Which speech to text works best with an Indian accent?

The best-tested option for Indian English is the open source Whisper model, in its large version. In the 2023 Svarah test, the large model had a word error rate of 7.2 percent ¹, the lowest of the systems the authors measured. The two commercial services in the same test were outperformed by the open source Whisper and NVIDIA Conformer models ². That is a starting point, not a verdict on your voice. The test is from 2023, and results depend on your first language and on the app that runs the model.

Svarah is the most useful public test of Indian English we found

Standard English test sets represent Indian speakers poorly. The Svarah authors say Indian speakers find a very poor representation in existing English benchmarks such as LibriSpeech ³. Svarah was built to fill that gap, with read speech and conversational speech from speakers of many regional first languages. Because it measures Indian speakers directly, its numbers say more about Indian English than the scores on US test sets do.

A benchmark is still one test. It has a limited number of speakers, it was run in 2023, and the models were run the way the authors set them up. Read the figures as a ranking for that test, not a guarantee for every speaker.

Medium Whisper also scored in single digits, but the large model may not be the one you can run

The authors found that the medium and large versions of Whisper gave single-digit error rates on Indian English ⁴. They also noted that, given the size of these models, they may be hard to deploy in the Indian context. For a dictation app, that matters. The app you use may run a smaller model, a cloud model or a different system altogether, and a smaller model can score worse on the same speech.

So the useful question is not "which Whisper is best?" It is "which system does this app run, and how does it score on my passage?"

Your first language changes the result

Results vary by first language. The authors report that the models performed poorly for speakers whose native language is Assamese, Bodo or Nepali, which they describe as low-resource languages from northeastern India ⁵. If your first language is one of these, expect weaker results and test more carefully before you depend on a tool.

Accent is one variable, and the language you dictate in is another. Whisper's own multilingual test shows the gap: the large-v2 model had a word error rate of 4.2 percent in English ⁶ and 21.5 percent in Hindi ⁷ on the FLEURS benchmark. If you mix Hindi and English in one message, read the guide to multilingual dictation before you choose a setting.

Set the English variant, then test the tools you have

Set the regional variant before you test. Apple's iPhone dictation languages list English (India) as a separate choice ⁸, so the setting is there to use. Then run a fair comparison:

1. Open the dictation settings for each app you plan to use, and select English (India) if it is offered. If it is not, choose the English variant closest to your speech.
2. Choose a passage of about 200 words that you would normally dictate. Keep the text in front of you as the reference.
3. Dictate the same passage into each tool, at your normal pace, with the microphone you use every day.
4. Count the words each tool gets wrong, including substitutions, missing words and extra words. Divide by 200. The lowest word error rate wins for your speech.

Use the same passage and the same microphone for every tool, or the comparison means little. If one tool is clearly better, keep it. If two are close, pick the one that handles your names and technical terms better, since those are the errors you will correct most often.

What to take from this

For Indian English, the strongest public evidence points to the open source Whisper family, and the large model did best of the systems tested in 2023. Your first language, the app you use and the variant you select can change the result. A short passage in your own voice will tell you more than any benchmark.

For the causes of errors, see why dictation does not understand your accent. If the tool allows you to adapt to your voice, the question on training dictation to your accent explains what that can and cannot do. For a broader English plan, read English voice typing for non-native speakers. Our guide to speech recognition and accents covers the wider research, and the languages overview lists every guide in this series.

Yap, which we make, is an on-device dictation app for Mac and Windows, so it can take part in the same passage test. For prices and downloads, see pricing and the download page.