File size: 2,687 Bytes
c3b1320 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 | ---
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).
|