File size: 12,068 Bytes
7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 b3f2891 7e17176 3a820e8 b3f2891 7e17176 3a820e8 8a5e901 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 b3f2891 7e17176 b3f2891 7e17176 b3f2891 3a820e8 7e17176 b3f2891 7e17176 b3f2891 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 3a820e8 7e17176 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 |
#!/usr/bin/python3
# -*- coding: utf-8 -*-
import argparse
import collections
import os
from typing import List
import matplotlib.pyplot as plt
import numpy as np
from scipy.io import wavfile
import torch
import webrtcvad
from project_settings import project_path
class FrameVoiceClassifier(object):
def predict(self, chunk: np.ndarray) -> float:
raise NotImplementedError
class WebRTCVoiceClassifier(FrameVoiceClassifier):
def __init__(self,
agg: int = 3,
sample_rate: int = 8000
):
self.agg = agg
self.sample_rate = sample_rate
self.model = webrtcvad.Vad(mode=agg)
def predict(self, chunk: np.ndarray) -> float:
if chunk.dtype != np.int16:
raise AssertionError("signal dtype should be np.int16, instead of {}".format(chunk.dtype))
audio_bytes = bytes(chunk)
is_speech = self.model.is_speech(audio_bytes, self.sample_rate)
return 1.0 if is_speech else 0.0
class SileroVoiceClassifier(FrameVoiceClassifier):
def __init__(self,
model_path: str,
sample_rate: int = 8000):
self.model_path = model_path
self.sample_rate = sample_rate
with open(self.model_path, "rb") as f:
model = torch.jit.load(f, map_location="cpu")
self.model = model
self.model.reset_states()
def predict(self, chunk: np.ndarray) -> float:
if self.sample_rate / len(chunk) > 31.25:
raise AssertionError("chunk samples number {} is less than {}".format(len(chunk), self.sample_rate / 31.25))
if chunk.dtype != np.int16:
raise AssertionError("signal dtype should be np.int16, instead of {}".format(chunk.dtype))
chunk = chunk / 32768
chunk = torch.tensor(chunk, dtype=torch.float32)
speech_prob = self.model(chunk, self.sample_rate).item()
return float(speech_prob)
class CallVoiceClassifier(FrameVoiceClassifier):
def __init__(self,
model_path: str,
sample_rate: int = 8000):
self.model_path = model_path
self.sample_rate = sample_rate
self.model = torch.jit.load(os.path.join(model_path, "cnn_voicemail.pth"))
def predict(self, chunk: np.ndarray) -> float:
if chunk.dtype != np.int16:
raise AssertionError("signal dtype should be np.int16, instead of {}".format(chunk.dtype))
chunk = chunk / 32768
inputs = torch.tensor(chunk, dtype=torch.float32)
inputs = torch.unsqueeze(inputs, dim=0)
try:
outputs = self.model(inputs)
except RuntimeError as e:
print(inputs.shape)
raise e
probs = outputs["probs"]
voice_prob = probs[0][2]
return float(voice_prob)
class Frame(object):
def __init__(self, signal: np.ndarray, timestamp_s: float):
self.signal = signal
self.timestamp_s = timestamp_s
class Vad(object):
def __init__(self,
model: FrameVoiceClassifier,
start_ring_rate: float = 0.5,
end_ring_rate: float = 0.5,
frame_length_ms: int = 30,
frame_step_ms: int = 30,
padding_length_ms: int = 300,
max_silence_length_ms: int = 300,
max_speech_length_s: float = 2.0,
min_speech_length_s: float = 0.3,
sample_rate: int = 8000
):
self.model = model
self.start_ring_rate = start_ring_rate
self.end_ring_rate = end_ring_rate
self.frame_length_ms = frame_length_ms
self.padding_length_ms = padding_length_ms
self.max_silence_length_ms = max_silence_length_ms
self.max_speech_length_s = max_speech_length_s
self.min_speech_length_s = min_speech_length_s
self.sample_rate = sample_rate
# frames
self.frame_length = int(sample_rate * (frame_length_ms / 1000.0))
self.frame_step = int(sample_rate * (frame_step_ms / 1000.0))
self.frame_timestamp_s = 0.0
self.signal_cache = np.zeros(shape=(self.frame_length,), dtype=np.int16)
# self.signal_cache = None
# segments
self.num_padding_frames = int(padding_length_ms / frame_step_ms)
self.ring_buffer = collections.deque(maxlen=self.num_padding_frames)
self.triggered = False
self.voiced_frames: List[Frame] = list()
self.segments = list()
# vad segments
self.is_first_segment = True
self.timestamp_start_s = 0.0
self.timestamp_end_s = 0.0
# speech probs
self.speech_probs: List[float] = list()
def signal_to_frames(self, signal: np.ndarray):
frames = list()
l = len(signal)
duration_s = float(self.frame_step) / self.sample_rate
for offset in range(0, l - self.frame_length + 1, self.frame_step):
sub_signal = signal[offset:offset+self.frame_length]
frame = Frame(sub_signal, self.frame_timestamp_s)
self.frame_timestamp_s += duration_s
frames.append(frame)
return frames
def segments_generator(self, signal: np.ndarray):
# signal rounding
if self.signal_cache is not None:
signal = np.concatenate([self.signal_cache, signal])
# rest
rest = (len(signal) - self.frame_length) % self.frame_step
if rest == 0:
self.signal_cache = None
signal_ = signal
else:
self.signal_cache = signal[-rest:]
signal_ = signal[:-rest]
