Maya Research's streaming speech synthesis model: supports 11 Indian languages and English, outputs 24 kHz mono audio, includes 20 built-in synthetic voices, and can also clone voices zero-shot using 3–10 second reference audio, with cross-lingual usage (e.g., speaking Tamil using a Hindi reference audio).
Upvotes 72 · Likes 72 · Downloads 0
Reasoning
- Officially stated to have 2.50B parameters (approx. 5.0 GB BF16 weights): the backbone is initialized from Qwen3-1.7B-Base, adds a depth decoder of approx. 555M, uses the Kyutai Mimi codec, outputs one audio frame every 80 ms, and supports playback while generating.
- Official evaluation uses 6,000 held-out texts (12 languages, 31 content types, totaling 20.3 hours of speech): the overall character error rate across three ASR evaluators ranges from 2.86% to 3.49% (scorer v1); URLs and code-switching between Indian languages and English remain difficult, with the ASR-1 character error rate for URL-type content at 29.06%.
- Officially stated to support 224 concurrent WebSocket streams on a single H100 NVL GPU, with a steady-state first-packet latency p50 of 111 ms; ranked 2 on the VoiceArena Hindi real-time TTS leaderboard as viewed on 10 27. The repository was created on 10 26, and it was not included in previous editions.
Usage
- Environment: Linux, Python 3.10–3.12, NVIDIA GPU supporting CUDA 12.8; pip install "maya3[server] @ git+https://github.com/MayaResearch/maya2-native" --extra-index-url https://download.pytorch.org/whl/cu128（包名和类名仍叫 maya3 / Maya3TTS).
- Python: tts = Maya3TTS.from_pretrained("maya-research/maya2-native", max_seats=1); tts.generate(text, lang="hi", voice="Ahana", normalize=True) returns a 24 kHz waveform; the Python interface requires manually passing normalize=True to convert numbers, currencies, and dates into spoken forms, and use tts.stream(...) for streaming.
- Cloning: tts.register_voice("my_voice", ref_audio=reference_audio, ref_text=verbatim_transcription, lang=reference_language), specify the target language in lang during generation; Server: python -m maya3.server --ckpt maya-research/maya2-native --port 8080 --max-seats 1, providing an OpenAI-compatible /v1/audio/speech endpoint and WebSocket streaming interface.
Risks
- Does not support Chinese; no built-in voices have more than 1 hours of training data in Assamese, which may result in a Bengali accent; voice stability is poor for short phrases of one or two words.
- Officially stated peak memory usage for a single stream is approx. 9.2 GB (plus CUDA context), recommending GPUs with 12 GB or more; the included server has no authentication, so you must add it yourself before exposing it publicly.
- No audio watermarking; officially requires cloning only your own or authorized voices; cloning similarity has not been specifically evaluated, all scores are based on ASR and automatic voiceprint scoring, with no human naturalness evaluation and no comparison with other TTS systems.