| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-07-24 | 897 Bytes | |
| voxtype v0.1.4 -- fast first transcription (MLX warmup) source code.tar.gz | 2026-07-24 | 55.5 MB | |
| voxtype v0.1.4 -- fast first transcription (MLX warmup) source code.zip | 2026-07-24 | 55.8 MB | |
| Totals: 3 Items | 111.4 MB | 0 | |
Fixes the slow first take
On the parakeet-mlx engine, the first dictation after launching used to take several seconds (then everything sped up). Cause: MLX compiles the model's Metal kernels on the first decode of a fresh process (~3-10s); every decode after is sub-second.
voxtype now warms the decoder up during startup — a throwaway decode runs on the worker thread while you still see "loading …", so that one-time compile happens off your first real take.
Result: the first real decode is now ~0.15s instead of ~3s; steady-state is unchanged (~0.12-0.4s, 16-52x realtime). The earlier MLX memory-cache fix was confirmed not the cause (decode timing is identical with and without it).
Upgrade:
:::bash
uv tool install --force --refresh voxtype
Full changelog: https://github.com/pchalasani/claude-code-tools/compare/voxtype-v0.1.3...voxtype-v0.1.4