File size: 10,682 Bytes
c0447ed fca9809 c0447ed fca9809 c0447ed |
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 |
import gc
import time
import numpy as np
import onnxruntime
from datetime import timedelta
from pydub import AudioSegment
from silero_vad import load_silero_vad, get_speech_timestamps, VADIterator
import os
import logging
class FixedVADIterator(VADIterator):
'''It fixes VADIterator by allowing to process any audio length, not only exactly 512 frames at once.
If audio to be processed at once is long and multiple voiced segments detected,
then __call__ returns the start of the first segment, and end (or middle, which means no end) of the last segment.
'''
def reset_states(self):
super().reset_states()
self.buffer = np.array([],dtype=np.float32)
def __call__(self, x, return_seconds=False):
self.buffer = np.append(self.buffer, x)
ret = None
while len(self.buffer) >= 512:
r = super().__call__(self.buffer[:512], return_seconds=return_seconds)
self.buffer = self.buffer[512:]
if ret is None:
ret = r
elif r is not None:
if 'end' in r:
ret['end'] = r['end'] # the latter end
if 'start' in r and 'end' in ret: # there is an earlier start.
# Remove end, merging this segment with the previous one.
del ret['end']
return ret if ret != {} else None
class SileroVADProcessor:
"""
A class for processing audio files using Silero VAD to detect voice activity
and extract voice segments from audio files.
"""
def __init__(self,
activate_threshold=0.5,
fusion_threshold=0.3,
min_speech_duration=0.25,
max_speech_duration=20,
min_silence_duration=250,
sample_rate=16000,
ort_providers=None):
"""
Initialize the SileroVADProcessor.
Args:
activate_threshold (float): Threshold for voice activity detection
fusion_threshold (float): Threshold for merging close speech segments (seconds)
min_speech_duration (float): Minimum duration of speech to be considered valid (seconds)
max_speech_duration (float): Maximum duration of speech (seconds)
min_silence_duration (int): Minimum silence duration (ms)
sample_rate (int): Sample rate of the audio (8000 or 16000 Hz)
ort_providers (list): ONNX Runtime providers for acceleration
"""
# VAD parameters
self.activate_threshold = activate_threshold
self.fusion_threshold = fusion_threshold
self.min_speech_duration = min_speech_duration
self.max_speech_duration = max_speech_duration
self.min_silence_duration = min_silence_duration
self.sample_rate = sample_rate
self.ort_providers = ort_providers if ort_providers else []
# Initialize logger
self.logger = logging.getLogger(__name__)
# Load Silero VAD model
self._init_onnx_session()
self.silero_vad = load_silero_vad(onnx=True)
def _init_onnx_session(self):
"""Initialize ONNX Runtime session with appropriate settings."""
session_opts = onnxruntime.SessionOptions()
session_opts.log_severity_level = 3
session_opts.inter_op_num_threads = 0
session_opts.intra_op_num_threads = 0
session_opts.enable_cpu_mem_arena = True
session_opts.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
session_opts.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
session_opts.add_session_config_entry("session.intra_op.allow_spinning", "1")
session_opts.add_session_config_entry("session.inter_op.allow_spinning", "1")
session_opts.add_session_config_entry("session.set_denormal_as_zero", "1")
# Set the session_opts to be used by silero_vad
# onnxruntime.capi._pybind_state.get_default_session_options(session_opts)
def load_audio(self, audio_path):
"""
Load audio file and prepare it for VAD processing.
Args:
audio_path (str): Path to the audio file
Returns:
numpy.ndarray: Audio data as numpy array
"""
self.logger.info(f"Loading audio from {audio_path}")
audio_segment = AudioSegment.from_file(audio_path)
audio_segment = audio_segment.set_channels(1).set_frame_rate(self.sample_rate)
# Convert to numpy array and normalize
dtype = np.float16 if self.use_gpu_fp16 else np.float32
audio_array = np.array(audio_segment.get_array_of_samples(), dtype=dtype) * 0.000030517578 # 1/32768
self.audio_segment = audio_segment # Store for later use
return audio_array
@property
def model(self):
return self.silero_vad
def process_timestamps(self, timestamps):
"""
Process VAD timestamps: filter short segments and merge close segments.
