
+    num_buckets=200,
+    min_secs=3,
+    max_secs=40,
+):
+    prompts_all = []
+
+    min_tokens = min_secs * target_sample_rate // hop_length
+    max_tokens = max_secs * target_sample_rate // hop_length
+
+    batch_accum = [0] * num_buckets
+    utts, ref_rms_list, ref_mels, ref_mel_lens, total_mel_lens, final_text_list = (
+        [[] for _ in range(num_buckets)] for _ in range(6)
+    )
+
+    mel_spectrogram = MelSpec(
+        n_fft=n_fft,
+        hop_length=hop_length,
+        win_length=win_length,
+        n_mel_channels=n_mel_channels,
+        target_sample_rate=target_sample_rate,
+        mel_spec_type=mel_spec_type,
+    )
+
+    for utt, prompt_text, prompt_wav, gt_text, gt_wav in tqdm(metainfo, desc="Processing prompts..."):
+        # Audio
+        ref_audio, ref_sr = torchaudio.load(prompt_wav)
+        ref_rms = torch.sqrt(torch.mean(torch.square(ref_audio)))
+        if ref_rms < target_rms:
+            ref_audio = ref_audio * target_rms / ref_rms
+        assert ref_audio.shape[-1] > 5000, f"Empty prompt wav: {prompt_wav}, or torchaudio backend issue."
+        if ref_sr != target_sample_rate:
+            resampler = torchaudio.transforms.Resample(ref_sr, target_sample_rate)
+            ref_audio = resampler(ref_audio)
+
+        # Text
+        if len(prompt_text[-1].encode("utf-8")) == 1:
+            prompt_text = prompt_text + " "
+        text = [prompt_text + gt_text]
+        if tokenizer == "pinyin":
+            text_list = convert_char_to_pinyin(text, polyphone=polyphone)
+        else:
+            text_list = text
+
+        # Duration, mel frame length
+        ref_mel_len = ref_audio.shape[-1] // hop_length
+        if use_truth_duration:
+            gt_audio, gt_sr = torchaudio.load(gt_wav)
+            if gt_sr != target_sample_rate:
+                resampler = torchaudio.transforms.Resample(gt_sr, target_sample_rate)
+                gt_audio = resampler(gt_audio)
+            total_mel_len = ref_mel_len + int(gt_audio.shape[-1] / hop_length / speed)
+
+            # # test vocoder resynthesis
+            # ref_audio = gt_audio
+        else:
+            ref_text_len = len(prompt_text.encode("utf-8"))
+            gen_text_len = len(gt_text.encode("utf-8"))
+            total_mel_len = ref_mel_len + int(ref_mel_len / ref_text_len * gen_text_len / speed)
+
+        # to mel spectrogram
+        ref_mel = mel_spectrogram(ref_audio)
+        ref_mel = ref_mel.squeeze(0)
+
+        # deal with batch
+        assert infer_batch_size > 0, "infer_batch_size should be greater than 0."
+        assert (
+            min_tokens <= total_mel_len <= max_tokens
+        ), f"Audio {utt} has duration {total_mel_len*hop_length//target_sample_rate}s out of range [{min_secs}, {max_secs}]."
+        bucket_i = math.floor((total_mel_len - min_tokens) / (max_tokens - min_tokens + 1) * num_buckets)
+
+        utts[bucket_i].append(utt)
+        ref_rms_list[bucket_i].append(ref_rms)
+        ref_mels[bucket_i].append(ref_mel)
+        ref_mel_lens[bucket_i].append(ref_mel_len)
+        total_mel_lens[bucket_i].append(total_mel_len)
+        final_text_list[bucket_i].extend(text_list)
+
+        batch_accum[bucket_i] += total_mel_len
+
+        if batch_accum[bucket_i] >= infer_batch_size:
+            # print(f"\n{len(ref_mels[bucket_i][0][0])}\n{ref_mel_lens[bucket_i]}\n{total_mel_lens[bucket_i]}")
+            prompts_all.append(
+                (
+                    utts[bucket_i],
+                    ref_rms_list[bucket_i],
+                    padded_mel_batch(ref_mels[bucket_i]),
+                    ref_mel_lens[bucket_i],
+                    total_mel_lens[bucket_i],
+                    final_text_list[bucket_i],
+                )
+            )
+            batch_accum[bucket_i] = 0
+            (
+                utts[bucket_i],
+                ref_rms_list[bucket_i],
+                ref_mels[bucket_i],
+                ref_mel_lens[bucket_i],
+                total_mel_lens[bucket_i],
+                final_text_list[bucket_i],
+            ) = [], [], [], [], [], []
+
+    # add residual
+    for bucket_i, bucket_frames in enumerate(batch_accum):
