mlx-audio-server_skill

Overview

This skill runs a local, always-on OpenAI-compatible API server for speech tasks on macOS with Apple Silicon. It uses the MLX audio stack to provide fast, low-latency STT (speech-to-text), TTS (text-to-speech), and speech-to-speech capabilities. Installation automates dependencies and configures a LaunchAgent so the service runs continuously.

How this skill works

The server exposes OpenAI-style endpoints locally and delegates heavy lifting to MLX audio models optimized for Apple Silicon. Installation scripts install Homebrew formulae, ffmpeg, and required utilities, then register the mlx_audio.server as a macOS service. Simple shell wrappers are provided to run STT and TTS operations via the local API or directly as example commands.

When to use it

• You need a private, offline-capable speech API on a Mac with Apple Silicon.
• Low-latency local STT/TTS for development, demos, or personal assistants.
• Batch transcription or programmatic TTS without cloud costs or data egress.
• Running continuous services that must restart automatically on macOS.
• Prototyping voice features while keeping audio data on-device.

Best practices

• Run on Apple Silicon macOS only — MLX models are built and optimized for this hardware.
• Use the provided install script to ensure ffmpeg and jq are available and the LaunchAgent is configured.
• Keep model and service updates managed via Homebrew formulae for reliable upgrades.
• Route input files through the wrapper scripts to ensure proper audio conversion to WAV.
• Test with small samples first to confirm model selection and output format before scaling.

Example use cases

• Transcribe meetings or interviews locally using the STT wrapper with glm-asr-nano by default.
• Generate natural-sounding speech for a desktop assistant using the bundled Qwen3-TTS model.
• Integrate the local API into a home automation system to avoid sending audio to external services.
• Batch-process a folder of videos into transcripts for archiving or compliance.
• Prototype a speech-to-speech pipeline that transforms voice inputs into synthesized outputs on-device.