maya2-native-mlx-bf16

maya-research/maya2-native converted for mlx-audio, so it runs on Apple Silicon Macs. BF16, the weights as released; 4.7 GB. All credit for the model goes to Maya Research.

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

Use

Maya-2 support is on mlx-audio's main branch and will be in the next release after 0.6.2. Until then:

pip install -U "mlx-audio @ git+https://github.com/Blaizzy/mlx-audio"
mlx_audio.tts.generate --model aruntemme/maya2-native-mlx-bf16 --text "வணக்கம்! நாளைக்கு மீட்டிங் இருக்கு." --voice Kayal --lang_code ta

from mlx_audio.tts.utils import load

model = load("aruntemme/maya2-native-mlx-bf16")
result = next(model.generate("नमस्ते, आपका ऑर्डर कल पहुँच जाएगा।", voice="Ritvik", lang_code="hi"))
audio = result.audio  # 24 kHz mono

• Voices: model.voices lists all 20; the default is Ahana.
• Streaming: stream=True.
• Cloning: pass ref_audio, ref_text and ref_lang (a clean 3-10 s clip). The output language can differ from the reference. Get the speaker's permission first.
• Text normalization (numbers, currency, dates, phone numbers, URLs) is on by default; normalize=False turns it off.

See the mlx-audio Maya-2 docs for all options. Mimi codec and ECAPA speaker-encoder weights download separately on first use.

Converted with mlx-audio's standard converter:

python -m mlx_audio.convert --hf-path maya-research/maya2-native --mlx-path maya2-native-mlx-bf16 --model-domain tts

Licenses 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 mlx-audio's layout
• no numeric change (BF16, as released)

Upstream components keep their own licenses: Qwen3-1.7B-Base (Apache-2.0, included in these weights); Kyutai Mimi (CC-BY-4.0) and SpeechBrain ECAPA-TDNN (Apache-2.0), both downloaded at runtime and not redistributed here.

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

Responsible use

Maya Research's guidance for the original model 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 safeguards.