Spaces:
Running
Running
File size: 9,630 Bytes
c5cffc6 dc94aa4 c5cffc6 63dd56a c5cffc6 63dd56a c5cffc6 b8db969 c5cffc6 b8f2975 0171728 63dd56a c5cffc6 dc94aa4 c5cffc6 63dd56a c5cffc6 d41970b c5cffc6 63dd56a c5cffc6 4a40b7f c5cffc6 63dd56a 7d18e1c c5cffc6 7d18e1c c5cffc6 63dd56a c5cffc6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 |
#!/usr/bin/python3
# -*- coding: utf-8 -*-
import argparse
import logging
import os
from pathlib import Path
import sys
import uuid
pwd = os.path.abspath(os.path.dirname(__file__))
sys.path.append(os.path.join(pwd, "../../"))
import librosa
import numpy as np
import pandas as pd
from scipy.io import wavfile
import torch
import torch.nn as nn
import torchaudio
from tqdm import tqdm
from toolbox.torchaudio.models.spectrum_dfnet.modeling_spectrum_dfnet import SpectrumDfNetPretrainedModel
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--valid_dataset", default="valid.xlsx", type=str)
parser.add_argument("--model_dir", default="serialization_dir/best", type=str)
parser.add_argument("--evaluation_audio_dir", default="evaluation_audio_dir", type=str)
parser.add_argument("--limit", default=10, type=int)
args = parser.parse_args()
return args
def logging_config():
fmt = "%(asctime)s - %(name)s - %(levelname)s %(filename)s:%(lineno)d > %(message)s"
logging.basicConfig(format=fmt,
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO)
stream_handler = logging.StreamHandler()
stream_handler.setLevel(logging.INFO)
stream_handler.setFormatter(logging.Formatter(fmt))
logger = logging.getLogger(__name__)
return logger
def mix_speech_and_noise(speech: np.ndarray, noise: np.ndarray, snr_db: float):
l1 = len(speech)
l2 = len(noise)
l = min(l1, l2)
speech = speech[:l]
noise = noise[:l]
# np.float32, value between (-1, 1).
speech_power = np.mean(np.square(speech))
noise_power = speech_power / (10 ** (snr_db / 10))
noise_adjusted = np.sqrt(noise_power) * noise / np.sqrt(np.mean(noise ** 2))
noisy_signal = speech + noise_adjusted
return noisy_signal
stft_power = torchaudio.transforms.Spectrogram(
n_fft=512,
win_length=200,
hop_length=80,
power=2.0,
window_fn=torch.hamming_window,
)
stft_complex = torchaudio.transforms.Spectrogram(
n_fft=512,
win_length=200,
hop_length=80,
power=None,
window_fn=torch.hamming_window,
)
istft = torchaudio.transforms.InverseSpectrogram(
n_fft=512,
win_length=200,
hop_length=80,
window_fn=torch.hamming_window,
)
def enhance(mix_spec_complex: torch.Tensor,
speech_spec_prediction: torch.Tensor,
speech_irm_prediction: torch.Tensor,
):
mix_spec_complex = mix_spec_complex.detach().cpu()
speech_spec_prediction = speech_spec_prediction.detach().cpu()
speech_irm_prediction = speech_irm_prediction.detach().cpu()
mask_speech = speech_irm_prediction
mask_noise = 1.0 - speech_irm_prediction
speech_spec = mix_spec_complex * mask_speech
noise_spec = mix_spec_complex * mask_noise
# print(f"speech_spec_prediction: {speech_spec_prediction.shape}")
# print(f"noise_spec: {noise_spec.shape}")
speech_wave = istft.forward(speech_spec_prediction)
# speech_wave = istft.forward(speech_spec)
noise_wave = istft.forward(noise_spec)
return speech_wave, noise_wave
def save_audios(noise_wave: torch.Tensor,
speech_wave: torch.Tensor,
mix_wave: torch.Tensor,
speech_wave_enhanced: torch.Tensor,
noise_wave_enhanced: torch.Tensor,
output_dir: str,
sample_rate: int = 8000,
):
basename = uuid.uuid4().__str__()
output_dir = Path(output_dir) / basename
output_dir.mkdir(parents=True, exist_ok=True)
filename = output_dir / "noise_wave.wav"
torchaudio.save(filename, noise_wave, sample_rate)
filename = output_dir / "speech_wave.wav"
torchaudio.save(filename, speech_wave, sample_rate)
filename = output_dir / "mix_wave.wav"
torchaudio.save(filename, mix_wave, sample_rate)
filename = output_dir / "speech_wave_enhanced.wav"
torchaudio.save(filename, speech_wave_enhanced, sample_rate)
filename = output_dir / "noise_wave_enhanced.wav"
torchaudio.save(filename, noise_wave_enhanced, sample_rate)
return output_dir.as_posix()
def main():
args = get_args()
logger = logging_config()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("GPU available count: {}; device: {}".format(n_gpu, device))
logger.info("prepare model")
model = SpectrumDfNetPretrainedModel.from_pretrained(
pretrained_model_name_or_path=args.model_dir,
)
model.to(device)
model.eval()
# optimizer
logger.info("prepare loss_fn")
irm_mse_loss = nn.MSELoss(
reduction="mean",
)
snr_mse_loss = nn.MSELoss(
reduction="mean",
)
logger.info("read excel")
df = pd.read_excel(args.valid_dataset)
total_loss = 0.
