Add stt-server.py: self-contained recording + VAD + transcription process
Replaces the old stdin/stdout transcription-only server. Now handles the full pipeline in Python: - Launches parec or arecord for mic capture - Runs Silero VAD (via silero-vad, already a faster-whisper dep — no sherpa-onnx needed) - Pre-roll ring buffer (0.2s) prepended to each segment for context - Transcribes with faster-whisper in a separate thread (GPU not blocking VAD) - Emits JSON line events to stdout: ready, vad_start, vad_end, transcript, error Event protocol is designed to map directly to WebSocket subscriptions later. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
205
stt-server.py
Executable file
205
stt-server.py
Executable file
@@ -0,0 +1,205 @@
|
||||
#!/usr/bin/env -S bash -c 'exec "$(dirname "$0")/venv/bin/python3" "$0" "$@"'
|
||||
"""
|
||||
STT process: records audio, runs Silero VAD, transcribes with faster-whisper.
|
||||
|
||||
Events (JSON lines on stdout):
|
||||
{"event": "ready"}
|
||||
{"event": "vad_start"}
|
||||
{"event": "vad_end", "duration": 1.23}
|
||||
{"event": "transcript", "text": "...", "words": [...], "duration": 1.23}
|
||||
{"event": "error", "message": "..."}
|
||||
|
||||
word format: {"word": "hello", "start": 0.12, "end": 0.45, "probability": 0.99}
|
||||
|
||||
All log/status messages go to stderr. Stdout is machine-readable events only.
|
||||
|
||||
Usage:
|
||||
./stt-server.py
|
||||
./stt-server.py --model large-v3 --device cuda --compute-type int8_float16
|
||||
"""
|
||||
|
||||
import sys
|
||||
import json
|
||||
import signal
|
||||
import argparse
|
||||
import threading
|
||||
import queue
|
||||
import subprocess
|
||||
import traceback
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
SAMPLE_RATE = 16000
|
||||
VAD_WINDOW = 512 # samples per VAD chunk (32ms at 16kHz)
|
||||
PRE_ROLL_SAMPLES = 3200 # 0.2s of audio prepended to each segment
|
||||
HISTORY_SAMPLES = 960000 # 60s ring buffer for pre-roll
|
||||
|
||||
|
||||
def log(msg):
|
||||
sys.stderr.write(f'[stt] {msg}\n')
|
||||
sys.stderr.flush()
|
||||
|
||||
|
||||
def emit(event):
|
||||
sys.stdout.write(json.dumps(event) + '\n')
|
||||
sys.stdout.flush()
|
||||
|
||||
|
||||
def find_mic():
|
||||
candidates = [
|
||||
['parec', ['--format=s16le', '--rate=16000', '--channels=1', '--latency-msec=50']],
|
||||
['arecord', ['-f', 'S16_LE', '-r', '16000', '-c', '1', '-t', 'raw', '-q']],
|
||||
]
|
||||
for cmd, args in candidates:
|
||||
try:
|
||||
subprocess.run(['which', cmd], check=True, capture_output=True)
|
||||
return cmd, args
|
||||
except subprocess.CalledProcessError:
|
||||
pass
|
||||
raise RuntimeError('no mic capture command found — need parec or arecord')
|
||||
|
||||
|
||||
def s16le_to_f32(data):
|
||||
return np.frombuffer(data, dtype=np.int16).astype(np.float32) / 32768.0
|
||||
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument('--model', default='base.en')
|
||||
parser.add_argument('--device', default='cuda')
|
||||
parser.add_argument('--compute-type', default='int8_float16')
|
||||
args = parser.parse_args()
|
||||
|
||||
log(f'loading faster-whisper {args.model} ({args.device}, {args.compute_type})...')
|
||||
from faster_whisper import WhisperModel
|
||||
try:
|
||||
model = WhisperModel(args.model, device=args.device, compute_type=args.compute_type)
|
||||
log(f'model ready on {args.device}')
|
||||
except Exception as e:
|
||||
log(f'{args.device} failed ({e}), falling back to cpu')
|
||||
model = WhisperModel(args.model, device='cpu', compute_type='int8')
|
||||
log('model ready on cpu')
|
||||
|
||||
log('loading silero VAD...')
