mlx-audio

What it is and what it does

MLX-Audio is a speech synthesis and processing library built on Apple's MLX framework, designed to run inference for text-to-speech, speech-to-text, and speech-to-speech tasks directly on Apple Silicon Macs. It wraps multiple pre-trained model architectures (Kokoro, Qwen3-TTS, OmniVoice, and others) with support for multilingual input, voice customization, and quantization options (3-bit through 8-bit). The package exposes both a command-line interface and a Python API, letting you generate audio, control speech speed, clone voices, and stream results during generation.

The library depends on huggingface_hub for model downloads, mlx and mlx-lm for inference, numpy and scipy for audio processing, transformers for tokenization, sounddevice for playback, miniaudio for audio I/O, and tqdm for progress reporting. It targets developers building voice applications on macOS who want local, on-device inference without cloud dependencies.

Use it for:

• Generate speech from text on your Mac using multiple TTS model architectures with voice selection and language hints.
• Build a local voice assistant or chatbot that speaks without sending audio to a cloud service.
• Clone a voice by providing sample audio and then synthesize new speech in that voice across different models.
• Stream audio output during generation for real-time playback or save multi-segment outputs as a single joined file.
• Integrate speech synthesis into a macOS or iOS app via the Swift package or REST API.
• Experiment with different quantized model variants (4-bit, 8-bit) to balance quality and inference speed on Apple Silicon.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

MLX-Audio provides text-to-speech, speech-to-text, and speech-to-speech inference optimized for Apple Silicon, with support for multiple model architectures and multilingual capabilities.

Yes, if you have Apple Silicon and need local TTS/STS inference. Active maintenance, permissive license, low install friction, no known vulnerabilities, and a large model ecosystem make it a solid choice. Verify whether STT is production-ready and whether the web interface requires optional extras before committing to a deployment.

Install

mlx-audio on PyPI

pip install mlx-audio

uv add mlx-audio

poetry add mlx-audio

Installing mlx-audio

Before you install

Low friction—pure Python wheel with no compiled dependencies. Active maintenance: last commit 2026-08-13, 7721 stars, released 2026-08-10 (4 days old). Requires Python >=3.10.

License in practice

MIT license (permissive). You can use, modify, and distribute this package freely in commercial and private projects with minimal restrictions.

Quickstart

pip install mlx-audio

from mlx_audio.tts.utils import load_model

model = load_model("mlx-community/Kokoro-82M-bf16")
for result in model.generate("Hello from MLX-Audio!", voice="Chelsie"):
    print(f"Generated {result.audio.shape[0]} samples")

Requires Python >=3.10 and Apple Silicon (M series) for optimal performance; runtime dependencies include mlx, numpy, scipy, transformers, huggingface_hub, and sounddevice.

Verify before relying

• Whether STT (speech-to-text) is fully implemented or if the package currently focuses primarily on TTS and STS.
• Performance characteristics and latency on different M-series chip generations.
• Whether the web interface and OpenAI-compatible API server are included in the base pip install or require the dev/server extras.

Package facts

Evidence: mlx_audio-0.4.8-py3-none-any.whl

Tags

More Artificial Intelligence packages

Further reading