Whisper Transcribe Transcribe

whisper-transcribe-transcribe

Internal transcription backend. Wraps the ~/.local/bin/whisper-transcribe binary built by /connect0459-agent-skills:whisper-transcribe-setup. Not intended for direct user invocation — called by /connect0459-agent-skills:youtube-transcribe-local.

CLI

whisper-transcribe <input.wav>
    [--format srt|txt|vtt|json]   default: srt
    [--model NAME]                default: large-v3-turbo
    [--output-prefix PATH]        default: next to input (stem only, no extension)
    [--language CODE]             default: auto-detect
    [--gpu|--no-gpu]              default: auto (GPU outside sandbox, CPU inside)

Behaviour

• Input must be a 16 kHz mono PCM WAV — produced by ffmpeg in youtube-transcribe-local.
• Model must exist at ~/.local/share/whisper-transcribe/models/ggml-<model>.bin.
• For --format srt|vtt|json: writes <prefix>.<ext> and prints wrote <path>.
• For --format txt: writes plain transcript to stdout.
• GPU auto-detection: CPU inside Claude Code sandbox (SANDBOX_RUNTIME=1 or CODEX_SANDBOX=seatbelt), GPU otherwise. Override with --gpu or --no-gpu.

Missing install

If the binary or model is missing, the binary exits with a hint to run /connect0459-agent-skills:whisper-transcribe-setup.