Maya Research opens an Indic speech model
On 25 October 2026, the day its model card gives as the checkpoint release, Maya Research published the weights of Maya 2 Native, a streaming text-to-speech model for 11 Indian languages and English with 20 built-in voices and zero-shot voice cloning from 3 to 10 seconds of reference audio. The weights and the inference code, with a Python API, an OpenAI-compatible speech endpoint and WebSocket streaming, are under Apache 2.0.
Why it matters
Speech synthesis in Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu is released with an open licence by a small team, so it can be hosted and adapted by those who need it, and a Hindi reference voice can speak Tamil. The card is frank about its limits: Assamese is the weakest language, there is no human listening study, and no zero-shot cloning scores are reported.
What the card says. 2.50 billion unique parameters, initialised from Qwen3-1.7B-Base, with audio tokens decoded by Kyutai's Mimi codec (CC-BY-4.0) and a SpeechBrain ECAPA-TDNN speaker encoder; 24 kHz mono output in 80 ms streaming frames; it needs Linux and an NVIDIA GPU with CUDA 12.8. Licence: the model weights are Apache-2.0, and so is the inference code; the upstream components keep their own licences. The package and class are still called maya3 and the commit that uploaded the weights says "Maya3 1.0"; the card explains that the model is Maya 2 Native. Dates. The card's table gives "Checkpoint release | 1.0 · October 25, 2026". The Hugging Face repository was created on 26 October at 02:58 UTC and the commit with the weights is at 06:20 UTC the same day (23:20 on 25 October in the US Pacific zone), so the record keeps the card's day and the repository's is one day later in UTC. What the record does not claim. The card ranks the model second for Hindi on the VoiceArena leaderboard (checked on 27 October) and prints a snapshot with different figures; the board moves and no rank is repeated here. The card says its evaluation used automated speech-recognition and speaker-embedding judges and reports no human naturalness study.