+        if bucket_frames > 0:
+            prompts_all.append(
+                (
+                    utts[bucket_i],
+                    ref_rms_list[bucket_i],
+                    padded_mel_batch(ref_mels[bucket_i]),
+                    ref_mel_lens[bucket_i],
+                    total_mel_lens[bucket_i],
+                    final_text_list[bucket_i],
+                )
+            )
+    # not only leave easy work for last workers
+    random.seed(666)
+    random.shuffle(prompts_all)
+
+    return prompts_all
+
+
+# get wav_res_ref_text of seed-tts test metalst
+# https://github.com/BytedanceSpeech/seed-tts-eval
+
+
+def get_seed_tts_test(metalst, gen_wav_dir, gpus):
+    f = open(metalst)
+    lines = f.readlines()
+    f.close()
+
+    test_set_ = []
+    for line in tqdm(lines):
+        if len(line.strip().split("|")) == 5:
+            utt, prompt_text, prompt_wav, gt_text, gt_wav = line.strip().split("|")
+        elif len(line.strip().split("|")) == 4:
+            utt, prompt_text, prompt_wav, gt_text = line.strip().split("|")
+
+        if not os.path.exists(os.path.join(gen_wav_dir, utt + ".wav")):
+            continue
+        gen_wav = os.path.join(gen_wav_dir, utt + ".wav")
+        if not os.path.isabs(prompt_wav):
+            prompt_wav = os.path.join(os.path.dirname(metalst), prompt_wav)
+
+        test_set_.append((gen_wav, prompt_wav, gt_text))
+
+    num_jobs = len(gpus)
+    if num_jobs == 1:
+        return [(gpus[0], test_set_)]
+
+    wav_per_job = len(test_set_) // num_jobs + 1
+    test_set = []
+    for i in range(num_jobs):
+        test_set.append((gpus[i], test_set_[i * wav_per_job : (i + 1) * wav_per_job]))
+
+    return test_set
+
+
+# get librispeech test-clean cross sentence test
+
+
+def get_librispeech_test(metalst, gen_wav_dir, gpus, librispeech_test_clean_path, eval_ground_truth=False):
+    f = open(metalst)
+    lines = f.readlines()
+    f.close()
+
+    test_set_ = []
+    for line in tqdm(lines):
+        ref_utt, ref_dur, ref_txt, gen_utt, gen_dur, gen_txt = line.strip().split("\t")
+
+        if eval_ground_truth:
+            gen_spk_id, gen_chaptr_id, _ = gen_utt.split("-")
+            gen_wav = os.path.join(librispeech_test_clean_path, gen_spk_id, gen_chaptr_id, gen_utt + ".flac")
+        else:
+            if not os.path.exists(os.path.join(gen_wav_dir, gen_utt + ".wav")):
+                raise FileNotFoundError(f"Generated wav not found: {gen_utt}")
+            gen_wav = os.path.join(gen_wav_dir, gen_utt + ".wav")
+
+        ref_spk_id, ref_chaptr_id, _ = ref_utt.split("-")
+        ref_wav = os.path.join(librispeech_test_clean_path, ref_spk_id, ref_chaptr_id, ref_utt + ".flac")
+
+        test_set_.append((gen_wav, ref_wav, gen_txt))
+
+    num_jobs = len(gpus)
+    if num_jobs == 1:
+        return [(gpus[0], test_set_)]
+
+    wav_per_job = len(test_set_) // num_jobs + 1
+    test_set = []
+    for i in range(num_jobs):
+        test_set.append((gpus[i], test_set_[i * wav_per_job : (i + 1) * wav_per_job]))
+
+    return test_set
+
+
+# load asr model
+
+
+def load_asr_model(lang, ckpt_dir=""):
+    if lang == "zh":
+        from funasr import AutoModel
+
+        model = AutoModel(
+            model=os.path.join(ckpt_dir, "paraformer-zh"),
+            # vad_model = os.path.join(ckpt_dir, "fsmn-vad"),
+            # punc_model = os.path.join(ckpt_dir, "ct-punc"),
+            # spk_model = os.path.join(ckpt_dir, "cam++"),
+            disable_update=True,
+        )  # following seed-tts setting
+    elif lang == "en":
+        from faster_whisper import WhisperModel
+
+        model_size = "large-v3" if ckpt_dir == "" else ckpt_dir
+        model = WhisperModel(model_size, device="cuda", compute_type="float16")
+    return model
+
+
+# WER Evaluation, the way Seed-TTS does
+
+
+def run_asr_wer(args):
+    rank, lang, test_set, ckpt_dir = args
+
+    if lang == "zh":
+        import zhconv
+
+        torch.cuda.set_device(rank)
+    elif lang == "en":
+        os.environ["CUDA_VISIBLE_DEVICES"] = str(rank)
+    else:
+        raise NotImplementedError(
+            "lang support only 'zh' (funasr paraformer-zh), 'en' (faster-whisper-large-v3), for now."