|
||||
from silero_vad import load_silero_vad, VADIterator
|
||||
vad_model = load_silero_vad()
|
||||
vad = VADIterator(vad_model, sampling_rate=SAMPLE_RATE,
|
||||
threshold=0.5, min_silence_duration_ms=500)
|
||||
log('VAD ready')
|
||||
|
||||
|
||||
# Ring buffer for pre-roll context
|
||||
history = np.zeros(HISTORY_SAMPLES, dtype=np.float32)
|
||||
history_pos = 0
|
||||
|
||||
def push_history(samples):
|
||||
global history_pos
|
||||
n = len(samples)
|
||||
base = history_pos % HISTORY_SAMPLES
|
||||
# May wrap around — handle both cases
|
||||
space = HISTORY_SAMPLES - base
|
||||
if n <= space:
|
||||
history[base:base + n] = samples
|
||||
else:
|
||||
history[base:] = samples[:space]
|
||||
history[:n - space] = samples[space:]
|
||||
history_pos += n
|
||||
|
||||
def get_preroll():
|
||||
start = max(0, history_pos - PRE_ROLL_SAMPLES)
|
||||
count = history_pos - start
|
||||
out = np.empty(count, dtype=np.float32)
|
||||
for i in range(count):
|
||||
out[i] = history[(start + i) % HISTORY_SAMPLES]
|
||||
return out
|
||||
|
||||
|
||||
# Transcription runs in a separate thread so VAD is never blocked by GPU
|
||||
transcription_queue = queue.Queue()
|
||||
|
||||
def transcription_worker():
|
||||
while True:
|
||||
item = transcription_queue.get()
|
||||
if item is None:
|
||||
break
|
||||
samples, duration = item
|
||||
try:
|
||||
segments, _ = model.transcribe(
|
||||
samples,
|
||||
language='en',
|
||||
word_timestamps=True,
|
||||
vad_filter=False,
|
||||
)
|
||||
text = ''
|
||||
words = []
|
||||
for seg in segments:
|
||||
text += seg.text
|
||||
for w in (seg.words or []):
|
||||
words.append({
|
||||
'word': w.word,
|
||||
'start': round(float(w.start), 4),
|
||||
'end': round(float(w.end), 4),
|
||||
'probability': round(float(w.probability), 4),
|
||||
})
|
||||
log(f'transcript: {json.dumps(text.strip())} ({len(words)} words)')
|
||||
if text.strip():
|
||||
emit({'event': 'transcript', 'text': text.strip(), 'words': words, 'duration': round(duration, 3)})
|
||||
except Exception:
|
||||
msg = traceback.format_exc()
|
||||
log(f'transcription error:\n{msg}')
|
||||
emit({'event': 'error', 'message': msg})
|
||||
finally:
|
||||
transcription_queue.task_done()
|
||||
|
||||
|
||||
threading.Thread(target=transcription_worker, daemon=True).start()
|
||||
|
||||
|
||||
# Main recording + VAD loop
|
||||
cmd, cmd_args = find_mic()
|
||||
log(f'mic: {cmd} {" ".join(cmd_args)}')
|
||||
mic = subprocess.Popen([cmd] + cmd_args, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL)
|
||||
|
||||
def shutdown(sig=None, frame=None):
|
||||
mic.terminate()
|
||||
transcription_queue.put(None)
|
||||
sys.exit(0)
|
||||
|
||||
signal.signal(signal.SIGTERM, shutdown)
|
||||
signal.signal(signal.SIGINT, shutdown)
|
||||
|
||||
emit({'event': 'ready'})
|
||||
|
||||
speech_samples = []
|
||||
speech_start = None
|
||||
pending = b''
|
||||
|
||||
for chunk in mic.stdout:
|
||||
pending += chunk
|
||||
while len(pending) >= VAD_WINDOW * 2:
|
||||
raw = pending[:VAD_WINDOW * 2]
|
||||
pending = pending[VAD_WINDOW * 2:]
|
||||
|
||||
f32 = s16le_to_f32(raw)
|
||||
push_history(f32)
|
||||
|
||||
result = vad(torch.from_numpy(f32), return_seconds=True)
|
||||
|
||||
if result is not None:
|
||||
if 'start' in result:
|
||||
speech_start = result['start']
|
||||
speech_samples = [get_preroll()]
|
||||
log(f'VAD start at {speech_start:.2f}s')
|
||||
emit({'event': 'vad_start'})
|
||||
|
||||
elif 'end' in result and speech_start is not None:
|
||||
duration = result['end'] - speech_start
|
||||
log(f'VAD end at {result["end"]:.2f}s (duration {duration:.2f}s)')
|
||||
emit({'event': 'vad_end', 'duration': round(duration, 3)})
|
||||
segment = np.concatenate(speech_samples)
|
||||
transcription_queue.put((segment, duration))
|
||||
speech_samples = []
|
||||
speech_start = None
|
||||
vad.reset_states()
|
||||
|
||||
if speech_start is not None:
|
||||
speech_samples.append(f32)
|
||||
Reference in New Issue
Block a user