Media Transcribe

Model strategy

• On a CUDA-capable machine, prefer model turbo, device cuda, compute type float16, and batched inference.
• On CPU-only machines, prefer model small, device cpu, compute type int8.
• If CUDA initialization fails, fall back to CPU automatically.

This is a pragmatic default: it favors high throughput on GPU while keeping CPU fallback usable.

Commands

The examples below assume the current directory is the skill root.

Prepare audio deterministically from a local video before transcription:

mkdir -p output/example/audio
.venv/bin/python scripts/prepare_audio.py --input "output/example/video/example.mp4" --output-dir output/example/audio

Transcribe a local audio file with automatic device/model selection:

mkdir -p downloads/media-transcribe/transcripts
.venv/bin/python subskills/media-transcribe/scripts/transcribe_audio.py --input "downloads/media-transcribe/audio/example.m4a" --output-dir downloads/media-transcribe/transcripts --language zh

Force CPU mode:

.venv/bin/python subskills/media-transcribe/scripts/transcribe_audio.py --input "downloads/media-transcribe/audio/example.m4a" --device cpu --compute-type int8 --model small

Output files

For an input named example.m4a, the script writes:

• example.txt
• example.srt
• example.transcription.json

Notes

• Prefer existing subtitle text over transcription; prefer transcription over danmaku when the user wants actual spoken content.
• Keep the raw audio file; do not delete it after transcription.
• Report the output paths and detected language back to the user.