+        )
+
+    asr_model = load_asr_model(lang, ckpt_dir=ckpt_dir)
+
+    from zhon.hanzi import punctuation
+
+    punctuation_all = punctuation + string.punctuation
+    wers = []
+
+    from jiwer import compute_measures
+
+    for gen_wav, prompt_wav, truth in tqdm(test_set):
+        if lang == "zh":
+            res = asr_model.generate(input=gen_wav, batch_size_s=300, disable_pbar=True)
+            hypo = res[0]["text"]
+            hypo = zhconv.convert(hypo, "zh-cn")
+        elif lang == "en":
+            segments, _ = asr_model.transcribe(gen_wav, beam_size=5, language="en")
+            hypo = ""
+            for segment in segments:
+                hypo = hypo + " " + segment.text
+
+        # raw_truth = truth
+        # raw_hypo = hypo
+
+        for x in punctuation_all:
+            truth = truth.replace(x, "")
+            hypo = hypo.replace(x, "")
+
+        truth = truth.replace("  ", " ")
+        hypo = hypo.replace("  ", " ")
+
+        if lang == "zh":
+            truth = " ".join([x for x in truth])
+            hypo = " ".join([x for x in hypo])
+        elif lang == "en":
+            truth = truth.lower()
+            hypo = hypo.lower()
+
+        measures = compute_measures(truth, hypo)
+        wer = measures["wer"]
+
+        # ref_list = truth.split(" ")
+        # subs = measures["substitutions"] / len(ref_list)
+        # dele = measures["deletions"] / len(ref_list)
+        # inse = measures["insertions"] / len(ref_list)
+
+        wers.append(wer)
+
+    return wers
+
+
+# SIM Evaluation
+
+
+def run_sim(args):
+    rank, test_set, ckpt_dir = args
+    device = f"cuda:{rank}"
+
+    model = ECAPA_TDNN_SMALL(feat_dim=1024, feat_type="wavlm_large", config_path=None)
+    state_dict = torch.load(ckpt_dir, weights_only=True, map_location=lambda storage, loc: storage)
+    model.load_state_dict(state_dict["model"], strict=False)
+
+    use_gpu = True if torch.cuda.is_available() else False
+    if use_gpu:
+        model = model.cuda(device)
+    model.eval()
+
+    sim_list = []
+    for wav1, wav2, truth in tqdm(test_set):
+        wav1, sr1 = torchaudio.load(wav1)
+        wav2, sr2 = torchaudio.load(wav2)
+
+        resample1 = torchaudio.transforms.Resample(orig_freq=sr1, new_freq=16000)
+        resample2 = torchaudio.transforms.Resample(orig_freq=sr2, new_freq=16000)
+        wav1 = resample1(wav1)
+        wav2 = resample2(wav2)
+
+        if use_gpu:
+            wav1 = wav1.cuda(device)
+            wav2 = wav2.cuda(device)
+        with torch.no_grad():
+            emb1 = model(wav1)
+            emb2 = model(wav2)
+
+        sim = F.cosine_similarity(emb1, emb2)[0].item()
+        # print(f"VSim score between two audios: {sim:.4f} (-1.0, 1.0).")
+        sim_list.append(sim)
+
+    return sim_list

src/f5_tts/infer/README.md

+# Inference
+
+The pretrained model checkpoints can be reached at [🤗 Hugging Face](https://huggingface.co/SWivid/F5-TTS) and [🤖 Model Scope](https://www.modelscope.cn/models/SWivid/F5-TTS_Emilia-ZH-EN), or will be automatically downloaded when running inference scripts.
+
+Currently support **30s for a single** generation, which is the **total length** including both prompt and output audio. However, you can provide `infer_cli` and `infer_gradio` with longer text, will automatically do chunk generation. Long reference audio will be **clip short to ~15s**.
+
+To avoid possible inference failures, make sure you have seen through the following instructions.
+
+- Use reference audio <15s and leave some silence (e.g. 1s) at the end. Otherwise there is a risk of truncating in the middle of word, leading to suboptimal generation.
