LiveTranslate / vad_processor.py
fasdfsa's picture
init
bd95217
import logging
import collections
import numpy as np
import torch
torch.set_num_threads(1)
log = logging.getLogger("LiveTrans.VAD")
class VADProcessor:
"""Voice Activity Detection with multiple modes."""
def __init__(
self,
sample_rate=16000,
threshold=0.50,
min_speech_duration=1.0,
max_speech_duration=15.0,
chunk_duration=0.032,
):
self.sample_rate = sample_rate
self.threshold = threshold
self.energy_threshold = 0.02
self.min_speech_samples = int(min_speech_duration * sample_rate)
self.max_speech_samples = int(max_speech_duration * sample_rate)
self._chunk_duration = chunk_duration
self.mode = "silero" # "silero", "energy", "disabled"
self._model, self._utils = torch.hub.load(
repo_or_dir="snakers4/silero-vad",
model="silero_vad",
trust_repo=True,
)
self._model.eval()
self._speech_buffer = []
self._confidence_history = [] # per-chunk confidence, synced with _speech_buffer
self._speech_samples = 0
self._is_speaking = False
self._silence_counter = 0
# Pre-speech ring buffer: capture onset consonants before VAD triggers
self._pre_speech_chunks = 3 # ~96ms at 32ms/chunk
self._pre_buffer = collections.deque(maxlen=self._pre_speech_chunks)
# Silence timing
self._silence_mode = "auto" # "auto" or "fixed"
self._fixed_silence_dur = 0.8
self._silence_limit = self._seconds_to_chunks(0.8)
# Progressive silence: shorter threshold when buffer is long
self._progressive_tiers = [
# (buffer_seconds, silence_multiplier)
(3.0, 1.0), # < 3s: use full silence_limit
(6.0, 0.5), # 3-6s: use half silence_limit
(10.0, 0.25), # 6-10s: use quarter silence_limit
]
# Adaptive silence tracking: recent pause durations (seconds)
self._pause_history = collections.deque(maxlen=50)
self._adaptive_min = 0.3
self._adaptive_max = 2.0
# Exposed for monitor
self.last_confidence = 0.0
def _seconds_to_chunks(self, seconds: float) -> int:
return max(1, round(seconds / self._chunk_duration))
def _update_adaptive_limit(self):
if len(self._pause_history) < 3:
return
pauses = sorted(self._pause_history)
# P75 of recent pauses × 1.2
idx = int(len(pauses) * 0.75)
p75 = pauses[min(idx, len(pauses) - 1)]
target = max(self._adaptive_min, min(self._adaptive_max, p75 * 1.2))
new_limit = self._seconds_to_chunks(target)
if new_limit != self._silence_limit:
log.debug(f"Adaptive silence: {target:.2f}s ({new_limit} chunks), P75={p75:.2f}s")
self._silence_limit = new_limit
def update_settings(self, settings: dict):
if "vad_mode" in settings:
self.mode = settings["vad_mode"]
if "vad_threshold" in settings:
self.threshold = settings["vad_threshold"]
if "energy_threshold" in settings:
self.energy_threshold = settings["energy_threshold"]
if "min_speech_duration" in settings:
self.min_speech_samples = int(
settings["min_speech_duration"] * self.sample_rate
)
if "max_speech_duration" in settings:
self.max_speech_samples = int(
settings["max_speech_duration"] * self.sample_rate
)
if "silence_mode" in settings:
self._silence_mode = settings["silence_mode"]
if "silence_duration" in settings:
self._fixed_silence_dur = settings["silence_duration"]
if self._silence_mode == "fixed":
self._silence_limit = self._seconds_to_chunks(self._fixed_silence_dur)
log.info(
f"VAD settings updated: mode={self.mode}, threshold={self.threshold}, "
f"silence={self._silence_mode} "
f"({self._silence_limit} chunks = {self._silence_limit * self._chunk_duration:.2f}s)"
)
def _silero_confidence(self, audio_chunk: np.ndarray) -> float:
window_size = 512 if self.sample_rate == 16000 else 256
chunk = audio_chunk[:window_size]
if len(chunk) < window_size:
chunk = np.pad(chunk, (0, window_size - len(chunk)))
tensor = torch.from_numpy(chunk).float()
return self._model(tensor, self.sample_rate).item()
def _energy_confidence(self, audio_chunk: np.ndarray) -> float:
rms = float(np.sqrt(np.mean(audio_chunk**2)))
return min(1.0, rms / (self.energy_threshold * 2))
def _get_confidence(self, audio_chunk: np.ndarray) -> float:
if self.mode == "silero":
return self._silero_confidence(audio_chunk)
elif self.mode == "energy":
return self._energy_confidence(audio_chunk)
else: # disabled
return 1.0
def _get_effective_silence_limit(self) -> int:
"""Progressive silence: accept shorter pauses as split points when buffer is long."""
