Model selection brief · 4 August 2026

Which speech-to-text model actually handles English and Indian languages?

A read of the blind Voice-of-India benchmark, an independent field-audio study, and the current vendor documentation — with the numbers that matter and the ones that mislead.

Short answer

If you can call a hosted API: Sarvam’s Saaras V3 (Sarvam’s audio model tops the leaderboard at 13.0% word error rate). It covers all 22 scheduled Indian languages plus English in one model, is the most accurate system on the only large blind Indic benchmark, handles Hinglish-style code-mixing, streams at sub-150 ms, and costs ₹30 (~US$0.34) per audio-hour.

If you need one model for the whole world, not just India: Google Gemini 3 Pro (15.1% — third overall, and the only global frontier model that stays close to the Indian specialists).

If you must self-host: AI4Bharat IndicConformer 600M (MIT licence, 16.0%) for the 22 Indian languages, paired with a separate English model — it has no English support of its own.

What to avoid for Indian languages: OpenAI’s transcription models and Whisper, AssemblyAI Universal, and Deepgram Nova 3. All are fine-to-excellent in English and fall apart on Indic scripts — sometimes past 100% error rate.

The benchmark that counts

Most “best speech-to-text 2026” lists rank models on clean English read speech, where the top ten are separated by less than one WER point and none of it predicts Indian performance. The Hugging Face Open ASR Leaderboard, the usual reference, contains zero Indian languages — its multilingual tab is German, French, Italian, Spanish and Portuguese.

The relevant yardstick is Voice of India (AI4Bharat at IIT Madras with Josh Talks, Interspeech 2026): 306,244 utterances, 536 hours, 36,691 speakers across 1,033 districts and 15 languages — unscripted two-party telephone speech, human-transcribed through a six-stage relay, and kept as a private, blind split so no vendor can train on it. It scores with OI-WER (orthographically-informed WER), which accepts legitimate spelling variants instead of punishing them.

Source: Voice of India leaderboard, arena.ai4bharat.org, read 4 August 2026. Bhojpuri and Chhattisgarhi are scored under the Hindi tag as a fallback.

What the ranking tells you that a single number cannot

India-native models lead, but not by much. The gap between the best API (13.0) and the best free, self-hostable model (16.0) is three points. The gap between them and OpenAI’s transcription endpoints is fifty points.
English-first models don’t merely degrade — they break. GPT-4o mini Transcribe returns 295.9% WER on Gujarati and 167.8% on Malayalam; AssemblyAI Universal exceeds 100% on six of fifteen languages; Whisper large-v3-turbo posts 124.8% on Maithili. Above 100% means the model inserts more wrong words than the reference contains — usually transliteration into the wrong script or outright hallucination. One Hindi clip came back as Hungarian text.
An MIT-licensed 600M model beats most of the industry. IndicConformer outranks ElevenLabs, Amazon, Azure, Deepgram, OpenAI and AssemblyAI on Indian languages while running on a single modest GPU.

Per-language reality check

“Supports Hindi” and “supports Malayalam” are very different promises. Hindi is close to solved; Dravidian and low-resource languages are not.

Word error rate (%) on Voice of India. Left: Hindi, the easiest case. Right: Malayalam, a hard case.
Model	Hindi	Malayalam	Comment
Sarvam (Saaras V3 family)	4.95	18.93	Best or tied-best in 14 of 15 languages
Sarvam Saarika 2.5	6.16	18.85	Previous generation; still best on Malayalam
Google Gemini 3 Pro	5.98	21.73	Best non-Indian system; wins on Bhojpuri/Chhattisgarhi dialects
Amazon Transcribe	6.78	—	Quietly competitive on Hindi
ElevenLabs Scribe v2	7.73	22.96	Broadest Indic coverage among global vendors (14 languages)
AI4Bharat IndicConformer	8.20	25.95	Open weights, MIT
Microsoft / Azure	11.42	—	Enterprise fallback
Deepgram Nova 3	12.97	—	Only 8 Indian languages; no Malayalam at all
Meta omniASR-LLM 7B	13.71	52.02	1,600+ languages — breadth, not accuracy
AssemblyAI Universal	19.28	107.50	Collapses on Indic scripts
OpenAI GPT-4o Transcribe	33.95	96.96 (Gujarati 98.2)	Not viable for Indian languages
OpenAI GPT-4o mini Transcribe	19.63	167.77	Gujarati 295.9%

Sarvam’s per-language spread on the same benchmark shows how much harder the tail is: Hindi 4.95, Bengali 5.96, Urdu 6.95, Marathi 9.42, Punjabi 11.24, Assamese 12.67, Gujarati 12.78, Odia 13.96, Tamil 14.16, Kannada 16.28, Telugu 18.17, Malayalam 18.93, Bhojpuri 20.88, Maithili 24.84. Budget for roughly 3–5× the Hindi error rate if your users speak Maithili, Bhojpuri or a Dravidian language.

Choosing: four situations, four answers

Default choice

Sarvam Saaras V3 (API)

22 Indian languages + English in one model; 1M+ hours of training audio; ~19.3% WER on the public IndicVoices set and 13.0% on the blind Voice-of-India set.

Streaming built in (Accurate / Balanced / Fast, sub-150 ms first token), automatic language ID with native-script output, speaker diarization in batch and realtime, numeral/format control.

₹30 per audio-hour (~$0.34); ₹45 with diarization. Indian company, so data-localisation and VPC/on-prem conversations are straightforward.

Weak spot: it is an Indic specialist — if you also need Thai, Swahili or Polish, look elsewhere.

Global coverage

Google Gemini 3 Pro / Chirp 3

Gemini 3 Pro is the strongest non-Indian system on Indic (15.1% overall, Hindi 5.98%) and best on Hindi dialects. It also preserves domain-critical terminology better than rivals — in the agricultural study its domain-weighted error was lower than its plain WER in all three languages tested.

