mirror of
https://github.com/R0m1k3/LiveFlow.git
synced 2026-10-12 01:37:29 +02:00
Claude
a8517c3597
- docker-compose : Caddy (HTTPS local) + app FastAPI + Qwen3-ASR via vLLM (GPU) - Backend : WebSocket audio, segmentation VAD (WebRTC), appel ASR compatible OpenAI, stockage SQLite, exports TXT/MD/SRT/JSON - Frontend : capture micro (AudioWorklet, rééchantillonnage 16 kHz), transcription live, historique des réunions, copie et téléchargement - README : démarrage, prérequis GPU, changement de moteur ASR https://claude.ai/code/session_01YHMp3EKzr4s6o8w1ygxuUe
101 lines
3.6 KiB
Python
101 lines
3.6 KiB
Python
"""Découpage du flux micro en segments de parole via WebRTC VAD.
|
|
|
|
Le client envoie du PCM 16 bits mono 16 kHz. On analyse des trames de 30 ms :
|
|
un segment démarre quand la parole domine la fenêtre récente (avec un
|
|
pré-roll pour ne pas couper le début de phrase) et se termine après un
|
|
silence prolongé ou une durée maximale.
|
|
"""
|
|
|
|
from collections import deque
|
|
from dataclasses import dataclass
|
|
|
|
import webrtcvad
|
|
|
|
SAMPLE_RATE = 16000
|
|
FRAME_MS = 30
|
|
FRAME_BYTES = SAMPLE_RATE * FRAME_MS // 1000 * 2 # 960 octets
|
|
|
|
VAD_AGGRESSIVENESS = 2
|
|
PREROLL_FRAMES = 10 # 300 ms conservées avant le déclenchement
|
|
TRIGGER_RATIO = 0.6 # part de trames "parole" du pré-roll pour démarrer
|
|
SILENCE_END_MS = 700 # silence qui clôt un segment
|
|
MIN_SPEECH_MS = 300 # en dessous, le segment est ignoré (bruit)
|
|
MAX_SEGMENT_S = 25 # coupe forcée pour garder une latence raisonnable
|
|
|
|
|
|
@dataclass
|
|
class Segment:
|
|
pcm: bytes
|
|
t0: float # secondes depuis le début de la réunion
|
|
t1: float
|
|
|
|
|
|
class SpeechSegmenter:
|
|
def __init__(self):
|
|
self._vad = webrtcvad.Vad(VAD_AGGRESSIVENESS)
|
|
self._pending = bytearray()
|
|
self._ring: deque[tuple[bytes, bool]] = deque(maxlen=PREROLL_FRAMES)
|
|
self._frame_index = 0
|
|
self._triggered = False
|
|
self._segment = bytearray()
|
|
self._segment_start_frame = 0
|
|
self._silence_frames = 0
|
|
self._speech_frames = 0
|
|
|
|
def feed(self, data: bytes) -> list[Segment]:
|
|
"""Ajoute de l'audio brut et renvoie les segments terminés."""
|
|
self._pending.extend(data)
|
|
segments = []
|
|
while len(self._pending) >= FRAME_BYTES:
|
|
frame = bytes(self._pending[:FRAME_BYTES])
|
|
del self._pending[:FRAME_BYTES]
|
|
seg = self._process_frame(frame)
|
|
if seg is not None:
|
|
segments.append(seg)
|
|
return segments
|
|
|
|
def flush(self) -> Segment | None:
|
|
"""Clôt le segment en cours (fin d'enregistrement)."""
|
|
seg = self._finish_segment() if self._triggered else None
|
|
self._ring.clear()
|
|
self._pending.clear()
|
|
return seg
|
|
|
|
def _process_frame(self, frame: bytes) -> Segment | None:
|
|
is_speech = self._vad.is_speech(frame, SAMPLE_RATE)
|
|
self._frame_index += 1
|
|
|
|
if not self._triggered:
|
|
self._ring.append((frame, is_speech))
|
|
voiced = sum(1 for _, s in self._ring if s)
|
|
if len(self._ring) == self._ring.maxlen and voiced >= TRIGGER_RATIO * self._ring.maxlen:
|
|
self._triggered = True
|
|
self._segment_start_frame = self._frame_index - len(self._ring)
|
|
self._segment = bytearray(b"".join(f for f, _ in self._ring))
|
|
self._speech_frames = voiced
|
|
self._silence_frames = 0
|
|
self._ring.clear()
|
|
return None
|
|
|
|
self._segment.extend(frame)
|
|
if is_speech:
|
|
self._speech_frames += 1
|
|
self._silence_frames = 0
|
|
else:
|
|
self._silence_frames += 1
|
|
|
|
too_long = len(self._segment) >= MAX_SEGMENT_S * SAMPLE_RATE * 2
|
|
ended = self._silence_frames * FRAME_MS >= SILENCE_END_MS
|
|
if ended or too_long:
|
|
return self._finish_segment()
|
|
return None
|
|
|
|
def _finish_segment(self) -> Segment | None:
|
|
self._triggered = False
|
|
seg, self._segment = self._segment, bytearray()
|
|
if self._speech_frames * FRAME_MS < MIN_SPEECH_MS:
|
|
return None
|
|
t0 = self._segment_start_frame * FRAME_MS / 1000
|
|
t1 = self._frame_index * FRAME_MS / 1000
|
|
return Segment(pcm=bytes(seg), t0=t0, t1=t1)
|