+- Uppercased letters will be uttered letter by letter, so use lowercased letters for normal words. 
+- Add some spaces (blank: " ") or punctuations (e.g. "," ".") to explicitly introduce some pauses.
+- Preprocess numbers to Chinese letters if you want to have them read in Chinese, otherwise in English.
+
+
+## Gradio App
+
+Currently supported features:
+
+- Basic TTS with Chunk Inference
+- Multi-Style / Multi-Speaker Generation
+- Voice Chat powered by Qwen2.5-3B-Instruct
+
+The cli command `f5-tts_infer-gradio` equals to `python src/f5_tts/infer/infer_gradio.py`, which launches a Gradio APP (web interface) for inference.
+
+The script will load model checkpoints from Huggingface. You can also manually download files and update the path to `load_model()` in `infer_gradio.py`. Currently only load TTS models first, will load ASR model to do transcription if `ref_text` not provided, will load LLM model if use Voice Chat.
+
+Could also be used as a component for larger application.
+```python
+import gradio as gr
+from f5_tts.infer.infer_gradio import app
+
+with gr.Blocks() as main_app:
+    gr.Markdown("# This is an example of using F5-TTS within a bigger Gradio app")
+
+    # ... other Gradio components
+
+    app.render()
+
+main_app.launch()
+```
+
+
+## CLI Inference
+
+The cli command `f5-tts_infer-cli` equals to `python src/f5_tts/infer/infer_cli.py`, which is a command line tool for inference.
+
+The script will load model checkpoints from Huggingface. You can also manually download files and use `--ckpt_file` to specify the model you want to load, or directly update in `infer_cli.py`.
+
+For change vocab.txt use `--vocab_file` to provide your `vocab.txt` file.
+
+Basically you can inference with flags:
+```bash
+# Leave --ref_text "" will have ASR model transcribe (extra GPU memory usage)
+f5-tts_infer-cli \
+--model "F5-TTS" \
+--ref_audio "ref_audio.wav" \
+--ref_text "The content, subtitle or transcription of reference audio." \
+--gen_text "Some text you want TTS model generate for you."
+
+# Choose Vocoder
+f5-tts_infer-cli --vocoder_name bigvgan --load_vocoder_from_local --ckpt_file <YOUR_CKPT_PATH, eg:ckpts/F5TTS_Base_bigvgan/model_1250000.pt>
+f5-tts_infer-cli --vocoder_name vocos --load_vocoder_from_local --ckpt_file <YOUR_CKPT_PATH, eg:ckpts/F5TTS_Base/model_1200000.safetensors>
+```
+
+And a `.toml` file would help with more flexible usage.
+
+```bash
+f5-tts_infer-cli -c custom.toml
+```
+
+For example, you can use `.toml` to pass in variables, refer to `src/f5_tts/infer/examples/basic/basic.toml`:
+
+```toml
+# F5-TTS | E2-TTS
+model = "F5-TTS"
+ref_audio = "infer/examples/basic/basic_ref_en.wav"
+# If an empty "", transcribes the reference audio automatically.
+ref_text = "Some call me nature, others call me mother nature."
+gen_text = "I don't really care what you call me. I've been a silent spectator, watching species evolve, empires rise and fall. But always remember, I am mighty and enduring."
+# File with text to generate. Ignores the text above.
+gen_file = ""
+remove_silence = false
+output_dir = "tests"
+```
+
+You can also leverage `.toml` file to do multi-style generation, refer to `src/f5_tts/infer/examples/multi/story.toml`.
+
+```toml
+# F5-TTS | E2-TTS
+model = "F5-TTS"
+ref_audio = "infer/examples/multi/main.flac"
+# If an empty "", transcribes the reference audio automatically.
+ref_text = ""
+gen_text = ""
+# File with text to generate. Ignores the text above.
+gen_file = "infer/examples/multi/story.txt"
+remove_silence = true
+output_dir = "tests"
+
+[voices.town]
+ref_audio = "infer/examples/multi/town.flac"
+ref_text = ""
+
+[voices.country]
+ref_audio = "infer/examples/multi/country.flac"
+ref_text = ""
+```
+You should mark the voice with `[main]` `[town]` `[country]` whenever you want to change voice, refer to `src/f5_tts/infer/examples/multi/story.txt`.
