Recommended Models
Models are downloaded automatically from Hugging Face (ggerganov/whisper.cpp) the first time you request them, verified atomically via a .part file, and cached in your user cache directory.
| Model | Approx. Size | Speed | Accuracy | Best For |
|---|---|---|---|---|
tiny / tiny.en |
~75 MB | Fastest | Low | Quick drafts, low-RAM machines |
base / base.en |
~140 MB | Fast | Medium (default) | Everyday videos, fast turnarounds |
small / small.en |
~460 MB | Medium | Good | Podcasts, talks, tutorials |
medium / medium.en |
~1.5 GB | Slow | Great | Noisy audio, accents, multilingual |
turbo (large-v3-turbo) |
~1.5 GB | Fast | Near-best | Best overall speed-to-accuracy ratio |
large (large-v3) |
~3.0 GB | Slowest | Best | Highest accuracy, archival transcription |
Quantized & Specialized Variants
Quantized models (-q5_0, -q5_1, -q8_0) reduce disk size and memory usage by 30–60% with minimal quality loss — ideal for laptops and parallel batch workers.
| Family | Available Identifiers |
|---|---|
| Tiny | tiny, tiny.en, tiny-q5_1, tiny.en-q5_1, tiny-q8_0, tiny.en-q8_0 |
| Base | base, base.en, base-q5_1, base.en-q5_1, base-q8_0, base.en-q8_0 |
| Small | small, small.en, small.en-tdrz, small-q5_1, small.en-q5_1, small-q8_0, small.en-q8_0 |
| Medium | medium, medium.en, medium-q5_0, medium.en-q5_0, medium-q8_0, medium.en-q8_0 |
| Large v1/v2 | large-v1, large-v2, large-v2-q5_0, large-v2-q8_0 |
| Large v3 | large-v3, large-v3-q5_0, large (alias), large-q5_0 (alias) |
| Large v3 Turbo | large-v3-turbo, large-v3-turbo-q5_0, large-v3-turbo-q8_0, turbo (alias) |
Managing Models
List all available models and see which ones are already cached locally:
subtitle models --list
Pre-download a model before going offline:
subtitle models --download large-v3-turbo-q5_0
Use a custom cache directory (per-command or via environment variable):
export SUBTITLE_MODELS_DIR=/mnt/fast-ssd/whisper-models
subtitle video.mp4 --model turbo