Fixed low-memory whisper import bug | Generated by AI
Fixed and verified. Here’s the full picture:
Root cause
ww is installed as a global console script (/home/lzw/.local/bin/ww) running under system python3.11, while the ML deps live in the project venv. The plain ww whisper worked because it shells out to the whisper CLI (installed in system python), but --low-memory did import whisper inside the ww interpreter — where openai-whisper was never installed. (ww whisper diarize had the same latent bug with whisperx.)
What I changed
ww/audio/whisper_low_mem.py — rewrote the low-memory path to use faster-whisper instead of openai-whisper, and added a _ensure_deps() guard that re-execs the script with the project venv python if deps are missing in the running interpreter. Same CLI surface (--model/--language/--task/--device) plus new --compute-type (auto/float16/int8) and --no-proxy (HF downloads can stall through your local proxy — same reason diarize unsets it). Still writes .txt/.srt/.vtt/.tsv/.json like before.
ww/audio/whisper_diarize.py — added the same venv re-exec guard (would have crashed the same way).
ww/main.py — updated the --low-memory help text.
Verified end-to-end through the real ww entry point; ruff/pyright clean (only pre-existing findings remain).
Your questions
Why was openai-whisper needed? Only for the default ww whisper path — it invokes the openai-whisper CLI, which you have installed globally, so that path works. It’s no longer needed for --low-memory.
Why not whisperx for low-memory? whisperx is faster-whisper under the hood — it’s just faster-whisper plus wav2vec2 word-level alignment (+~1.2 GB VRAM and an extra download) plus optional pyannote diarization (HF token + models). Plain transcription needs none of that; the engine alone is the memory-efficient part, so that’s what --low-memory now uses directly. Use ww whisper diarize when you actually want speaker labels.
Note: large-v3 will now be downloaded from HuggingFace (Systran/faster-whisper-large-v3, ~3 GB) on first run — separate cache from your existing ~/.cache/whisper/*.pt files. If the download stalls behind your proxy, add --no-proxy.
One heads-up: your first real run with --low-memory on ruixiu2.mp4 will be slower only because of that model download.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