+
+## Speech Editing
+
+To test speech editing capabilities, use the following command:
+
+```bash
+python src/f5_tts/infer/speech_edit.py
+```
+
+## Socket Realtime Client
+
+To communicate with socket server you need to run 
+```bash
+python src/f5_tts/socket_server.py
+```
+
+<details>
+<summary>Then create client to communicate</summary>
+
+``` python
+import socket
+import numpy as np
+import asyncio
+import pyaudio
+
+async def listen_to_voice(text, server_ip='localhost', server_port=9999):
+    client_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    client_socket.connect((server_ip, server_port))
+
+    async def play_audio_stream():
+        buffer = b''
+        p = pyaudio.PyAudio()
+        stream = p.open(format=pyaudio.paFloat32,
+                        channels=1,
+                        rate=24000,  # Ensure this matches the server's sampling rate
+                        output=True,
+                        frames_per_buffer=2048)
+
+        try:
+            while True:
+                chunk = await asyncio.get_event_loop().run_in_executor(None, client_socket.recv, 1024)
+                if not chunk:  # End of stream
+                    break
+                if b"END_OF_AUDIO" in chunk:
+                    buffer += chunk.replace(b"END_OF_AUDIO", b"")
+                    if buffer:
+                        audio_array = np.frombuffer(buffer, dtype=np.float32).copy()  # Make a writable copy
+                        stream.write(audio_array.tobytes())
+                    break
+                buffer += chunk
+                if len(buffer) >= 4096:
+                    audio_array = np.frombuffer(buffer[:4096], dtype=np.float32).copy()  # Make a writable copy
+                    stream.write(audio_array.tobytes())
+                    buffer = buffer[4096:]
+        finally:
+            stream.stop_stream()
+            stream.close()
+            p.terminate()
+
+    try:
+        # Send only the text to the server
+        await asyncio.get_event_loop().run_in_executor(None, client_socket.sendall, text.encode('utf-8'))
+        await play_audio_stream()
+        print("Audio playback finished.")
+
+    except Exception as e:
+        print(f"Error in listen_to_voice: {e}")
+
+    finally:
+        client_socket.close()
+
+# Example usage: Replace this with your actual server IP and port
+async def main():
+    await listen_to_voice("my name is jenny..", server_ip='localhost', server_port=9998)
+
+# Run the main async function
+asyncio.run(main())
+```
+
+</details>
+

src/f5_tts/infer/examples/basic/basic.toml

+# F5-TTS | E2-TTS
+model = "F5-TTS"
+ref_audio = "infer/examples/basic/basic_ref_en.wav"
+# If an empty "", transcribes the reference audio automatically.
+ref_text = "Some call me nature, others call me mother nature."
+gen_text = "I don't really care what you call me. I've been a silent spectator, watching species evolve, empires rise and fall. But always remember, I am mighty and enduring."
+# File with text to generate. Ignores the text above.
+gen_file = ""
+remove_silence = false
+output_dir = "tests"
\ No newline at end of file

src/f5_tts/infer/examples/multi/story.toml

+# F5-TTS | E2-TTS
+model = "F5-TTS"
+ref_audio = "infer/examples/multi/main.flac"
+# If an empty "", transcribes the reference audio automatically.
+ref_text = ""
+gen_text = ""
+# File with text to generate. Ignores the text above.
+gen_file = "infer/examples/multi/story.txt"
+remove_silence = true
+output_dir = "tests"
+
+[voices.town]
+ref_audio = "infer/examples/multi/town.flac"
+ref_text = ""
+
+[voices.country]
+ref_audio = "infer/examples/multi/country.flac"
+ref_text = ""
+

src/f5_tts/infer/examples/multi/story.txt

+A Town Mouse and a Country Mouse were acquaintances, and the Country Mouse one day invited his friend to come and see him at his home in the fields. The Town Mouse came, and they sat down to a dinner of barleycorns and roots, the latter of which had a distinctly earthy flavour. The fare was not much to the taste of the guest, and presently he broke out with [town] “My poor dear friend, you live here no better than the ants. Now, you should just see how I fare! My larder is a regular horn of plenty. You must come and stay with me, and I promise you you shall live on the fat of the land.” [main] So when he returned to town he took the Country Mouse with him, and showed him into a larder containing flour and oatmeal and figs and honey and dates. The Country Mouse had never seen anything like it, and sat down to enjoy the luxuries his friend provided: but before they had well begun, the door of the larder opened and someone came in. The two Mice scampered off and hid themselves in a narrow and exceedingly uncomfortable hole. Presently, when all was quiet, they ventured out again; but someone else came in, and off they scuttled again. This was too much for the visitor. [country] “Goodbye,” [main] said he, [country] “I’m off. You live in the lap of luxury, I can see, but you are surrounded by dangers; whereas at home I can enjoy my simple dinner of roots and corn in peace.”
\ No newline at end of file