Maya 2 explained: A free open weight voice AI model for 22 Indian languages
Follow Us
India has a population of more than 1.5 billion, 22 languages officially recognized, and a typing issue. Typing Tamil and Odia on a phone’s keyboard is a difficult task, making voice the gateway for the usage of AI in India. Maya Research aims at becoming that voice. The CEO of Maya Research, Dheemanth Reddy, claims on X that Maya 2 surpasses the ElevenLabs v3, Cartesia Sonic 3.5 and Bulbul v3 in Indian languages, and would be open-weighted and free for everyone in a while. It sounds pretty ambitious. Here is the work that the company has shown so far,
SurveyAlso read: Fortnite is coming to Linux: Epic Games rebuilds Easy Anti-Cheat and vibe codes a new launcher
Rankings
Maya 2 Native, the main offering, is ranked one for real-time Hindi on Voice Arena and two in total terms with an Elo of 1086. Maya 2 Global, optimized for all markets, is ranked third for real-time English and fourth in total terms with an Elo of 1024. Maya 2 Flash, optimized for speed, has an Elo of 1089 on the Artificial Analysis leaderboards.
These are impressive scores for a seed-stage company. But they are not a reflection of how Maya beats ElevenLabs v3 on Indian languages. The ElevenLabs, Cartesia, and Bulbul benchmark was taken from Reddy’s post. And the results do not show how the test was conducted. It will have to be independently repeated in Marathi, Malayalam and Punjabi to be taken seriously.
Fast, and built for live calls
It’s where the noise comes out loud and clear in Maya’s offerings. The company claims its response time is 75ms for Time To First Byte, with a margin of error of 7.3ms between invocations, calculated from Bangalore to the Mumbai region. The audio begins streaming even while the model is generating, such that a lengthy paragraph speaks as fast as a short line of text. Maya offers an API endpoint that returns a 24 kHz PCM stream, hosted in Mumbai and Iowa.
Also read: Nothing smart glasses leak hints at two models, one with AI and a camera: Report
So what does that matter? It’s customer support, BFSI, IVR and e-commerce applications, all of which require live conversation. Imagine making a call reminding someone to repay their loan in Marathi. Half a second lag means a dead connection, while the flat robotic voice sounds like a fraud. Maya claims that its models convey accent, rhythm and emotion – something anyone can check themselves, since the site features audio clips in 13 languages.
There are ten Indian languages plus English, Arabic and Indonesian in those 13. They fall far short of the total 22 scheduled languages, although 10+ is stated on the site itself. There is no Assamese, Urdu or Dogri among them.
Open weight, with a catch
Maya knows how to do it. While Maya 1 is available on Hugging Face under Apache 2.0 license, Veena, the first Hindi and English model, continues to be downloaded daily, according to the firm. But there is more to Maya 2 than that, for now. It is offered as hosted models by the site, but, according to Reddy, it will soon become available freely to all builders. “Free” is not just talk at the moment but rather a concrete offer.
The funds to back the project are minimal. Maya secured a $1.9M seed round from South Park Commons, and its competitors are described as labs that are orders of magnitude larger than the startup. If the weights come to pass, Indian developers will have a voice model they can fine-tune and run on their stack, which is what the entire purpose of open weight is about. Keep an eye on Maya’s Hugging Face page for the release. In the meantime, it is an API you can use.
Also read: Workday CEO Aneel Bhusri calls Oracle, SAP as AI ‘dinosaurs’
