Transcribe Local

Runs faster-whisper (Whisper large-v3-turbo, int8 on CPU by default) on your own machine. Audio never leaves it.

Setup (one time)

pip install faster-whisper

ffmpeg must be on PATH. The first run downloads the model once (large-v3-turbo, about 1.6 GB; pass --model small for a small, faster one). If you already have a faster-whisper model in the Hugging Face cache, or the Codex desktop app's bundled faster-whisper runtime under ~/.cache/codex-runtimes, the script reuses it and nothing is installed.

Run it

python scripts/transcribe_local.py "<audio-or-video-file>" [--start T] [--end T] [--language en] [--model NAME_OR_DIR] [--out notes.txt]

• Output is [MM:SS] text lines on stdout (progress goes to stderr).
• --start / --end accept SS, MM:SS or HH:MM:SS; timestamps stay absolute.
• Language is auto-detected; pass --language xx to force it.
• --device cuda is opt-in and untested here; the default CPU path is the verified one.
• Speed on CPU is several times faster than real time. For a very long file, confirm with the user first, or transcribe a section.
• For word-level timestamps (karaoke captions) use talking-head-polish/scripts/word_times.py.

Limits

• Whisper can hallucinate on silence or music and mishears names and mixed-language speech; flag doubtful stretches instead of smoothing them over, and proofread captions against the audio.
• Romanised Tanglish or Hinglish is not transcribed well without a prompt (see word_times.py --prompts).
• No speaker labels.