How do you transcribe Tamil, Urdu, Bengali or other South Asian audio?

Set the spoken language to Tamil, Urdu or Bengali yourself, transcribe the recording with a tool that supports that language, and then read the output against the audio. How well this works depends on the language far more than on the file format. In a published test on read sentences, OpenAI's Whisper large-v2 had a word error rate of 17.5 percent in Tamil ¹ and 22.6 percent in Urdu ². In Bengali it was 104.1 percent ³, which means the output had more wrong words than the reference contained. Tamil and Urdu are workable with checking. Bengali, on that model, is not.

This page explains those numbers, the cases where the answer changes, and the steps to take. For the same approach to Hindi, see the question on Hindi audio. For the wider method, start with the guide to transcribing audio in other languages.

Bengali and Malayalam are far harder than Tamil and Urdu

The 2022 Whisper test covered many languages, and the spread across South Asian languages is very wide. Whisper large-v2 had a word error rate of 100.7 percent in Malayalam ⁴, so Malayalam and Bengali were both at or above 100 percent. Values above 100 happen because the error count includes inserted words. A model that adds words that were never spoken can score worse than 100 percent even when every real word is accounted for. For the term, see word error rate.

Tamil and Urdu sit in a much better range on the same table. Treat those as the best case for read speech. A conversation, a voice note or a recording with background noise will usually be harder.

Fine-tuned Indian-language models won most tests in one 2023 study

Some Indian-language speech projects fine-tune Whisper on local data rather than using it as released. One such project, Vistaar, built IndicWhisper, a Whisper model fine-tuned on Indian-language data, and compared it with other systems. In its own benchmark, IndicWhisper had the lowest word error rate in 39 of 59 benchmarks ⁵.

That result is encouraging but limited. The benchmark is written by the same group that built the model, and a count of wins does not tell you how much better it is for your audio. If you can run an open model, look for one fine-tuned on the language you need, and test it on your own recordings before you rely on it.

Urdu results depend on whether the speech is read or spoken

Urdu shows a split that matters for most recordings. A 2024 Urdu benchmark found that Seamless-large was best on read speech, while Whisper-large was best on conversational speech ⁶. If you are transcribing dictated notes, choose the tool that wins on read speech. If you are transcribing a conversation, the Whisper result is the more relevant one.

Urdu is written right to left, in a script derived from Arabic, so the same direction check described in the question on Arabic and Hebrew audio applies. Read the first and last word of each line, and check numbers and English names separately, because the direction changes where mixed text ends up. For the term, see right-to-left script.

Choose the regional setting, because the setting changes what the tool expects

Many tools list South Asian languages by region rather than by language. Microsoft's Azure Speech table includes Bengali (India) and Urdu (India) as separate locales, under the codes bn-IN and ur-IN ⁷. A locale in that table is not an app feature, so confirm that the tool you use offers your locale.

Urdu is listed by region in some products too. Google's current Docs help page lists Urdu (India) and Urdu (Pakistan) as separate voice typing entries ⁸. Pick the entry that matches the speaker's region, since spelling and vocabulary can differ between them.

Test on your own audio before you rely on a tool

A published figure for read speech is a starting point, not a guarantee. The steps below work for any tool and any of these languages.

1. Set the language before you start. For Urdu, choose the regional entry that fits the speaker. If the tool offers no entry for the language, look for one that does before you start.
2. Transcribe one or two minutes and read the result. If the words or the script look wrong, change the setting and run it again.
3. Build a small test set. Choose three to five recordings that include the hardest ones: fast speech, background noise, and any English words mixed in.
4. Correct one recording by hand, word by word, to create a reference, and count the errors against it. That gives your own word error rate.
5. Decide what a wrong word would cost. If an error could change the meaning, plan a human review.

For a recording that mixes English into Tamil, Urdu or Bengali, expect more errors at each switch. Mixed-language speech is harder for speech models ⁹, and the term for it explains why. Read the switch points closely.

When to use a human transcriber

An AI draft saves most of the typing time, but some work needs a person who speaks the language. Use a human transcriber, or a fluent reviewer, for legal records, medical instructions, interviews that will be quoted in print, and any recording where a wrong word could change the meaning. A word error rate above 100 percent, as in the Bengali test, is unlikely to be a usable starting point for that work.

For a recording you can upload in a browser, our free audio to text tool works from the file itself. Check its output with the steps above before you use it, whatever the tool.

Other questions in this guide: Arabic or Hebrew audio, Spanish, German or Dutch audio, translating a voice recording to English and whether AI can transcribe African languages. For the wider method, see how to transcribe audio to text and how accurate AI transcription is compared with people.