inference.py 10.4 KB
Newer Older
1
2
3
4
5
6
7
8
"""
Text-to-speech pipeline using Tacotron2.
"""

import argparse
import os
import random
import sys
9
from functools import partial
10

11
import numpy as np
12
13
14
15
16
17
18
19
20
import torch
import torchaudio
from datasets import InverseSpectralNormalization
from text.text_preprocessing import (
    available_symbol_set,
    available_phonemizers,
    get_symbol_list,
    text_to_sequence,
)
21
22
23
from torchaudio.models import Tacotron2
from torchaudio.models import tacotron2 as pretrained_tacotron2
from utils import prepare_input_sequence
24
25
26
27
28
29


def parse_args():
    r"""
    Parse commandline arguments.
    """
30
    from torchaudio.models.tacotron2 import _MODEL_CONFIG_AND_URLS as tacotron2_config_and_urls
31
32
33
34
    from torchaudio.models.wavernn import _MODEL_CONFIG_AND_URLS as wavernn_config_and_urls

    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
35
        "--checkpoint-name",
36
37
38
        type=str,
        default=None,
        choices=list(tacotron2_config_and_urls.keys()),
39
        help="[string] The name of the checkpoint to load.",
40
    )
41
42
    parser.add_argument("--checkpoint-path", type=str, default=None, help="[string] Path to the checkpoint file.")
    parser.add_argument("--output-path", type=str, default="./audio.wav", help="[string] Path to the output .wav file.")
43
    parser.add_argument(
44
45
        "--input-text",
        "-i",
46
47
        type=str,
        default="Hello world",
48
        help="[string] Type in something here and TTS will generate it!",
49
50
    )
    parser.add_argument(
51
52
53
        "--vocoder",
        default="nvidia_waveglow",
        choices=["griffin_lim", "wavernn", "nvidia_waveglow"],
54
55
56
57
        type=str,
        help="Select the vocoder to use.",
    )
    parser.add_argument(
58
        "--jit", default=False, action="store_true", help="If used, the model and inference function is jitted."
59
60
    )

61
    preprocessor = parser.add_argument_group("text preprocessor setup")
62
    preprocessor.add_argument(
63
64
        "--text-preprocessor",
        default="english_characters",
65
66
        type=str,
        choices=available_symbol_set,
67
        help="select text preprocessor to use.",
68
69
    )
    preprocessor.add_argument(
70
        "--phonemizer",
71
72
73
        default="DeepPhonemizer",
        type=str,
        choices=available_phonemizers,
74
        help='select phonemizer to use, only used when text-preprocessor is "english_phonemes"',
75
76
    )
    preprocessor.add_argument(
77
        "--phonemizer-checkpoint",
78
79
        default="./en_us_cmudict_forward.pt",
        type=str,
80
81
        help="the path or name of the checkpoint for the phonemizer, "
        'only used when text-preprocessor is "english_phonemes"',
82
83
    )
    preprocessor.add_argument(
84
        "--cmudict-root", default="./", type=str, help="the root directory for storing CMU dictionary files"
85
86
    )

87
88
89
90
91
92
    audio = parser.add_argument_group("audio parameters")
    audio.add_argument("--sample-rate", default=22050, type=int, help="Sampling rate")
    audio.add_argument("--n-fft", default=1024, type=int, help="Filter length for STFT")
    audio.add_argument("--n-mels", default=80, type=int, help="")
    audio.add_argument("--mel-fmin", default=0.0, type=float, help="Minimum mel frequency")
    audio.add_argument("--mel-fmax", default=8000.0, type=float, help="Maximum mel frequency")
93
94

