c28e08b3aa
Cauza: am rulat `find transcripts -size -1k -delete` cat transcribe.py scria in acelasi director. faster_whisper scrie bufferat, deci o transcriere de o ora sta sub 1KB minute in sir; find a sters fisierul de sub proces, Python a logat "Transcribed" pe un inode disparut, si audio-ul a fost sters. Lectia 11 (25.11.2025) pierduta complet, re-descarcata. - transcribe_file() intoarce bool, verifica nr. cuvinte vs durata audio - pull_moltbot.sh: staging + mutare doar peste 1KB, in loc de find -delete - PROGRESS.md: documentat, ca sa nu se repete Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MtSTyTmt6AbCL9j5ajEDm1
88 lines
3.4 KiB
Python
88 lines
3.4 KiB
Python
"""
|
|
Transcribe audio/<module>/*.mp3 -> transcripts/<module>/*.txt using
|
|
faster-whisper (CPU, CTranslate2). Deletes the source audio after a
|
|
successful transcript (nobody wants the audio kept — text only).
|
|
Resumable: skips lectures with an existing transcript.
|
|
"""
|
|
|
|
import argparse
|
|
import logging
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
from faster_whisper import WhisperModel
|
|
|
|
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
|
|
log = logging.getLogger(__name__)
|
|
|
|
|
|
# Vorbirea normală dă ~100-150 cuvinte/minut. Sub 15% din asta înseamnă
|
|
# halucinație, tăcere, sau o scriere care a eșuat — nu o transcriere bună.
|
|
MIN_WORDS_PER_SECOND = 100 / 60 * 0.15
|
|
|
|
|
|
def transcribe_file(model: WhisperModel, audio_path: Path, txt_path: Path) -> bool:
|
|
"""Întoarce True doar dacă transcrierea rezultată pare completă.
|
|
Apelantul șterge audio-ul DOAR pe True — altfel un fișier pierdut
|
|
înseamnă re-descărcare, nu doar re-transcriere."""
|
|
segments, info = model.transcribe(str(audio_path), language="ro", beam_size=5, vad_filter=True)
|
|
txt_path.parent.mkdir(parents=True, exist_ok=True)
|
|
with open(txt_path, "w", encoding="utf-8") as f:
|
|
for seg in segments:
|
|
f.write(seg.text.strip() + "\n")
|
|
|
|
if not txt_path.exists():
|
|
log.error(f" {txt_path} a dispărut în timpul scrierii — păstrez audio-ul")
|
|
return False
|
|
words = len(txt_path.read_text(encoding="utf-8").split())
|
|
expected = info.duration * MIN_WORDS_PER_SECOND
|
|
if words < expected:
|
|
log.error(f" {txt_path.name}: doar {words} cuvinte pentru {info.duration:.0f}s "
|
|
f"(minim așteptat {expected:.0f}) — păstrez audio-ul")
|
|
return False
|
|
log.info(f" Transcribed: {txt_path.name} (lang={info.language}, duration={info.duration:.0f}s, {words} cuvinte)")
|
|
return True
|
|
|
|
|
|
def main():
|
|
p = argparse.ArgumentParser(description="Transcribe ATM course audio")
|
|
p.add_argument("--audio-dir", default="audio")
|
|
p.add_argument("--out", default="transcripts")
|
|
p.add_argument("--model", default="medium", help="faster-whisper model size")
|
|
p.add_argument("--only", default=None, help="Limit to one module dir name (for piloting)")
|
|
p.add_argument("--keep-audio", action="store_true", help="Don't delete source mp3 after transcribing")
|
|
args = p.parse_args()
|
|
|
|
audio_dir = Path(args.audio_dir)
|
|
out_dir = Path(args.out)
|
|
mp3_files = sorted(audio_dir.glob("*/*.mp3"))
|
|
if args.only:
|
|
mp3_files = [f for f in mp3_files if f.parent.name == args.only]
|
|
if not mp3_files:
|
|
log.error(f"No audio files found under {audio_dir}")
|
|
sys.exit(1)
|
|
|
|
log.info(f"Loading faster-whisper model '{args.model}' (CPU, int8)...")
|
|
model = WhisperModel(args.model, device="cpu", compute_type="int8")
|
|
|
|
done = skipped = 0
|
|
for audio_path in mp3_files:
|
|
txt_path = out_dir / audio_path.parent.name / (audio_path.stem + ".txt")
|
|
if txt_path.exists() and txt_path.stat().st_size > 50:
|
|
log.info(f" Skipping (exists): {txt_path.name}")
|
|
skipped += 1
|
|
continue
|
|
|
|
log.info(f"Transcribing: {audio_path}")
|
|
ok = transcribe_file(model, audio_path, txt_path)
|
|
if ok and not args.keep_audio:
|
|
audio_path.unlink()
|
|
done += 1
|
|
|
|
log.info("=" * 60)
|
|
log.info(f"Transcribed {done}, skipped {skipped}.")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|