| --- |
| license: apache-2.0 |
| library_name: mlx-audio |
| pipeline_tag: text-to-speech |
| tags: |
| - mlx |
| base_model: Zyphra/ZONOS3-Mini |
| --- |
| |
| # ZONOS3-Mini |
|
|
| [ZONOS3-Mini](https://huggingface.co/Zyphra/ZONOS3-Mini) is Zyphra's multilingual |
| text-to-speech model with 100 million active parameters. It generates 44.1 kHz |
| mono audio with 27 built-in voices, streaming, and voice cloning from reference |
| audio without a transcript. |
|
|
| ## Supported Models |
|
|
| - [mlx-community/ZONOS3-Mini](https://huggingface.co/mlx-community/ZONOS3-Mini) |
|
|
| ## CLI Example |
|
|
| ```bash |
| python -m mlx_audio.tts.generate \ |
| --model mlx-community/ZONOS3-Mini \ |
| --text "Hello from ZONOS3 Mini!" \ |
| --voice jett |
| ``` |
|
|
| Use `--ref_audio reference.wav` to clone a voice, or add `--stream` for |
| streaming playback. |
|
|
| ## Python Example |
|
|
| ```python |
| from mlx_audio.audio_io import write as audio_write |
| from mlx_audio.tts import load |
| |
| model = load("mlx-community/ZONOS3-Mini") |
| |
| result = next(model.generate( |
| text="Hello from ZONOS3 Mini!", |
| voice="jett", |
| )) |
| audio_write("speech.wav", result.audio, result.sample_rate) |
| |
| # Voice cloning does not require a reference transcript. |
| result = next(model.generate( |
| text="Welcome back.", |
| ref_audio="reference.wav", |
| )) |
| audio_write("cloned.wav", result.audio, result.sample_rate) |
| ``` |
|
|
| Use `model.voices` to list the included voices. `voice="none"` generates without |
| a speaker prompt. Reference audio paths are resampled automatically; array |
| inputs must be mono PCM at 44,100 Hz. `speaker_embedding` accepts a |
| 192-dimensional vector or a NumPy file, and `speaker_cfg_scale` controls |
| speaker guidance (default `1.0`). |
|
|
| Set `stream=True` for incremental synthesis, with `streaming_interval` in |
| seconds. `speed` controls speaking rate. Acoustic controls are available through |
| `condition_values={"loudness": -20, "fmax": 20000, "speaking_rate": 16}` or |
| `condition_buckets`. `max_tokens` limits generated delayed audio rows at 15 Hz. |
|
|
| Lightweight English number normalization is enabled by default; other languages |
| pass through as written. Set `text_normalization=False` for text already |
| normalized by another frontend. Speech editing, accent adapters, hidden-state |
| steering, and raw reference-code encoding are not exposed by this implementation. |
|
|
| ## License |
|
|
| Model weights are released under [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0). |
| The backbone, codec, and speaker encoder are adapted from |
| [Zyphra's MIT-licensed reference implementation](https://github.com/Zyphra/ZONOS3-Mini/blob/6ff97cac0c684732de5f2bc6e3671aa783c15054/LICENSE). |
| The Gated DeltaNet kernel includes MIT-licensed code by Apple, also available |
| in [mlx-lm](https://github.com/ml-explore/mlx-lm/blob/main/LICENSE). |
|
|
