Company profile · 2 min read
Maya Research: real-time voice models for global languages
Maya Research builds low-latency voice models, open weights and a streaming API for multilingual voice applications.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
What it does
Maya Research builds voice models for languages and contexts underserved by mainstream speech systems. Its public product surface shows real-time Hindi and English models, open-weight releases, a voice API and streamed audio from Mumbai and Iowa. The SPC profile calls Maya the voice interface for the rest of the world; the homepage shows the model and API surface.
The fit is a developer or product team building voice experiences for multilingual users who need low latency, natural prosody and control over where the model runs. Maya is more interesting for teams that need language-specific behavior than for a generic text-to-speech demo.
Why I’d look closer
The public page gives an unusual amount of technical detail. Maya reports 75ms time to first byte, 7.3ms variation between calls, support for more than ten languages, streamed 24 kHz PCM and regions in Mumbai and Iowa. It names Maya 2 Native, Maya 2 Global and Maya 2 Flash, and offers Apache 2.0 open weights for Maya 1 and Veena. These are company-reported technical claims; a buyer should benchmark their own language, device and network conditions.
The SPC listing provides founder context: Dheemanth Reddy BS and Bharath Kumar are building the voice platform, with Bharath described as CTO working on voice intelligence and emerging technology. The open-weight and API split is strategically useful: a team can start with hosted inference and later evaluate self-hosting or fine-tuning.
What I’d ask
Which languages and voices have enough training/evaluation coverage, how are consent and data retention handled, and what are the latency tails under load? I’d test pronunciation, code-switching, names and medical/financial terms, then compare hosted API behavior with the open model before promising a production experience.
My editorial take
Maya Research is a strong fit for multilingual voice products that care about latency and local context. The open models make the work more inspectable than a pure API. The proof should be language-specific quality and reliability, not one global voice score.
Quick facts
| Field | Sourced detail |
|---|---|
| Buyer fit | Developers building multilingual voice interfaces |
| Product | Hosted TTS API, real-time models and open-weight releases |
| Technical claims | 75ms first byte, 7.3ms call variation and 10+ languages; company-reported |
| Public pricing | Not exposed in the sources checked |
Sources checked
Checked 2026-09-19.
| Source | Used for |
|---|---|
| SPC company profile | Company definition and founders |
| Maya Research homepage | Current models, API, latency and open-weight surface |
| Maya Research team | Public team surface |
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