Sarvam AI Releases Indic DiarBench for Multi-speaker Speech Recognition

itsOmSarraf_ · x · 2026-08-11

Sarvam AI has introduced Indic DiarBench, an open-source dataset designed to evaluate how well speech systems understand multi-speaker conversations across all 22 scheduled Indian languages.

The dataset tackles real-world conversational dynamics such as interruptions, overlapping speech, and rapid turn-taking. It features roughly 108 hours of audio with 2 to 9 speakers per session. Both transcripts and speaker turns are annotated together, allowing for the joint testing of diarization and ASR on the same audio. The accompanying paper has been accepted at Interspeech 2026.

Original post →

More from Research

Research channel →