total_examples = 0.
progress_bar = tqdm(total=len(df), desc="Evaluation")
for idx, row in df.iterrows():
noise_filename = row["noise_filename"]
noise_offset = row["noise_offset"]
noise_duration = row["noise_duration"]
speech_filename = row["speech_filename"]
speech_offset = row["speech_offset"]
speech_duration = row["speech_duration"]
snr_db = row["snr_db"]
noise_wave, _ = librosa.load(
noise_filename,
sr=8000,
offset=noise_offset,
duration=noise_duration,
)
speech_wave, _ = librosa.load(
speech_filename,
sr=8000,
offset=speech_offset,
duration=speech_duration,
)
mix_wave: np.ndarray = mix_speech_and_noise(
speech=speech_wave,
noise=noise_wave,
snr_db=snr_db,
)
noise_wave = torch.tensor(noise_wave, dtype=torch.float32)
speech_wave = torch.tensor(speech_wave, dtype=torch.float32)
mix_wave: torch.Tensor = torch.tensor(mix_wave, dtype=torch.float32)
noise_wave = noise_wave.unsqueeze(dim=0)
speech_wave = speech_wave.unsqueeze(dim=0)
mix_wave = mix_wave.unsqueeze(dim=0)
noise_spec: torch.Tensor = stft_power.forward(noise_wave)
speech_spec: torch.Tensor = stft_power.forward(speech_wave)
mix_spec: torch.Tensor = stft_power.forward(mix_wave)
speech_spec_complex: torch.Tensor = stft_complex.forward(speech_wave)
mix_spec_complex: torch.Tensor = stft_complex.forward(mix_wave)
# mix_spec_complex shape: [batch_size, freq_dim (257), time_steps, 2]
noise_spec = noise_spec[:, :-1, :]
speech_spec = speech_spec[:, :-1, :]
mix_spec = mix_spec[:, :-1, :]
speech_spec_complex = speech_spec_complex[:, :-1, :]
mix_spec_complex = mix_spec_complex[:, :-1, :]
speech_irm = speech_spec / (noise_spec + speech_spec)
speech_irm = torch.pow(speech_irm, 1.0)
snr_db: torch.Tensor = 10 * torch.log10(
speech_spec / (noise_spec + 1e-8)
)
snr_db = torch.clamp(snr_db, min=1e-8)
snr_db = torch.mean(snr_db, dim=1, keepdim=True)
# snr_db shape: [batch_size, 1, time_steps]
speech_spec_complex = speech_spec_complex.to(device)
mix_spec_complex = mix_spec_complex.to(device)
mix_spec = mix_spec.to(device)
speech_irm_target = speech_irm.to(device)
snr_db_target = snr_db.to(device)
with torch.no_grad():
speech_spec_prediction, speech_irm_prediction, lsnr_prediction = model.forward(mix_spec_complex)
speech_spec_prediction = torch.view_as_complex(speech_spec_prediction)
irm_loss = irm_mse_loss.forward(speech_irm_prediction, speech_irm_target)
# snr_loss = snr_mse_loss.forward(lsnr_prediction, snr_db_target)
# loss = irm_loss + 0.1 * snr_loss
loss = irm_loss
# mix_spec_complex shape: [batch_size, freq_dim (257), time_steps, 2]
# speech_irm_prediction shape: [batch_size, freq_dim (256), time_steps]
batch_size, _, time_steps = speech_irm_prediction.shape
mix_spec_complex = torch.concat(
[
mix_spec_complex,
torch.zeros(size=(batch_size, 1, time_steps), dtype=mix_spec_complex.dtype).to(device)
],
dim=1,
)
speech_spec_prediction = torch.concat(
[
speech_spec_prediction,
torch.zeros(size=(batch_size, 1, time_steps), dtype=speech_spec_prediction.dtype).to(device)
],
dim=1,
)
speech_irm_prediction = torch.concat(
[
speech_irm_prediction,
0.5 * torch.ones(size=(batch_size, 1, time_steps), dtype=speech_irm_prediction.dtype).to(device)
],
dim=1,
)
# speech_irm_prediction shape: [batch_size, freq_dim (257), time_steps]
speech_wave_enhanced, noise_wave_enhanced = enhance(mix_spec_complex, speech_spec_prediction, speech_irm_prediction)
save_audios(noise_wave, speech_wave, mix_wave, speech_wave_enhanced, noise_wave_enhanced, args.evaluation_audio_dir)
total_loss += loss.item()
total_examples += mix_spec.size(0)
evaluation_loss = total_loss / total_examples
evaluation_loss = round(evaluation_loss, 4)
progress_bar.update(1)
progress_bar.set_postfix({
"evaluation_loss": evaluation_loss,
})
if idx > args.limit:
break
return
if __name__ == '__main__':
main()
|