Metadata-Version: 2.4
Name: notepipe-asr-speech-swift
Version: 0.1.0
Classifier: Operating System :: MacOS :: MacOS X
Summary: Speech recognition adapter for notepipe on Apple silicon, using speech-swift (MLX and Core ML)
License-Expression: MIT
Requires-Python: >=3.9
Description-Content-Type: text/markdown; charset=UTF-8; variant=GFM
Project-URL: Repository, https://github.com/3ae3ae/notepipe

# notepipe-asr-speech-swift

A [notepipe](https://github.com/3ae3ae/notepipe) speech adapter that runs speech recognition locally on Apple silicon
with [speech-swift](https://github.com/soniqo/speech-swift): Qwen3-ASR 1.7B on MLX by default, or Qwen3-ASR 0.6B and
Whisper on Core ML.

This adapter is built in CI but has not yet been verified on hardware.

## Install

macOS 15 or later on Apple silicon:

```sh
uv tool install "notepipe[speech-swift]"
```

The package installs `mlx.metallib` next to the executable, where MLX looks for it. Building from source needs Xcode with
its Metal toolchain (`xcodebuild -downloadComponent MetalToolchain`); a plain `cargo install` does not copy
`mlx.metallib`.

## Use

```toml
[speech]
adapter = ["notepipe-asr-speech-swift"]
concurrency = 1
```

`--model` selects `qwen3-1.7b` (default), `qwen3-0.6b` or `whisper`. The first run downloads the model.
`[transcript].language` is passed to the model.

The Qwen models write at most 448 tokens per segment, roughly a minute of speech. Keep
`[speech.detection].max_segment` at its default of 20 seconds, or below 60, so that long segments are not cut short.

