---

name: tts-generation description: "Use MLX Audio TTS workflows for generation, voice cloning, streaming, batching, and model-specific controls." disable-model-invocation: true metadata: disco-role: operating license: MIT

TTS Generation

Use this sub-skill when the user wants to synthesize speech from text, clone a voice from a reference clip, stream audio during generation, or choose the right TTS model family and flags.

Route Here For

• Basic text-to-speech generation from CLI or Python.
• Voice cloning with ref_audio and optional ref_text.
• Streaming output, chunk joining, save behavior, and playback planning.
• Model-family questions for Kokoro, Qwen3-TTS, OmniVoice, Higgs Audio, CSM / MisoTTS, Spark, Ming Omni, Voxtral TTS, Chatterbox, and other TTS models shipped by the package.
• Safe command planning for mlx_audio.tts.generate.

Route Elsewhere

• For transcription, forced alignment, streaming ASR, or WER, use ../stt-transcription/.
• For enhancement, separation, VAD, or audio I/O, use ../speech-transforms-vad/.
• For server endpoints or conversion, use ../server-and-conversion/.

Fast Paths

• See references/workflows.md for the main TTS recipes.
• See references/api-reference.md for stable kwargs and model-loading facts.
• See references/troubleshooting.md for import, playback, and cloning failures.
• Use scripts/tts_command_builder.py to shape a safe CLI command before a long run.

Default Safety Policy

Prefer a small reproducible recipe first: confirm the model id, confirm the reference clip path, confirm whether the model requires ref_text, and then decide whether streaming or file saving is actually needed. If the task is only about a command plan, do not start an inference run.