Which speech to text API works best for Indian languages?

There is no single best speech to text API for Indian languages that the public research settles. The results depend on the language, the accent and the kind of speech. The usable answer is a shortlist of APIs that support your language, then a test on recordings that match your users. This page gives the evidence for each option and a way to run that test.

Google Cloud lists a Hindi model, and the evidence is in its table

Google Cloud Speech-to-Text lists Hindi (India), code hi-IN, for its chirp_2 model ¹. The table is a list of supported features and languages, so it tells you the language is available, not how accurate it is. For accuracy, see the Hindi error rates below.

Azure lists Bengali and Urdu locales

Microsoft's Azure Speech table of speech to text locales includes Bengali (India) and Urdu (India) ². Azure gives each language a locale code, such as ur-IN for Urdu (India), and you should use the code that matches your speakers. As with Google, a listed locale is not proof of accuracy.

AI4Bharat publishes an open Hindi model under the MIT license

AI4Bharat's Hindi IndicConformer model is listed on AIKosh, the Indian government's AI portal. It is described as a Conformer-Large model with 120 million parameters ³, and AIKosh lists its license as MIT ⁴. Because it is open, you can run it yourself, which suits teams that cannot send audio to a cloud provider. The listing gives no error rate, so you must measure it.

The same group built the IndicVoices dataset, which has 7,348 hours of speech in 22 languages ⁵, and the authors say their IndicASR model supports all 22 Eighth Schedule languages ⁶. Support for 22 languages is a claim about coverage, not accuracy.

Bhashini is a government platform with its own documentation

Bhashini is India's government language platform. In the 2025 Maha Kumbh material, its translation supported a chatbot in 11 languages, including Hindi ⁷. We could not read Bhashini's current developer documentation, so check its endpoints, pricing and terms directly before you plan around it.

How the models compare on Hindi is a test you need to run

Public Hindi numbers come from different tests, so they are not directly comparable. On the FLEURS read-speech benchmark, Whisper large-v2 had 21.5 percent word error rate in Hindi ⁸. In a 2025 table from the Voxtral paper, Gemini 2.5 Flash had 6.76 percent, GPT-4o mini Transcribe 12.36 percent, and Whisper large-v3 28.87 percent ^{9 10 11}. Those are read-speech results. On the crowd-recorded Common Voice Hindi test, Whisper large-v3 reached 46.75 percent ¹².

Accents change the picture. In the LAHAJA benchmark, existing open-source and commercial models performed poorly on the LAHAJA benchmark ¹³. The authors also found declines for speakers from North-East and South India, especially with named entities and specialized terms ¹⁴. If your users are in those regions or use technical vocabulary, test on that speech.

A test plan that gives you a usable answer

1. Shortlist two or three APIs with your exact locale code.
2. Record 30 to 60 minutes of real speech from your users, including names and terms.
3. Transcribe it with each option and measure word error rate against a careful human transcript. The glossary entry on word error rate explains the calculation.
4. Count errors on Hinglish separately. Mixed speech raises error rates: recent evaluations report a 30 to 50 percent relative increase in word error rate on code-switched speech compared with single-language speech ¹⁵.
5. Compare cost and data handling after accuracy, and choose the lowest error rate you can afford.

For the broader language picture, see the Hindi voice typing guide and the guide to Whisper languages and accuracy.