    # parameters for WaveRNN
95
    wavernn = parser.add_argument_group("WaveRNN parameters")
96
    wavernn.add_argument(
97
        "--wavernn-checkpoint-name",
98
99
        default="wavernn_10k_epochs_8bits_ljspeech",
        choices=list(wavernn_config_and_urls.keys()),
100
        help="Select the WaveRNN checkpoint.",
101
102
103
104
105
106
107
108
109
110
111
112
    )
    wavernn.add_argument(
        "--wavernn-loss",
        default="crossentropy",
        choices=["crossentropy"],
        type=str,
        help="The type of loss the WaveRNN pretrained model is trained on.",
    )
    wavernn.add_argument(
        "--wavernn-no-batch-inference",
        default=False,
        action="store_true",
113
        help="Don't use batch inference for WaveRNN inference.",
114
115
    )
    wavernn.add_argument(
116
        "--wavernn-no-mulaw", default=False, action="store_true", help="Don't use mulaw decoder to decode the signal."
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
    )
    wavernn.add_argument(
        "--wavernn-batch-timesteps",
        default=11000,
        type=int,
        help="The time steps for each batch. Only used when batch inference is used",
    )
    wavernn.add_argument(
        "--wavernn-batch-overlap",
        default=550,
        type=int,
        help="The overlapping time steps between batches. Only used when batch inference is used",
    )

    return parser


def unwrap_distributed(state_dict):
    r"""torch.distributed.DistributedDataParallel wraps the model with an additional "module.".
    This function unwraps this layer so that the weights can be loaded on models with a single GPU.

    Args:
        state_dict: Original state_dict.

    Return:
        unwrapped_state_dict: Unwrapped state_dict.
    """

145
    return {k.replace("module.", ""): v for k, v in state_dict.items()}
146
147
148


def nvidia_waveglow_vocode(mel_specgram, device, jit=False):
149
    waveglow = torch.hub.load("NVIDIA/DeepLearningExamples:torchhub", "nvidia_waveglow", model_math="fp16")
150
151
152
153
154
155
156
157
158
159
160
161
162
    waveglow = waveglow.remove_weightnorm(waveglow)
    waveglow = waveglow.to(device)
    waveglow.eval()

    if args.jit:
        raise ValueError("Vocoder option `nvidia_waveglow is not jittable.")

    with torch.no_grad():
        waveform = waveglow.infer(mel_specgram).cpu()

    return waveform


163
164
165
166
167
168
169
170
171
172
173
def wavernn_vocode(
    mel_specgram,
    wavernn_checkpoint_name,
    wavernn_loss,
    wavernn_no_mulaw,
    wavernn_no_batch_inference,
    wavernn_batch_timesteps,
    wavernn_batch_overlap,
    device,
    jit,
):
174
    from torchaudio.models import wavernn
175

176
177
    sys.path.append(os.path.join(os.path.dirname(__file__), "../pipeline_wavernn"))
    from processing import NormalizeDB
178
    from wavernn_inference_wrapper import WaveRNNInferenceWrapper
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200

    wavernn_model = wavernn(wavernn_checkpoint_name).eval().to(device)
    wavernn_inference_model = WaveRNNInferenceWrapper(wavernn_model)

    if jit:
        wavernn_inference_model = torch.jit.script(wavernn_inference_model)

    # WaveRNN spectro setting for default checkpoint
    # n_fft = 2048
    # n_mels = 80
    # win_length = 1100
    # hop_length = 275
    # f_min = 40
    # f_max = 11025

    transforms = torch.nn.Sequential(
        InverseSpectralNormalization(),
        NormalizeDB(min_level_db=-100, normalization=True),
    )
    mel_specgram = transforms(mel_specgram.cpu())

    with torch.no_grad():
201
202
203
204
205
206
207
208
        waveform = wavernn_inference_model(
            mel_specgram.to(device),
            loss_name=wavernn_loss,
            mulaw=(not wavernn_no_mulaw),
            batched=(not wavernn_no_batch_inference),
            timesteps=wavernn_batch_timesteps,
            overlap=wavernn_batch_overlap,
        )
209
210
211
    return waveform.unsqueeze(0)


212
213
214
215
216
217
218
219
220
def griffin_lim_vocode(
    mel_specgram,
    n_fft,
    n_mels,
    sample_rate,
    mel_fmin,
    mel_fmax,
    jit,
):
221
222
223
224
225
226
227
228
229
230
    from torchaudio.transforms import GriffinLim, InverseMelScale

    inv_norm = InverseSpectralNormalization()
    inv_mel = InverseMelScale(
        n_stft=(n_fft // 2 + 1),
        n_mels=n_mels,
        sample_rate=sample_rate,
        f_min=mel_fmin,
        f_max=mel_fmax,
        mel_scale="slaney",
231
        norm="slaney",
232
233
234
235
236
237
238
239
    )
    griffin_lim = GriffinLim(
        n_fft=n_fft,
        power=1,
        hop_length=256,
        win_length=1024,
    )

