Apple Silicon's audio Swiss Army knife

What it does

MLX-Audio wraps a sprawling zoo of text-to-speech, speech-to-text, and speech-to-speech models under Apple’s MLX framework. The pitch is simple: pick a model, run it fast on M-series chips, optionally quantize it down to 3-bit, and expose it through an OpenAI-compatible API or a web UI with 3D audio visualization. It also ships a Swift package for iOS/macOS embedding.

The interesting bit

The breadth is almost comical. The README lists 18 TTS architectures (Kokoro, Qwen3-TTS, Voxtral, a 7B diffusion model called KugelAudio), 13 STT variants including Whisper, Moonshine, and IBM’s Granite Speech, plus VAD and diarization. Someone clearly enjoys porting models. The “Mega-ASR” model even auto-routes between clean and degraded audio modes with LoRA switching — a nice touch of actual engineering amid the model-collecting.

Key highlights

• Runs fully offline on Apple Silicon with aggressive quantization (3-bit through 8-bit)
• OpenAI-compatible REST API means drop-in replacement for cloud TTS/STT
• Voice cloning, speaker diarization, streaming generation, and speed control
• Swift package for native iOS/macOS integration
• 40+ models including obscure entries like “Ming Omni TTS” that generates speech, music, and sound events

Caveats

• README claims “the best audio processing library” without benchmarks to back it up
• Many models link to HuggingFace repos with no performance numbers or comparison tables
• The web UI’s “3D audio visualization” sounds like chrome; no screenshot or demo is shown

Verdict

Grab this if you want to prototype voice features on a MacBook without cloud bills or API keys. Skip it if you’re on CUDA hardware — MLX is Apple-only, and you’ll get better ecosystem support elsewhere.