aruntemme/maya2-native-mlx-bf16

maya-research/maya2-native converted for mlx-audio on Apple Silicon: BF16, the reference numerics. All credit for the model goes to Maya Research.

Streaming text-to-speech for 11 Indic languages and English, with 20 voices and zero-shot cross-lingual voice cloning.

pip install -U mlx-audio
mlx_audio.tts.generate --model aruntemme/maya2-native-mlx-bf16 --text "வணக்கம்! நாளைக்கு 2 மணிக்கு meeting இருக்கு." --voice Kayal

from mlx_audio.tts.utils import load_model

model = load_model("aruntemme/maya2-native-mlx-bf16")
for result in model.generate("नमस्ते, आपका ऑर्डर कल पहुँच जाएगा।", voice="Ritvik"):
    audio = result.audio  # 24 kHz

• Language: inferred from the text's script; lang_code picks between scripts' languages (hi / mr, bn / as).
• Cloning: ref_audio, ref_text and ref_lang clone a 3-10 s clip into any language. Get the speaker's permission first.
• Streaming: stream=True.
• Text normalisation: numbers, currency, dates and URLs are spoken correctly in every language (normalize=True, the default).

See the mlx-audio docs page for options, voices and fidelity checks against the reference implementation.

Converted with python -m mlx_audio.tts.models.maya2.convert.

License and attribution

The weights are Maya Research's, released under the Apache License 2.0; see LICENSE in this repo.

Changes from the original:

• converted from the release's safetensors to the mlx-audio layout (tensor names; tokenizer moved to tokenizer/)
• no numeric change (BF16, as released)

Upstream components keep their own licenses:

• Qwen3-1.7B-Base (Apache-2.0): the backbone the weights were fine-tuned from, included in these weights.
• Kyutai Mimi (CC-BY-4.0): the audio codec, downloaded from Kyutai at runtime and not redistributed here.
• SpeechBrain ECAPA-TDNN (Apache-2.0): the speaker encoder, downloaded at runtime and not redistributed here.

This conversion is not affiliated with or endorsed by Maya Research.

Responsible use

This is Maya Research's guidance for the original model, and it applies equally here:

• Clone only with permission. Use voices you own or have explicit permission to use.
• Disclose synthetic speech to listeners, especially in calls and media.
• No harmful use. Never use it for deceptive impersonation, fraud, harassment, non-consensual content, disinformation, or to defeat voice authentication.
• The built-in voices are synthetic personas, not real people.
• No watermark. The model adds no audio watermark, so applications should add their own consent, disclosure and abuse safeguards.