Args:
timestamps (list): List of (start, end) tuples
Returns:
list: Processed list of (start, end) tuples
"""
# Filter out short durations
filtered_timestamps = [(start, end) for start, end in timestamps
if (end - start) >= self.min_speech_duration]
# Fuse timestamps in two passes for better merging
fused_timestamps_1st = []
for start, end in filtered_timestamps:
if fused_timestamps_1st and (start - fused_timestamps_1st[-1][1] <= self.fusion_threshold):
fused_timestamps_1st[-1] = (fused_timestamps_1st[-1][0], end)
else:
fused_timestamps_1st.append((start, end))
fused_timestamps_2nd = []
for start, end in fused_timestamps_1st:
if fused_timestamps_2nd and (start - fused_timestamps_2nd[-1][1] <= self.fusion_threshold):
fused_timestamps_2nd[-1] = (fused_timestamps_2nd[-1][0], end)
else:
fused_timestamps_2nd.append((start, end))
return fused_timestamps_2nd
def format_time(self, seconds):
"""
Convert seconds to VTT time format 'hh:mm:ss.mmm'.
Args:
seconds (float): Time in seconds
Returns:
str: Formatted time string
"""
td = timedelta(seconds=seconds)
td_sec = td.total_seconds()
total_seconds = int(td_sec)
milliseconds = int((td_sec - total_seconds) * 1000)
hours = total_seconds // 3600
minutes = (total_seconds % 3600) // 60
seconds = total_seconds % 60
return f"{hours:02}:{minutes:02}:{seconds:02}.{milliseconds:03}"
def detect_speech(self, audio:np.array):
"""
Run VAD on the audio file to detect speech segments.
Args:
audio_path (str): Path to the audio file
Returns:
list: List of processed timestamps as (start, end) tuples
"""
self.logger.info("Starting VAD process")
start_time = time.time()
# Get speech timestamps
raw_timestamps = get_speech_timestamps(
audio,
model=self.silero_vad,
threshold=self.activate_threshold,
max_speech_duration_s=self.max_speech_duration,
min_speech_duration_ms=int(self.min_speech_duration * 1000),
min_silence_duration_ms=self.min_silence_duration,
return_seconds=True
)
# Convert to simple format and process
timestamps = [(item['start'], item['end']) for item in raw_timestamps]
processed_timestamps = self.process_timestamps(timestamps)
# Clean up
del audio
gc.collect()
self.logger.info(f"VAD completed in {time.time() - start_time:.3f} seconds")
return processed_timestamps
"""
Save timestamps in both second and sample indices formats.
Args:
timestamps (list): List of (start, end) tuples
output_prefix (str): Prefix for output files
"""
# Save timestamps in seconds (VTT format)
seconds_path = f"{output_prefix}_timestamps_second.txt"
with open(seconds_path, "w", encoding='UTF-8') as file:
self.logger.info("Saving timestamps in seconds format")
for start, end in timestamps:
s_time = self.format_time(start)
e_time = self.format_time(end)
line = f"{s_time} --> {e_time}\n"
file.write(line)
# Save timestamps in sample indices
indices_path = f"{output_prefix}_timestamps_indices.txt"
with open(indices_path, "w", encoding='UTF-8') as file:
self.logger.info("Saving timestamps in indices format")
for start, end in timestamps:
line = f"{int(start * self.sample_rate)} --> {int(end * self.sample_rate)}\n"
file.write(line)
self.logger.info(f"Timestamps saved to {seconds_path} and {indices_path}")
def extract_speech_segments(self, audio_segment, timestamps):
"""
Extract speech segments from the audio and combine them into a single audio file.
Args:
timestamps (list): List of (start, end) tuples indicating speech segments
Returns:
AudioSegment: The combined speech segments
"""
audio_segment = audio_segment.numpy()
combined_speech = np.array([], dtype=np.float32)
# Extract and combine each speech segment
for i, (start, end) in enumerate(timestamps):
# Convert seconds to milliseconds for pydub
start_ms = int(start * 1000)
end_ms = int(end * 1000)
# Ensure the end time does not exceed the length of the audio segment
if end_ms > len(audio_segment):
end_ms = len(audio_segment)
# Extract the segment
segment = audio_segment[start_ms:end_ms]
# Add to combined audio
combined_speech = np.append(combined_speech, segment)
return combined_speech
def process_audio(self, audio_array:np.array):
"""
Complete processing pipeline: detect speech, save timestamps, and optionally extract speech.
Returns:
tuple: (timestamps, output_speech_path if extract_speech else None)
"""
# Run VAD to detect speech
timestamps = self.detect_speech(audio_array)
combined_speech = self.extract_speech_segments(audio_array, timestamps)
return timestamps, combined_speech
|