Voice of India - Bengali: leaderboard

Speech recognition on the Bengali part of Voice of India (2026), a closed benchmark of unscripted telephone conversations in Indian languages, scored with an orthographically informed word error rate that accepts permissible spelling variants. Percent, can exceed 100; lower is better.

Metric: Orthographically informed word error rate (OIWER, %; can exceed 100) on the Bengali unscripted telephone-conversation utterances of the closed Voice of India benchmark, transcripts with lattices of permissible spelling variants; systems run at default inference settings and only where they support the language; lower is better. Source: arxiv.org. Saturation forecast: Around March 2028. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Sarvam Audio6.1
2Saarika 2.58.2
3Gemini 3 Pro8.5#145
4Amazon Transcribe9.1
5elevenlabs-scribe_v210
6IndicConformer10.7
7Gemini 3 Flash12.6#175
8gpt-4o-mini-transcribe21.1
9omniASR_LLM_7B22.8
10Microsoft Speech-to-Text25.4
11Deepgram Nova 328.9
12omniASR_LLM_1B29.7
13GPT-4o Transcribe44.9
14AssemblyAI Universal103.8

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=voice-of-india-bengali · How It Works · Data refreshed daily, snapshot 2026-11-21.