How Accurate Is Free AI Speech-to-Text in 10 Indian Languages? We Measured It

With the language picked, Whisper large-v3-turbo wrote Hindi and Urdu with about 10% of characters wrong: good enough for subtitles you check before posting. Tamil, Kannada and Marathi came out with 16–25% of characters wrong: readable, but every line needs fixing. Bengali, Telugu and Gujarati were worse, and Malayalam and Punjabi were unusable.

On auto-detect, the model heard Telugu, Kannada, Malayalam and Punjabi as other languages on all 15 clips each, and Bengali, Gujarati and Marathi on most of theirs. If you make videos in an Indian language, pick the language before you make subtitles.

The clips are 15 test sentences of each language from FLEURS, Google's public speech dataset (CC BY 4.0): 150 clips and 1,831 seconds of speech, read aloud by native speakers, each with its exact written text. We measured on 2 November 2026.

Each clip went through the same path a video's sound takes on ClipUtils' Auto Subtitles: read in the browser, made louder if it was quiet, and sent as Ogg Opus audio to two services running OpenAI's Whisper large-v3-turbo, the model behind most free subtitle tools: Groq and Cloudflare Workers AI. First with the language picked, as you'd pick it from the menu, then on auto-detect.

We scored each transcript against the real text with the two standard measures: word error rate (WER), the share of words that had to be changed, added or removed, and character error rate (CER), the same for single characters. Case and punctuation don't count, and the text is Unicode-normalized first, so the same letters typed two ways count as the same.

Both services run the same model, and on Hindi, Urdu and Marathi their results were nearly the same. Groq was clearly better on Tamil, Kannada and Malayalam, Cloudflare on Telugu and Gujarati; Bengali and Punjabi were close. (On Groq, one Kannada clip failed to come back, so Kannada there is 14 clips.) On Malayalam and Punjabi the word error rate passed 100%: the model repeated itself and made up words, which added more errors than there were words.

Most subtitle tools start on auto-detect. For Indian languages that's the biggest single cause of bad subtitles we found: Whisper heard Telugu, Kannada and Malayalam as Tamil, Punjabi as Hindi, and Bengali and Gujarati mostly as Hindi, then wrote them in the wrong language. Only Tamil was recognized every time.

Tamil, Telugu, Kannada and Malayalam join many ideas into one long word, and Bengali, Marathi and Gujarati add endings to words. One wrong letter makes the whole long word count as wrong, so word error rates run high even when most of the text is right. Character error rate is the fairer measure for these languages, and it's the one to read first.

For Hindi and Urdu, a character error rate of about 10% means roughly one letter in ten needs fixing, while a quarter to a third of words have at least one mistake.

Pick the spoken language before you make subtitles: on auto-detect, eight of these ten languages often came out in the wrong language altogether.

For Hindi and Urdu, free AI subtitles are a good first draft: read them through and fix names and the odd word. For Tamil, Kannada and Marathi, expect to edit most lines. For Bengali, Telugu and Gujarati, typing them yourself may be quicker than correcting; for Malayalam and Punjabi, these free models aren't ready.

Whisper's tiny model, which runs inside a browser and is offered by some tools (ClipUtils too) to keep video on your device, isn't usable for these languages: see the next section. Use the large model on a server for them.

We also ran Whisper tiny, the model small enough to work inside a browser (about 41 MB, so nothing has to be uploaded), with the language picked. It repeated phrases over and over and wrote more wrong text than there was speech, so its error rates passed 100%. We stopped after Hindi, Bengali and Tamil.

Fifteen clips per language is a small sample: a few hard sentences can move a language's numbers by several points, so read small differences between languages and services as noise. The clips are clear read speech recorded for a dataset; real videos, with music, noise, overlapping speakers, slang and Hinglish, will usually do worse. Our main results are for one model (Whisper large-v3-turbo) through two services, as free subtitle tools use it; paid services trained for Indian languages may do better. We also ran a third service, DeepInfra's Whisper large-v3, but left it out: its transcripts came back almost empty, which looks like a problem with how audio reaches it rather than the model, and we're checking it before publishing.

The audio and reference text are from FLEURS (Conneau et al., 2022), Google, licensed CC BY 4.0. Our scores per language, service and setting are free to use with a link to this page: https://www.cliputils.com/research/indian-stt-2026-11-02.json

To cite: ClipUtils (2026), "How Accurate Is Free AI Speech-to-Text in 10 Indian Languages?", 2 November 2026, https://www.cliputils.com/blog/indian-language-speech-to-text-accuracy