mlx-audio-server

# MLX Audio Server

`mlx-audio`: The best audio processing library built on Apple's MLX framework, providing fast and efficient text-to-speech (TTS), speech-to-text (STT), and speech-to-speech (STS) on Apple Silicon.

This skill will run it as a OpenAI-compatible API server on macOS in background, and provide scripts/examples for AI agents to use the api.

Default Models:

- Speech-To-Text: `mlx-community/glm-asr-nano-2512-8bit`
- Text-To-Speech: `mlx-community/Qwen3-TTS-12Hz-1.7B-VoiceDesign-bf16`

The server will download these models when needed, so first run will be a bit slow.

More choices here: https://github.com/Blaizzy/mlx-audio?tab=readme-ov-file#supported-models

## Requirements

- `mlx`: macOS with Apple Silicon
- `brew`: used to install deps if not available

## Installation

```bash
bash ${baseDir}/install.sh
```
This script will:
- clone (forked) mlx-audio repo into `~/opt/mlx-audio`
- use `uv` to create a venv and install deps in it: `~/opt/mlx-audio/.venv`
- create a plist file to run mlx-audio server as a launchd service in background in user domain
- run as a OpenAI compatible API server, on port 8899 by default.

## Usage

STT/Speech-To-Text:
```bash
# input will be converted to wav with ffmpeg, if not yet.
# output will be transcript text only.
bash ${baseDir}/run_stt.sh <audio_or_video_path>
```

TTS/Text-To-Speech:
```bash
# audio will be saved into a tmp dir, with default name `speech.wav`, and print to stdout.
bash ${baseDir}/run_tts.sh "Hello, Human!"
# or you can specify a output dir
bash ${baseDir}/run_tts.sh "Hello, Human!" ./output
# output will be audio path only.
```
You can use both scripts directly, or as example/reference.