# Indian voice AI is getting seriously competitive. Has anyone actually tested these models in production? — retained record

> Retained reddit record from reddit:developersIndia.

Canonical: https://highsignal.app/data/records/27995adb26036e5b

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:developersIndia)  published 2026-10-23 10:19 ingested 2026-10-24 02:36

[original source ↗](https://www.reddit.com/r/developersIndia/comments/1xda6ag/indian_voice_ai_is_getting_seriously_competitive/)

## retained text



```
Been looking into the Indian speech AI ecosystem recently, and it's interesting how many homegrown companies are developing their own models rather than relying entirely on OpenAI, Google or ElevenLabs.

A few interesting ones:

* **Sarvam AI:** Saaras v4 for STT and Bulbul v4 Flash for TTS. Focused on Indian languages, regional accents and code switching.
* **ConvoZen:** Akshara/Akshara Pro for STT and the Ragini family for TTS. Focused on Indian telephony, multilingual customer conversations and enterprise voice agents.
* **Gnani.ai:** Prisma for STT, Timbre for TTS and Warp for speech-to-speech. Interesting work around telephony, regional accents and real time voice applications.
* **Smallest AI:** Lightning v3.1 and v3.1 Pro for TTS, with support for multiple Indian languages and Hindi English code switching.
* **AI4Bharat (IIT Madras):** IndicConformer for STT and Indic TTS for speech generation. Open research models covering several Indian languages.
* **Project Vaani (IISc/ARTPARK):** SraVaani for STT and DhVaani for TTS. Particularly interesting for underrepresented Indian languages and dialects.

What's interesting is how differently these companies and projects are approaching the same problem. Some are building foundational speech models, others are developing complete voice agent infrastructure, while research initiatives are focusing on languages that commercial providers have historically overlooked.

But how much of this progress actually translates into production performance?

A model might perform extremely well on clean Hindi audio but struggle when someone switches between Hindi, English and a regional language mid-sentence, mentions a local place name, corrects themselves or interrupts the agent, all over a noisy 8 kHz phone connection.

And STT accuracy is only part of the equation. TTS naturalness, turn-taking, interruption handling and end to end latency can completely change how usable a voice agent feels.

A few things I'm curious about:

* Which Indian STT models actually handle Hinglish, regional accents and code switching reliably?
* How do Indian TTS models compare with ElevenLabs, Cartesia or Google in terms of naturalness, pronunciation and latency?
* Has anyone benchmarked these models using actual telephony recordings rather than clean datasets?
* How do they compare on API pricing, reliability and deployment at scale?
* Are open source models like IndicConformer or SraVaani viable alternatives to commercial APIs?

I'm particularly interested in whether Indian models are beginning to outperform international providers on Indian language use cases, rather than just catching up with them.

Has anyone here deployed or benchmarked any of these? Would be interesting to hear actual experiences, including where they fail, rather than just looking at demo videos and vendor benchmarks.
```



document fetched 2026-10-24 00:25 · canonical https://reddit.com/r/developersIndia/comments/1xda6ag/indian_voice_ai_is_getting_seriously_competitive