240
    vocoder = torch.nn.Sequential(inv_norm, inv_mel, griffin_lim)
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258

    if jit:
        vocoder = torch.jit.script(vocoder)

    waveform = vocoder(mel_specgram.cpu())
    return waveform


def main(args):
    torch.manual_seed(0)
    random.seed(0)
    np.random.seed(0)

    device = "cuda" if torch.cuda.is_available() else "cpu"

    if args.checkpoint_path is None and args.checkpoint_name is None:
        raise ValueError("Either --checkpoint-path or --checkpoint-name must be specified.")
    elif args.checkpoint_path is not None and args.checkpoint_name is not None:
259
        raise ValueError("Both --checkpoint-path and --checkpoint-name are specified, " "can only specify one.")
260
261
262
263
264
265
266
267
268
269
270
271
272

    n_symbols = len(get_symbol_list(args.text_preprocessor))
    text_preprocessor = partial(
        text_to_sequence,
        symbol_list=args.text_preprocessor,
        phonemizer=args.phonemizer,
        checkpoint=args.phonemizer_checkpoint,
        cmudict_root=args.cmudict_root,
    )

    if args.checkpoint_path is not None:
        tacotron2 = Tacotron2(n_symbol=n_symbols)
        tacotron2.load_state_dict(
273
274
            unwrap_distributed(torch.load(args.checkpoint_path, map_location=device)["state_dict"])
        )
275
276
277
278
279
        tacotron2 = tacotron2.to(device).eval()
    elif args.checkpoint_name is not None:
        tacotron2 = pretrained_tacotron2(args.checkpoint_name).to(device).eval()

        if n_symbols != tacotron2.n_symbols:
280
281
282
283
284
            raise ValueError(
                "the number of symbols for text_preprocessor ({n_symbols}) "
                "should match the number of symbols for the"
                "pretrained tacotron2 ({tacotron2.n_symbols})."
            )
285
286
287
288

    if args.jit:
        tacotron2 = torch.jit.script(tacotron2)

289
    sequences, lengths = prepare_input_sequence([args.input_text], text_processor=text_preprocessor)
290
291
292
293
294
295
296
297
    sequences, lengths = sequences.long().to(device), lengths.long().to(device)
    with torch.no_grad():
        mel_specgram, _, _ = tacotron2.infer(sequences, lengths)

    if args.vocoder == "nvidia_waveglow":
        waveform = nvidia_waveglow_vocode(mel_specgram=mel_specgram, device=device, jit=args.jit)

    elif args.vocoder == "wavernn":
298
299
300
301
302
303
304
305
306
307
308
        waveform = wavernn_vocode(
            mel_specgram=mel_specgram,
            wavernn_checkpoint_name=args.wavernn_checkpoint_name,
            wavernn_loss=args.wavernn_loss,
            wavernn_no_mulaw=args.wavernn_no_mulaw,
            wavernn_no_batch_inference=args.wavernn_no_batch_inference,
            wavernn_batch_timesteps=args.wavernn_batch_timesteps,
            wavernn_batch_overlap=args.wavernn_batch_overlap,
            device=device,
            jit=args.jit,
        )
309
310

    elif args.vocoder == "griffin_lim":
311
312
313
314
315
316
317
318
319
        waveform = griffin_lim_vocode(
            mel_specgram=mel_specgram,
            n_fft=args.n_fft,
            n_mels=args.n_mels,
            sample_rate=args.sample_rate,
            mel_fmin=args.mel_fmin,
            mel_fmax=args.mel_fmax,
            jit=args.jit,
        )
320
321
322
323
324
325
326
327
328

    torchaudio.save(args.output_path, waveform, args.sample_rate)


if __name__ == "__main__":
    parser = parse_args()
    args, _ = parser.parse_known_args()

    main(args)