MLX-Audio
Open-source library for running text-to-speech, speech-to-text, speech-to-speech and music generation models locally on Apple Silicon using MLX.
At a Glance
MLX-Audio library for running TTS, STT, STS and music generation models on Apple Silicon, released under the MIT License.
Engagement
Available On
Listed Nov 2026
About MLX-Audio
MLX-Audio is an open-source Python library from Neywa Labs for running audio models on Apple Silicon Macs. It covers text-to-speech, speech-to-text, speech-to-speech and music generation, all built on Apple's MLX framework and running locally.
It is installed with pip or uv, used through command-line tools or a Python API, and can also be served through a local web interface and OpenAI-compatible API.
What It Is
MLX-Audio is a model-runtime library for audio workloads on Apple Silicon. Developers load models, mostly MLX-converted checkpoints hosted on Hugging Face, and generate speech, transcribe audio, separate or enhance sound, or generate music. Neywa Labs describes itself as the local AI company behind the Nativ app, and lists mlx-audio among the libraries it builds.
Supported Model Families
The README documents a broad set of models by task:
- Text-to-speech: Kokoro, Qwen3-TTS, Chatterbox, Dia, OuteTTS, Spark, OmniVoice, Voxtral TTS, KugelAudio and others, with voice cloning and multilingual support on many of them.
- Speech-to-text: Whisper, Qwen3-ASR, Parakeet, Voxtral Realtime, VibeVoice-ASR, Granite Speech and more, including diarization, timestamps and streaming variants.
- Voice activity detection and diarization: Silero VAD and Sortformer variants.
- Speech-to-speech: source separation (SAM-Audio), speech enhancement (MossFormer2, DeepFilterNet) and conversational audio models.
- Music generation: MiniMax Music 3.
Workflow and Outputs
The command line tool mlx_audio.tts.generate produces audio files, can play them immediately, stream during generation, or join segments into one file. The Python API returns waveforms as MLX arrays and transcription results with text, segments and timestamps. A server started with mlx_audio.server exposes /v1/audio/speech and /v1/audio/transcriptions endpoints. A convert script supports quantization (affine, mxfp4, mxfp8, nvfp4) and dtype conversion, with optional upload to Hugging Face.
Setup Path
It requires Python 3.10+ and an Apple Silicon Mac. ffmpeg is needed for MP3, FLAC, OGG, Opus and Vorbis output, while WAV works without it. A separate mlx-audio-swift package targets macOS and iOS integration. The code is MIT licensed.
Community Discussions
Be the first to start a conversation about MLX-Audio
Share your experience with MLX-Audio, ask questions, or help others learn from your insights.
Pricing
Open Source (MIT)
MLX-Audio library for running TTS, STT, STS and music generation models on Apple Silicon, released under the MIT License.
- MIT License
- Text-to-speech, speech-to-text, speech-to-speech and music generation
- Optimized for Apple Silicon (MLX)
- OpenAI-compatible REST API
- Quantization support (3-bit, 4-bit, 6-bit, 8-bit, and more)
Capabilities
Key Features
- Text-to-speech, speech-to-text, speech-to-speech and music generation
- Voice cloning and voice customization
- Multilingual model support
- Adjustable speech speed
- Streaming audio generation and transcription
- Speaker diarization and word-level timestamps
- Quantization (3-bit to 8-bit, mxfp4, mxfp8, nvfp4)
- OpenAI-compatible REST API
- Web interface with 3D audio visualization
- Command-line tools and Python API
- Swift package for iOS/macOS