# frames
frames = self.signal_to_frames(signal_)
for frame in frames:
speech_prob = self.model.predict(frame.signal)
self.speech_probs.append(speech_prob)
if not self.triggered:
self.ring_buffer.append((frame, speech_prob))
num_voiced = sum([p for _, p in self.ring_buffer])
if num_voiced > self.start_ring_rate * self.ring_buffer.maxlen:
self.triggered = True
for f, _ in self.ring_buffer:
self.voiced_frames.append(f)
continue
self.voiced_frames.append(frame)
self.ring_buffer.append((frame, speech_prob))
num_voiced = sum([p for _, p in self.ring_buffer])
if num_voiced < self.end_ring_rate * self.ring_buffer.maxlen:
segment = [
np.concatenate([f.signal for f in self.voiced_frames]),
self.voiced_frames[0].timestamp_s,
self.voiced_frames[-1].timestamp_s,
]
yield segment
self.triggered = False
self.ring_buffer.clear()
self.voiced_frames = []
continue
def vad_segments_generator(self, segments_generator):
segments = list(segments_generator)
for i, segment in enumerate(segments):
start = round(segment[1], 4)
end = round(segment[2], 4)
if self.timestamp_start_s is None and self.timestamp_end_s is None:
self.timestamp_start_s = start
self.timestamp_end_s = end
continue
if self.timestamp_end_s - self.timestamp_start_s > self.max_speech_length_s:
end_ = self.timestamp_start_s + self.max_speech_length_s
vad_segment = [self.timestamp_start_s, end_]
yield vad_segment
self.timestamp_start_s = end_
silence_length_ms = (start - self.timestamp_end_s) * 1000
if silence_length_ms < self.max_silence_length_ms:
self.timestamp_end_s = end
continue
if self.timestamp_end_s - self.timestamp_start_s < self.min_speech_length_s:
self.timestamp_start_s = start
self.timestamp_end_s = end
continue
vad_segment = [self.timestamp_start_s, self.timestamp_end_s]
yield vad_segment
self.timestamp_start_s = start
self.timestamp_end_s = end
def vad(self, signal: np.ndarray) -> List[list]:
segments = self.segments_generator(signal)
vad_segments = self.vad_segments_generator(segments)
vad_segments = list(vad_segments)
return vad_segments
def last_vad_segments(self) -> List[list]:
# last segments
if len(self.voiced_frames) == 0:
segments = []
else:
segment = [
np.concatenate([f.signal for f in self.voiced_frames]),
self.voiced_frames[0].timestamp_s,
self.voiced_frames[-1].timestamp_s
]
segments = [segment]
# last vad segments
vad_segments = self.vad_segments_generator(segments)
vad_segments = list(vad_segments)
if self.timestamp_end_s > 1e-5 and self.timestamp_end_s > 1e-5:
vad_segments = vad_segments + [[self.timestamp_start_s, self.timestamp_end_s]]
return vad_segments
def process_speech_probs(signal: np.ndarray, speech_probs: List[float], frame_step: int) -> np.ndarray:
speech_probs_ = list()
for p in speech_probs[1:]:
speech_probs_.extend([p] * frame_step)
pad = (signal.shape[0] - len(speech_probs_))
speech_probs_ = speech_probs_ + [0.0] * pad
speech_probs_ = np.array(speech_probs_, dtype=np.float32)
if len(speech_probs_) != len(signal):
raise AssertionError
return speech_probs_
def make_visualization(signal: np.ndarray, speech_probs, sample_rate: int, vad_segments: list):
time = np.arange(0, len(signal)) / sample_rate
plt.figure(figsize=(12, 5))
plt.plot(time, signal / 32768, color='b')
plt.plot(time, speech_probs, color='gray')
for start, end in vad_segments:
plt.axvline(x=start, ymin=0.15, ymax=0.85, color="g", linestyle="--", label="开始端点")
plt.axvline(x=end, ymin=0.15, ymax=0.85, color="r", linestyle="--", label="结束端点")
plt.show()
return
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--wav_file",
default=(project_path / "data/early_media/62/3300999628999191096.wav").as_posix(),
type=str,
)
parser.add_argument(
"--model_path",
default=(project_path / "pretrained_models/silero_vad/silero_vad.jit").as_posix(),
type=str,
)
args = parser.parse_args()
return args
SAMPLE_RATE = 8000
def main():
args = get_args()
sample_rate, signal = wavfile.read(args.wav_file)
if SAMPLE_RATE != sample_rate:
raise AssertionError
# model = SileroVoiceClassifier(model_path=args.model_path, sample_rate=SAMPLE_RATE)
# model = WebRTCVoiceClassifier(agg=1, sample_rate=SAMPLE_RATE)
model = CallVoiceClassifier(model_path=(project_path / "trained_models/cnn_voicemail_common_20231130").as_posix())
vad = Vad(model=model,
start_ring_rate=0.2,
end_ring_rate=0.1,
frame_length_ms=300,
frame_step_ms=30,
padding_length_ms=300,
max_silence_length_ms=300,
sample_rate=SAMPLE_RATE,
)
print(vad)
vad_segments = list()
segments = vad.vad(signal)
vad_segments += segments
for segment in segments:
print(segment)
# last vad segment
segments = vad.last_vad_segments()
vad_segments += segments
for segment in segments:
print(segment)
print(vad.speech_probs)
print(len(vad.speech_probs))
# speech_probs
speech_probs = process_speech_probs(
signal=signal,
speech_probs=vad.speech_probs,
frame_step=vad.frame_step,
)
# plot
make_visualization(signal, speech_probs, SAMPLE_RATE, vad_segments)
return
if __name__ == '__main__':
main()
|