bhetherman
263f39e0a3
Upgrade PyTorch to 2.7+ with cu128 wheels for Blackwell (sm_120) GPU support. Replace silero-vad (which depends on torchaudio) with a direct ONNX Runtime implementation of the same Silero VAD model, eliminating the torchaudio dependency entirely. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
91 lines
3.0 KiB
Python
91 lines
3.0 KiB
Python
import numpy as np
|
|
import onnxruntime
|
|
|
|
|
|
class SileroVADOnnx:
|
|
"""Silero VAD model loaded via ONNX Runtime (no torchaudio dependency)."""
|
|
|
|
SAMPLE_RATE = 16000
|
|
WINDOW_SIZE = 512 # 32ms at 16kHz
|
|
|
|
def __init__(self, model_path: str):
|
|
opts = onnxruntime.SessionOptions()
|
|
opts.inter_op_num_threads = 1
|
|
opts.intra_op_num_threads = 1
|
|
self.session = onnxruntime.InferenceSession(
|
|
model_path, sess_options=opts
|
|
)
|
|
self._reset_state()
|
|
|
|
def _reset_state(self):
|
|
self._h = np.zeros((2, 1, 64), dtype=np.float32)
|
|
self._c = np.zeros((2, 1, 64), dtype=np.float32)
|
|
|
|
def reset_states(self):
|
|
self._reset_state()
|
|
|
|
def __call__(self, chunk: np.ndarray) -> float:
|
|
"""Run VAD on a single audio chunk. Returns speech probability."""
|
|
input_data = chunk[np.newaxis, :] # add batch dim
|
|
sr = np.array(self.SAMPLE_RATE, dtype=np.int64)
|
|
|
|
ort_inputs = {
|
|
"input": input_data,
|
|
"sr": sr,
|
|
"h": self._h,
|
|
"c": self._c,
|
|
}
|
|
out, self._h, self._c = self.session.run(None, ort_inputs)
|
|
return float(out.squeeze())
|
|
|
|
|
|
class StreamingVAD:
|
|
"""Wraps Silero VAD (ONNX) for streaming chunk-by-chunk speech detection."""
|
|
|
|
def __init__(self, model: SileroVADOnnx, threshold: float = 0.5,
|
|
min_silence_ms: int = 400):
|
|
self.model = model
|
|
self.threshold = threshold
|
|
self.min_silence_samples = int(
|
|
SileroVADOnnx.SAMPLE_RATE * min_silence_ms / 1000
|
|
)
|
|
self.audio_buffer: list[np.ndarray] = []
|
|
self.is_speaking = False
|
|
self._silence_samples = 0
|
|
|
|
def process_chunk(self, chunk_16k: np.ndarray) -> np.ndarray | None:
|
|
"""Feed a 512-sample chunk at 16kHz.
|
|
|
|
Returns the complete utterance as a numpy array when speech ends,
|
|
or None if still accumulating.
|
|
"""
|
|
prob = self.model(chunk_16k.astype(np.float32))
|
|
|
|
if prob >= self.threshold:
|
|
self._silence_samples = 0
|
|
if not self.is_speaking:
|
|
self.is_speaking = True
|
|
self.audio_buffer = []
|
|
self.audio_buffer.append(chunk_16k.copy())
|
|
elif self.is_speaking:
|
|
self._silence_samples += len(chunk_16k)
|
|
self.audio_buffer.append(chunk_16k.copy())
|
|
if self._silence_samples >= self.min_silence_samples:
|
|
self.is_speaking = False
|
|
self._silence_samples = 0
|
|
if self.audio_buffer:
|
|
result = np.concatenate(self.audio_buffer)
|
|
self.audio_buffer = []
|
|
self.model.reset_states()
|
|
return result
|
|
self.model.reset_states()
|
|
|
|
return None
|
|
|
|
def reset(self):
|
|
"""Reset VAD state for a new conversation turn."""
|
|
self.audio_buffer = []
|
|
self.is_speaking = False
|
|
self._silence_samples = 0
|
|
self.model.reset_states()
|