Maya Research 发布支持12种语言的开源权重语音模型 Maya 2 Native
Maya Research Releases Open-Weight Speech Model Supporting 12 Languages
Maya Research 发布开源权重文本转语音模型 Maya 2 Native，支持包括11种印度语言和英语在内的12种语言，采用 Apache 2.0 许可。模型支持实时音频流式输出，以 80 毫秒帧生成 24 kHz 单声道音频，并内置20种合成音色，可用约3至10秒参考音频进行声音克隆。本地运行需 NVIDIA GPU，文档建议单会话至少 12 GB 显存；模型卡也指出混合语言文本、URL、数字及部分跨语言音色组合存在局限，且未包含音频水印。
正文
Maya Research has released Maya 2 Native, an open-weight text-to-speech model that generates speech in 12 languages, including 11 Indian languages and English. The model supports real-time audio streaming and voice cloning, allowing developers to build multilingual voice applications.
The release gives developers access to a model that can generate speech, imitate a reference voice, and produce audio in another supported language. Licensed under Apache 2.0, Maya 2 Native can be integrated into applications or deployed on suitable infrastructure, making it an option for developers building voice assistants and regional-language AI services.
Maya Research's model documentation describes the release as “Streaming text-to-speech for 11 Indic languages and English.” The model generates 24 kHz mono audio and streams it in 80-millisecond frames, letting applications start playing speech while generation continues.
Supported languages include Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu. The model also includes 20 built-in synthetic voices that developers can select for different applications.
Maya 2 Native supports voice cloning using a reference audio clip of around three to 10 seconds. Developers can use a sample recorded in one language to generate speech in another supported language. The model's documentation notes that cloned voices approximate the reference, with similarity and accent varying across outputs.
The release also includes inference code and interfaces for Python, an OpenAI-compatible speech endpoint and WebSocket streaming. Its Apache 2.0 licence lets developers use and modify the released model and code, subject to the licence terms. Running the model locally requires an NVIDIA GPU, with the documentation recommending at least 12 GB of GPU memory for a single-session setup.
Maya Research evaluated the model on 6,000 text samples across 12 languages, generating 20.3 hours of speech. Three automated speech-recognition systems assessed intelligibility, with character error rates varying by language and evaluation system. The documentation reports lower error rates for several languages, including Hindi, Marathi and Gujarati, while Assamese presented greater challenges.
The model card also identifies limitations with code-mixed text, URLs, numbers and some cross-lingual voice combinations. It does not report a human naturalness study or a controlled comparison with other text-to-speech models.
The company recommends obtaining permission before cloning a voice and disclosing to listeners that the speech is synthetic, particularly in calls and media. The release does not include an audio watermark, leaving developers to implement appropriate consent, disclosure and abuse-monitoring measures.