2697e01714
- новый api/: /pipeline/process (yt-dlp, субтитры или Vosk, LLM/экстрактивная суммаризация), /videos, /llm; старый код перенесён в api_legacy/ - web/: Flet UI (flet 0.28.3 + flet-web, порт 8550) - Dockerfile.api: ffmpeg слоем из mwader/static-ffmpeg, pip с кэш-маунтом - requirements/pyproject: убраны moviepy, imageio-ffmpeg, битый asyncio - README с инструкцией запуска и загрузкой Vosk-модели Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
120 lines
3.4 KiB
Python
120 lines
3.4 KiB
Python
"""Vosk-based fallback transcription: stream PCM from ffmpeg into a Vosk recognizer."""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import logging
|
|
import os
|
|
import shutil
|
|
import subprocess
|
|
import sys
|
|
import threading
|
|
from pathlib import Path
|
|
|
|
try:
|
|
from vosk import KaldiRecognizer, Model
|
|
except Exception:
|
|
Model = None
|
|
KaldiRecognizer = None
|
|
|
|
|
|
logger = logging.getLogger("transcribe")
|
|
|
|
|
|
class VoskModelManager:
|
|
"""Loads a Vosk model. Captures the model's stderr to Python logs while loading."""
|
|
|
|
def __init__(self, model_path: Path):
|
|
if Model is None or KaldiRecognizer is None:
|
|
raise RuntimeError("vosk is not installed in runtime")
|
|
self.model = self._load_with_stderr_capture(model_path)
|
|
|
|
@staticmethod
|
|
def _load_with_stderr_capture(path: Path):
|
|
log = logging.getLogger("vosk.loader")
|
|
r_fd, w_fd = os.pipe()
|
|
saved_stderr = os.dup(sys.stderr.fileno())
|
|
os.dup2(w_fd, sys.stderr.fileno())
|
|
os.close(w_fd)
|
|
|
|
def reader(fd: int) -> None:
|
|
with os.fdopen(fd, "rb") as fh:
|
|
for raw in iter(fh.readline, b""):
|
|
log.info(raw.decode(errors="ignore").rstrip())
|
|
|
|
t = threading.Thread(target=reader, args=(r_fd,), daemon=True)
|
|
t.start()
|
|
try:
|
|
return Model(str(path))
|
|
finally:
|
|
try:
|
|
os.dup2(saved_stderr, sys.stderr.fileno())
|
|
finally:
|
|
os.close(saved_stderr)
|
|
t.join(timeout=2)
|
|
|
|
|
|
def transcribe_via_ffmpeg(media_path: Path, model_path: Path, sample_rate: int = 16000) -> str:
|
|
"""Decode the source media to PCM s16le mono via ffmpeg and feed it to Vosk in chunks."""
|
|
if Model is None or KaldiRecognizer is None:
|
|
raise RuntimeError("vosk is not installed in runtime")
|
|
if shutil.which("ffmpeg") is None:
|
|
raise RuntimeError("ffmpeg is not available in PATH")
|
|
if not media_path.exists():
|
|
raise FileNotFoundError(media_path)
|
|
|
|
mgr = VoskModelManager(model_path)
|
|
recognizer = KaldiRecognizer(mgr.model, sample_rate)
|
|
try:
|
|
recognizer.SetWords(True)
|
|
except Exception:
|
|
pass
|
|
|
|
cmd = [
|
|
"ffmpeg",
|
|
"-hide_banner",
|
|
"-loglevel",
|
|
"error",
|
|
"-i",
|
|
str(media_path),
|
|
"-f",
|
|
"s16le",
|
|
"-acodec",
|
|
"pcm_s16le",
|
|
"-ac",
|
|
"1",
|
|
"-ar",
|
|
str(sample_rate),
|
|
"-vn",
|
|
"-",
|
|
]
|
|
logger.info("Spawning ffmpeg streaming decode for %s", media_path.name)
|
|
|
|
proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
|
|
if proc.stdout is None:
|
|
proc.kill()
|
|
raise RuntimeError("ffmpeg did not provide stdout")
|
|
|
|
try:
|
|
bytes_read = 0
|
|
last_logged_mb = 0
|
|
while True:
|
|
chunk = proc.stdout.read(4000)
|
|
if not chunk:
|
|
break
|
|
recognizer.AcceptWaveform(chunk)
|
|
bytes_read += len(chunk)
|
|
mb = bytes_read // (1024 * 1024)
|
|
if mb >= last_logged_mb + 5:
|
|
last_logged_mb = mb
|
|
logger.info("Transcribed %d MB of audio so far", mb)
|
|
|
|
result = json.loads(recognizer.FinalResult())
|
|
text = result.get("text", "")
|
|
logger.info("Transcription finished, length=%d", len(text))
|
|
return text
|
|
finally:
|
|
try:
|
|
proc.kill()
|
|
except Exception:
|
|
pass
|