Chirp 3 covers 111 locales but only Hindi and Indian English are GA; the other eleven Indian languages are Preview (no SLA, diarization limited to 14 locales). No India region. ~$0.96/audio-hour standard, $0.18 dynamic batch.

Self-hosted

IndicConformer 600M + an English model

MIT licence, 600M parameters, hybrid CTC/RNNT, 22 Indian languages, 16.0% on Voice of India. Runs on ONNX Runtime, CoreML and sherpa-onnx; int8 builds exist.

Three hard limits: no English (there is no en RNNT head), no language ID (you pass the code), no streaming. So you need a language-ID front end and a second model for English — NVIDIA nemotron-3.5-asr-streaming-0.6b (English + Hindi, 8.1% FLEURS Hindi, streaming) or Parakeet-TDT-0.6B-v3 / Voxtral Mini Realtime for English.

One global vendor

ElevenLabs Scribe v2

The pragmatic compromise if you want a single non-Indian vendor: 90+ languages, 14 Indian languages (the broadest of the global three), and the only vendor publishing per-language accuracy bands — English/Kannada/Malayalam ≤5%; Hindi, Bengali, Gujarati, Marathi, Nepali, Odia, Tamil, Telugu 5–10%; Assamese and Punjabi 10–20%; Urdu and Sindhi 25–50%.

$0.22/audio-hour batch, $0.39 realtime, ~150 ms partials, diarization to 32 speakers included.

Route by data-residency first, then by how much of the world beyond India you need.

Cost and latency, side by side

Model / API	Indian languages	VOI WER	Batch $/audio-hr	Realtime $/hr	Streaming latency
Sarvam Saaras V3	22 + English	13.0	≈0.34 (₹30)	≈0.34 (₹30)	<150 ms (Fast mode)
Google Gemini 3 Pro	wide (audio input to the LLM)	15.1	token-billed	token-billed	n/a (LLM call)
Google Chirp 3	13, only 2 GA	—	0.96 / 0.18 batch	0.96	not published
ElevenLabs Scribe v2	14	17.6	0.22	0.39	~150 ms
Amazon Transcribe	most major ones	17.6	1.44	1.44	~250 ms
Deepgram Nova 3	8	39.7	0.26	0.29	Flux: ~260 ms end-of-turn
IndicConformer 600M	22, no English	16.0	GPU cost only	—	no streaming mode
OpenAI gpt-4o-transcribe	nominal only	63.1	0.36	0.36	sub-150 ms

₹ converted at ≈₹88/US$. Prices are pay-as-you-go list rates read from vendor pricing pages on 4 Aug 2026; volume tiers cut most of them 20–45%. The overall market spread for STT is roughly 30× ($0.13/hr to $4.50/hr), so accuracy — not price — should drive this decision.

Five traps in the published numbers

Read speech flatters everyone. FLEURS Hindi scores of 6–10% become 13–21% on the Vaani benchmark’s code-switched, real-world Hindi. Same models, same language, double the errors.
Code-switching is the real workload in India, and it is priced in errors. Independent provider analysis puts the penalty bluntly: a model at 5% WER on monolingual audio routinely lands at 15–20% on Hindi-English mixed calls. If your users say “form submit kar diya, but OTP nahi aaya”, benchmark exactly that.
Character error rate is not word error rate. Meta’s omniASR shows a lovely 3.6% CER on Hindi and a 13.7% WER on the same language’s real speech, with 52% on Malayalam. Vendors quoting CER on Indic scripts are quoting the easier metric.
Diarization quality varies far more than transcription quality. On Sarvam’s open indic-diarbench (1,164 samples, all 22 languages), diarization error rate runs 16.0 (Sarvam) · 23.5 (AWS) · 32.0 (Deepgram) · 34.8 (Azure) · 35.0 (ElevenLabs) · 40.5 (AssemblyAI) · 74.0 (Gemini 3 Pro). Gemini transcribes words well and attributes them badly.
Multi-speaker field audio is a different problem. In the agricultural study, running diarization and keeping only the primary speaker cut Hindi WER from 88.3% to 30.0% for Chirp 3 and 53.5% to 18.5% for Gemini 2.5 Pro. If your recordings have bystanders, that pipeline step is worth more than swapping models.

A 90-minute evaluation you should actually run

Collect 60–100 real recordings from your own channel — the phone codec, the noise, the accents, the code-mixing. Public benchmarks cannot substitute for this; Appen’s private Indian-accent test sets reorder the public English leaderboard outright when switched on.
Transcribe them with three candidates: Sarvam Saaras V3, Gemini 3 Pro, and either ElevenLabs Scribe v2 or IndicConformer depending on your constraint.
Score three things, not one: WER; entity accuracy on the terms that carry your business (names, amounts, order IDs, drug or crop names — general English entity accuracy of 95% drops to 50–70% on alphanumerics); and P95 latency, not median.
Test the failure mode deliberately: feed silence, English-only audio, and a mid-sentence language switch. Watch for hallucinated text and wrong-script output.
If you are on noisy telephony, measure a denoise or voice-isolation front end separately — it typically removes 10–30% of errors and is cheaper than a model upgrade.

Bottom line

For English plus Indian languages in August 2026 the honest ranking is: Sarvam Saaras V3 if you can use an API, Gemini 3 Pro if you need global coverage in the same model, IndicConformer + an English model if the audio cannot leave your servers, and ElevenLabs Scribe v2 if you want one global vendor at the lowest per-hour price. Whisper and the OpenAI transcription endpoints, whatever their reputation in English, are the wrong tool for Indian languages — and the further you go from Hindi, the more expensive that mistake becomes.