subsprint — offline subtitle generator for any video

CLI that creates .srt subtitles from a video file, fully offline after the first model download. Tuned for Apple Silicon (M2 Pro tested).

Install

cd ~/Developers/projects/subsprint
pip install -e .
# optional, faster on M-series via Metal:
pip install -e ".[mlx]"
# optional, for `translate` command:
pip install -e ".[translate]"

Requires system ffmpeg (brew install ffmpeg).

Use

# transcribe English audio to English subs
# (pass -l en for English audio, omit -l for auto-detect)
subsprint transcribe movie.mp4 -l en

# higher quality, still <10 min with Metal backend:
subsprint transcribe movie.mp4 -l en --backend mlx --model small

# best quality offline (slower on CPU, faster on Metal):
subsprint transcribe movie.mp4 -l en --model large-v3-turbo --beam-size 1

# custom output dir / suffix / format
subsprint transcribe movie.mp4 -l en -o ./subs --suffix .en.auto --format srt

# translate an existing .srt (e.g. Italian -> English), keeps timings
subsprint translate subs.srt --from it --to en -o subs.eng.srt

Outputs by default sit next to the video: <video-stem>.<lang>.auto.srt (e.g. movie.en.auto.srt).

Example timings (English subs, measured on M2 Pro 10-core / 16 GB)

Fast defaults: model small, beam_size=1, VAD on, faster-whisper CPU backend.

Video length	Audio language	Model / backend	Subs generation time
1 min	English	small / faster-whisper (CPU)	~5 s transcribe (+ one-time model load)
10 min	English	small / faster-whisper (CPU)	~50 s (scales linearly)
98 min	English	small / faster-whisper (CPU)	~5 min transcribe + ~20 s audio extract, 1,474 segments

Notes:

• First run downloads the model once (small ≈ 500 MB); afterwards 100 % offline.
• Model load is extra on first run only (~45 s with download, ~1 s cached).
• tiny is faster but noticeably less accurate; medium / large-v3-turbo are more accurate and slower (turbo is much faster on the Metal mlx backend).
• mlx-whisper (--backend mlx) uses the Metal GPU: ~3 s inference per 60 s of audio with small after the one-time model download.
• translate uses Helsinki-NLP/opus-mt-<src>-<tgt>; first run downloads it too.