anibon-local-transcription

Anibon Local Audio Transcription

Overview & Triggers

Use when YouTube has no subtitles or auto-captions for the target video.

1. Audio Extraction

Download audio stream directly as mono 16kHz 16-bit PCM WAV (avoids temporary files and >4GB WAV header overflow on long 8h+ streams):

# Direct single-pass extraction via yt-dlp + ffmpeg:
yt-dlp -f ba -x --audio-format wav --postprocessor-args "ExtractAudio:-ar 16000 -ac 1 -c:a pcm_s16le" -o "audio_16k.%(ext)s" "VIDEO_URL"

[!NOTE] If YouTube returns HTTP Error 403: Forbidden, update yt-dlp first:

yt-dlp --update

Or convert existing audio:

ffmpeg -i audio.wav -ar 16000 -ac 1 -c:a pcm_s16le audio_16k.wav

2. Local Transcription

Run GPU-accelerated whisper.cpp build:

Check for whisper-cli in system PATH or local $HOME/whisper.cpp build directory:

• $HOME/whisper.cpp/build/bin/whisper-cli (or .exe on Windows)
• $HOME/whisper.cpp/main

Model path defaults to $HOME/whisper.cpp/models/ggml-large-v3-turbo.bin.

Execution:

# Linux/macOS
whisper-cli -m models/ggml-large-v3-turbo.bin -f audio_16k.wav -l th -ot 240000 -t 8 --output-json -of whisper_output

# Windows (Vulkan Device 0: RX 7600, or Device 1: Tesla P100):
C:\Users\peter\whisper.cpp\build\bin\whisper-cli.exe -m C:\Users\peter\whisper.cpp\models\ggml-large-v3-turbo.bin -f audio_16k.wav -ot 240000 -l th -t 8 --output-json -of whisper_output --print-progress

(-ot 240000 offset skips the first 4 minutes of standby/intro silence to avoid repetition loop bugs on silence).

For full build options and platform configurations, see BUILD_WHISPERCPP_GUILD.md.

3. Format Conversion & Windows Invariants

Convert whisper_output.json directly to pipeline-standard raw_transcript.json:

[!IMPORTANT] Windows UTF-8 Invariant:

1. Windows default console encoding (cp1252/charmap) crashes when printing Thai text. Always run scripts with python -X utf8.
2. MSVC whisper-cli output can split multibyte Thai UTF-8 characters across segment buffer cuts. Always decode JSON with errors="replace".

Run inline Python converter:

python -X utf8 -c "
import json

def fmt_ts(seconds):
    h = int(seconds // 3600)
    m = int((seconds % 3600) // 60)
    s = int(seconds % 60)
    return f'{h:02d}:{m:02d}:{s:02d}'

with open('whisper_output.json', 'rb') as f:
    data = json.loads(f.read().decode('utf-8', errors='replace'))

items = []
for seg in data.get('transcription', []):
    text = seg.get('text', '').strip()
    if not text:
        continue
    offsets = seg.get('offsets', {})
    start_sec = offsets.get('from', 0) / 1000.0
    end_sec = offsets.get('to', 0) / 1000.0
    items.append({
        'text': text,
        'start': round(start_sec, 2),
        'duration': round(end_sec - start_sec, 2),
        'timestamp': fmt_ts(start_sec)
    })

with open('raw_transcript.json', 'w', encoding='utf-8') as out:
    json.dump(items, out, ensure_ascii=False, indent=2)
"

4. Hallucination Detection & Recovery

On livestreams $\ge 2\text{h}$, Whisper may loop on music or silence. Run parallel Divide-and-Conquer recovery:

python -X utf8 ..\whisper-corruption-recovery\scripts\fix_hallucinations.py whisper_output.json audio_16k.wav --devices 0 -w 3 -o raw_transcript.json

Then run chunking with prepare_video.py:

python -X utf8 scripts\prepare_video.py <VIDEO_URL> --workspace <WORKSPACE> --format txt --block 300 --overlap 30