Whisper Low Memory Module Wired Up

Whisper Low Memory Module Wired Up | Generated by AI

Home 2026.08

All done. ww whisper <file> --low-memory is wired up. Summary of changes:

Changes

1. New module — ww/audio/whisper_low_mem.py Adapted your transcribe.py memory-safe loader (CPU checkpoint → GPU fp16 only, ~3.2 GB for large-v3) into a proper main() module. whisper is imported lazily with a friendly error if openai-whisper isn’t installed. CLI:

ww whisper file.mp4 --low-memory [--model large-v3] [--language zh] [--task transcribe|translate] [--device cuda]

Writes <stem>.txt/.srt/.json/.vtt/.tsv next to the audio, same as your script.

2. ww/main.py

Routing: --low-memory anywhere in argv is stripped and dispatched to whisper_low_mem (before the default whisper_translate fallback).
Help text: added ww whisper <file> --low-memory line under the Whisper section.

3. ~/.zsh/completions/_ww The whisper file-completion case now also offers --low-memory, --model, --language, --task, --device flags. (Reload with autoload -Uz compinit && compinit.)

4. pyproject.toml Added openai-whisper>=20240930 — uv sync already run, installed openai-whisper==20250625 cleanly (torch was already present via whisperx).

Verified

uv run ruff check — pass; pyright — 0 errors
uv run ww whisper --low-memory → routes correctly (argparse shows the new module’s usage)
Missing-file error path works
zsh -n completion syntax OK

Note: the model must be cached at ~/.cache/whisper/large-v3.pt (download once with whisper large-v3); the loader will tell you if it’s missing. The full transcription path needs your CUDA GPU, so test it on the workstation.

Back Donate