← sources

Has anyone tested SST ( speech to text ) models like Saaras, Scribe, Akshara or Deepgram on actual Indian voice calls?

source: reddit (reddit:developersIndia)published 2026-09-25 09:09ingested 2026-09-26 02:36

retained text

I’m trying to look at different speech to text models for Indian voice agents under real call conditions, rather than clean audio benchmarks.

They all seem to be solving slightly different versions of the same problem. Sarvam Saaras V3 is designed around Indian languages and code mixed speech, Convozen Akshara is focused on compressed 8 kHz telephony audio, ElevenLabs Scribe takes a broader multilingual approach; and Deepgram Nova 3 offers mature real time streaming with support for Hindi and several other Indian languages.

Has anyone tested them on actual Indian calls?

I’d especially like to compare:

Transcription accuracy

Hinglish and regional code-switching

Streaming latency

Indian names, addresses and numbers

Noisy 8 kHz audio

Interruptions and overlapping speech

For Indian voice agents, does telephony specific training make a noticeable difference in production, or do broader multilingual models perform similarly after deployment?

( sources:- Sarvam saaras docs, convozen akshara model page, Elvenlabs scribe docs, Deepgram Nova 3 multilingual/code switching docs)

document fetched 2026-09-26 00:25 · canonical https://reddit.com/r/developersIndia/comments/1wprhbi/has_anyone_tested_sst_speech_to_text_models_like