Hindi speech-to-text APIs compared: prices and features, October 2026

What does each Hindi speech-to-text API publish?

Every price below is the vendor’s list price, read on 16 October 2026 from the linked page, in the vendor’s currency. Volume and enterprise contracts are not shown. A dash means the pages read did not say.

Provider and model	Published price	Hindi, as published	Live streaming	Source
mirai minds, Mira Transcribe	₹18 per hour of recorded audio	Built for Hindi and Hinglish phone calls; detects the main language	Standalone streaming in private preview; live inside every agent	Docs
Sarvam, Saaras	₹0.49 a minute real-time or batch, pay as you go; ₹0.41 on Growth and Business; batch with diarization ₹0.59	23 languages, 22 Indian plus English; a codemix output mode	WebSocket streaming	Pricing, models
ElevenLabs, Scribe v2; Scribe v2 Realtime	$0.22; $0.39 per hour	Hindi among 90+ languages	Scribe v2 Realtime	Pricing, languages
Deepgram, Nova-3	Pre-recorded $0.0043 a minute (multilingual $0.0052); streaming $0.0048 (limited-time; regular $0.0077)	Hindi (hi); the multilingual mode covers Hindi with English	WebSocket streaming	Pricing, languages
Google Cloud, Speech-to-Text V2	Standard $0.016 a minute for the first 500,000 minutes a month; dynamic batch $0.003	hi-IN on chirp_3 (us, eu regions), chirp_2 and telephony models	—	Pricing, languages
Microsoft Azure, Speech	Real-time $1 per hour; fast $0.36; batch $0.18	hi-IN listed, including fast transcription	Real-time transcription	Pricing, languages
Amazon Transcribe	US East examples: streaming $0.01 a minute, batch $0.006; tiered and varies by region	Hindi (hi-IN), batch and streaming	Streaming	Pricing, languages
OpenAI, gpt-transcribe; gpt-4o-transcribe; gpt-4o-mini-transcribe	$0.0045; $0.006; $0.003 a minute (estimated)	Language hints by ISO code; whisper-1 follows Whisper’s 98 languages	gpt-live-transcribe, $0.017 a minute	Pricing, guide

Speaker labels are priced differently from one vendor to the next. Ours are included in the ₹18. Sarvam lists batch with diarization as its own line, Deepgram and Azure list diarization as an add-on, and OpenAI points to a separate diarization model, gpt-4o-transcribe-diarize.

What does an hour of Hindi audio cost?

At list price, before tax and free tiers, for one hour of recorded audio:

Provider and model	Working	Cost
Mira Transcribe	1 × ₹18	₹18
Sarvam Saaras, batch, pay as you go	60 × ₹0.49	₹29.40
ElevenLabs Scribe v2	1 × $0.22	$0.22
Deepgram Nova-3 Monolingual, pre-recorded	60 × $0.0043	$0.258
Google Speech-to-Text V2, dynamic batch	60 × $0.003	$0.18
Azure batch transcription	1 × $0.18	$0.18
Amazon Transcribe batch, US East example	60 × $0.006	$0.36
OpenAI gpt-transcribe	60 × $0.0045	$0.27

Convert the dollar figures at the rate you pay. Free allowances change the small-volume picture: Azure gives 5 hours of real-time transcription a month on its free tier, Amazon gives 60 minutes a month for a year, and Deepgram gives $200 of credit.

How accurate is Hindi speech recognition?

Vendors publish their own accuracy figures, measured their own way, and we do not repeat them here. The one comparison we can give is our own. On all 10,000 utterances of the public RinggAI Hindi benchmark, six corpora in all, our report scored normalised word error rate for five systems with the same references and the same normaliser:

System	Normalised WER, all 10,000 utterances
Mira STT	9.12%
Ringg Parrot	9.72%
ElevenLabs	11.11%
Sarvam	12.83%
Deepgram	15.79%

Read the limits with the numbers. The vendor scores come from outputs Ringg bundled with the dataset, so the model versions are Ringg’s choice and may not be what each vendor serves today. Mira STT is a fine-tune of the open SraVaani-1.0 model from ARTPARK at IISc, which was trained on the training splits of these corpora, so the test is in-domain for us. Mira STT was not lowest everywhere: Ringg scored lower on CommonVoice and IndicTTS, and ElevenLabs on FLEURS. The report also scores the CTC-head build that ships in our package at 9.85%. On Lahaja, a corpus outside SraVaani-1.0’s training list, Mira STT scored 11.63% against 16.55% for the open Qwen3-ASR-1.7B, the only other system run there.

All of this is read and prompted speech, not phone calls, so it says nothing about accuracy on an 8 kHz call. Mira Transcribe, our product for recordings, adds several recognition passes and a language model on top, so the table describes the recogniser and not the whole pipeline. The full method is on the Mira STT page.

How should you test Hindi speech to text on your own audio?

1. Pick twenty to fifty recordings that look like production: the same phone lines, accents, background noise and call lengths.
2. Have a Hindi speaker write a reference transcript for each, and agree first how English words are written. Devanagari “ऑर्डर” and Latin “order” are the same word to a listener and two different words to a WER script.
3. Normalise spelling variants that sound the same, such as अन्दर and अंदर, before scoring. Do not normalise grammar: है and हैं should still count as different words.
4. Send the same files to each service with the same language setting. Give every service the names and terms it can take as vocabulary or keyterms.
5. Score word error rate, then read the worst ten transcripts yourself. Names, numbers and amounts usually decide whether a transcript is usable.
6. If you need speaker labels, check them separately, since a perfect transcript with the agent and customer swapped is still wrong for call review.

On our side, upload each file to POST /v2/transcriptions with a line of context and your vocabulary, then read the job until it completes. The transcription guide has the requests.

If you need transcription because you are building a calling agent, the per-minute plans are the better comparison: our agents are ₹1 a minute on T1 with a ₹1,00,000 monthly commitment, ₹3 on T3 with ₹50,000 a month, and ₹5 pay as you go, with recognition, the model and the voice included.