buf_seconds = self._speech_samples / self.sample_rate
multiplier = 1.0
for tier_sec, tier_mult in self._progressive_tiers:
if buf_seconds < tier_sec:
break
multiplier = tier_mult
effective = max(1, round(self._silence_limit * multiplier))
return effective
def process_chunk(self, audio_chunk: np.ndarray):
confidence = self._get_confidence(audio_chunk)
self.last_confidence = confidence
effective_threshold = self.threshold if self.mode == "silero" else 0.5
eff_silence_limit = self._get_effective_silence_limit()
log.debug(
f"VAD conf={confidence:.3f} ({self.mode}), speaking={self._is_speaking}, "
f"buf={self._speech_samples / self.sample_rate:.1f}s, "
f"silence_cnt={self._silence_counter}, limit={eff_silence_limit} "
f"(base={self._silence_limit})"
)
if confidence >= effective_threshold:
# Record pause duration for adaptive mode
if self._is_speaking and self._silence_counter > 0:
pause_dur = self._silence_counter * self._chunk_duration
if pause_dur >= 0.1:
self._pause_history.append(pause_dur)
if self._silence_mode == "auto":
self._update_adaptive_limit()
if not self._is_speaking:
# Speech onset: prepend pre-speech buffer to capture leading consonants
# Use threshold as confidence so these chunks don't create false valleys
for pre_chunk in self._pre_buffer:
self._speech_buffer.append(pre_chunk)
self._confidence_history.append(effective_threshold)
self._speech_samples += len(pre_chunk)
self._pre_buffer.clear()
self._is_speaking = True
self._silence_counter = 0
self._speech_buffer.append(audio_chunk)
self._confidence_history.append(confidence)
self._speech_samples += len(audio_chunk)
elif self._is_speaking:
self._silence_counter += 1
self._speech_buffer.append(audio_chunk)
self._confidence_history.append(confidence)
self._speech_samples += len(audio_chunk)
else:
# Not speaking: feed pre-speech ring buffer
self._pre_buffer.append(audio_chunk)
# Force segment if max duration reached — backtrack to find best split point
if self._speech_samples >= self.max_speech_samples:
return self._split_at_best_pause()
# End segment after enough silence (progressive threshold)
if self._is_speaking and self._silence_counter >= eff_silence_limit:
if self._speech_samples >= self.min_speech_samples:
return self._flush_segment()
else:
self._reset()
return None
return None
def _find_best_split_index(self) -> int:
"""Find the best chunk index to split at using smoothed confidence.
A sliding window average reduces single-chunk noise, then we find
the center of the lowest valley. Works even when the speaker never
fully pauses (e.g. fast commentary).
Returns -1 if no usable split point found."""
n = len(self._confidence_history)
if n < 4:
return -1
# Smooth confidence with a sliding window (~160ms = 5 chunks at 32ms)
smooth_win = min(5, n // 2)
smoothed = []
for i in range(n):
lo = max(0, i - smooth_win // 2)
hi = min(n, i + smooth_win // 2 + 1)
smoothed.append(sum(self._confidence_history[lo:hi]) / (hi - lo))
# Search in the latter 70% of the buffer (avoid splitting too early)
search_start = max(1, n * 3 // 10)
# Find global minimum in smoothed curve
min_val = float("inf")
min_idx = -1
for i in range(search_start, n):
if smoothed[i] <= min_val:
min_val = smoothed[i]
min_idx = i
if min_idx <= 0:
return -1
# Check if this is a meaningful dip
avg_conf = sum(smoothed[search_start:]) / max(1, n - search_start)
dip_ratio = min_val / max(avg_conf, 1e-6)
effective_threshold = self.threshold if self.mode == "silero" else 0.5
if min_val < effective_threshold or dip_ratio < 0.8:
log.debug(
f"Split point at chunk {min_idx}/{n}: "
f"smoothed={min_val:.3f}, avg={avg_conf:.3f}, dip_ratio={dip_ratio:.2f}"
)
return min_idx
# Fallback: any point below average is better than hard cut
if min_val < avg_conf:
log.debug(
f"Split point (fallback) at chunk {min_idx}/{n}: "
f"smoothed={min_val:.3f}, avg={avg_conf:.3f}"
)
return min_idx
return -1
def _split_at_best_pause(self):
"""When hitting max duration, backtrack to find the best pause point.
Flushes the first part and keeps the remainder for continued accumulation."""
if not self._speech_buffer:
return None
split_idx = self._find_best_split_index()
if split_idx <= 0:
# No good split point — hard flush everything
log.info(
f"Max duration reached, no good split point, "
f"hard flush {self._speech_samples / self.sample_rate:.1f}s"
)
return self._flush_segment()
# Split: emit first part, keep remainder
first_bufs = self._speech_buffer[:split_idx]
remain_bufs = self._speech_buffer[split_idx:]
remain_confs = self._confidence_history[split_idx:]
first_samples = sum(len(b) for b in first_bufs)
remain_samples = sum(len(b) for b in remain_bufs)
log.info(
f"Max duration split at {first_samples / self.sample_rate:.1f}s, "
f"keeping {remain_samples / self.sample_rate:.1f}s remainder"
)
segment = np.concatenate(first_bufs)
# Keep remainder in buffer for next segment
self._speech_buffer = remain_bufs
self._confidence_history = remain_confs
self._speech_samples = remain_samples
self._is_speaking = True
self._silence_counter = 0
return segment
def _flush_segment(self):
if not self._speech_buffer:
return None
# Speech density check: discard segments where most chunks are below threshold
if len(self._confidence_history) >= 4:
effective_threshold = self.threshold if self.mode == "silero" else 0.5
voiced = sum(1 for c in self._confidence_history if c >= effective_threshold)
density = voiced / len(self._confidence_history)
if density < 0.25:
dur = self._speech_samples / self.sample_rate
log.debug(
f"Low speech density {density:.0%} ({voiced}/{len(self._confidence_history)}), "
f"discarding {dur:.1f}s segment"
)
self._reset()
return None
segment = np.concatenate(self._speech_buffer)
self._reset()
return segment
def _reset(self):
self._speech_buffer = []
self._confidence_history = []
self._speech_samples = 0
self._is_speaking = False
self._silence_counter = 0
def flush(self):
if self._speech_samples >= self.min_speech_samples:
return self._flush_segment()
self._reset()
return None