Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
68 lines
3 KiB
Markdown
68 lines
3 KiB
Markdown
# Maya1
|
|
|
|
`maya1` provides native offline inference for
|
|
[Maya Research Maya1](https://huggingface.co/maya-research/maya1), an English
|
|
text-to-speech model with natural-language voice design and inline emotion
|
|
controls. Audio is decoded at 24 kHz with the SNAC speech codec.
|
|
|
|
## Install
|
|
|
|
```bash
|
|
python tools/model_manager_v2.py install maya1_orig --models-root models
|
|
# Or install the smaller Q8_0 package:
|
|
python tools/model_manager_v2.py install maya1_q8_0 --models-root models
|
|
```
|
|
|
|
## Run
|
|
|
|
```bash
|
|
audiocpp_cli --task tts --family maya1 \
|
|
--model models/Maya1-GGUF/maya1-orig.gguf \
|
|
--backend cuda \
|
|
--text "Welcome back. <laugh> Today is going to be interesting." \
|
|
--request-option 'instruct=Warm female voice in her 30s with an American accent and conversational pacing.' \
|
|
--out maya1.wav
|
|
```
|
|
|
|
`--task vdes` is an alias to the same instruction-conditioned synthesis path.
|
|
|
|
Emotion tags such as `<laugh>`, `<whisper>`, `<sigh>`, and `<gasp>` may be
|
|
placed directly in the input text. See the upstream Maya1 model card for its
|
|
full prompt and emotion-tag guidance.
|
|
|
|
### Request options (use with `--request-option`)
|
|
|
|
| Option | Values | Default | Description |
|
|
|---|---|---:|---|
|
|
| `instruct` | text | required | Natural-language voice, accent, timbre, age, and delivery description. |
|
|
| `max_tokens` | positive integer | `2048` | Maximum generated audio-token count per text chunk. |
|
|
| `min_tokens` | non-negative integer | `28` | Minimum token count before EOS may stop generation. |
|
|
| `temperature` | positive float | `0.4` | Autoregressive sampling temperature. |
|
|
| `top_p` | `(0, 1]` | `0.9` | Nucleus sampling probability. |
|
|
| `repetition_penalty` | positive float | `1.1` | Autoregressive repetition penalty. |
|
|
| `seed` | non-negative integer | random | Sampling and codec-noise seed. |
|
|
| `text_chunk_size` | positive integer | `300` | Maximum Unicode codepoints per long-form chunk. |
|
|
| `text_chunk_mode` | `default`, `tag_aware`, `japanese`, `endline` | `tag_aware` | Framework long-form text chunking mode. |
|
|
### Session options (use with `--session-option`)
|
|
|
|
| Option | Values | Default | Description |
|
|
|---|---|---:|---|
|
|
| `weight_type` | framework-supported storage type | `native` | Shared AR and SNAC decoder storage type. |
|
|
| `ar_weight_type` | same as above | `weight_type` | AR backbone storage override. |
|
|
| `codec_weight_type` | `native`, `f32` | `weight_type` | SNAC decoder storage override. Set `maya1.codec_weight_type=f32` when the shared `weight_type` is neither `native` nor `f32`; AR storage remains independently configurable. |
|
|
|
|
## Convert
|
|
|
|
Convert the original Maya1 and SNAC checkpoints into a self-contained GGUF:
|
|
|
|
```bash
|
|
conda run --no-capture-output -n qwen3-tts \
|
|
python tests/maya1/convert_gguf.py \
|
|
--model-dir /path/to/maya1 \
|
|
--codec-dir /path/to/snac_24khz \
|
|
--output-dir /path/to/Maya1-GGUF
|
|
```
|
|
|
|
Maya1 is Apache-2.0 licensed. The bundled SNAC decoder weights are MIT
|
|
licensed; retain the notices for both upstream projects when redistributing the
|
|
combined GGUF.
|