We played 240 public test recordings, 20 in each of Hindi, Urdu, Bengali, Tamil, Telugu, Gujarati, Marathi, Kannada, Malayalam, Punjabi, Nepali and Odia, to a speech service and to speech models that run on your own device, and counted every word and letter each got wrong. Here are the numbers, how we got them, what they do not tell you, and every clip to check for yourself.
Live preview, rendered on this page by the same engine that burns your video.
Lower is better. Speech service: Microsoft's MAI-Transcribe, run in the EU, through our server, finding the language by itself, measured 16 October 2026 (the same numbers as our first run on 9 October). On the device: AI4Bharat's IndicConformer, one open-source model made for each language, the exact files Burncaps downloads, measured 26 September 2026.
| Language | Speech service: words / letters wrong | On your device: words / letters wrong |
|---|---|---|
| Hindi | 9.7 / 4.4 | 10.6 / 4.4 |
| Punjabi | 13.3 / 5.6 | 18.6 / 5.4 |
| Gujarati | 14.6 / 5.3 | 24.8 / 8.3 |
| Bengali | 14.9 / 3.8 | 15.8 / 3.8 |
| Urdu | 15.0 / 5.0 | 21.2 / 8.2 |
| Kannada | 17.0 / 4.3 | 20.5 / 5.6 |
| Odia | 17.5 / 7.8 | 25.7 / 7.7 |
| Marathi | 18.5 / 4.4 | 23.3 / 5.2 |
| Malayalam | 20.5 / 4.4 | 40.0 / 9.9 |
| Telugu | 20.6 / 3.1 | 25.7 / 7.2 |
| Nepali | 20.6 / 4.7 | 32.3 / 9.2 |
| Tamil | 21.2 / 8.6 | 28.2 / 11.1 |
| Average of the twelve | 17.0 / 5.1 | 23.9 / 7.2 |
20 clips a language, 240 in all, 48.6 minutes, 4,709 reference words. Over all 4,709 words together the service got 16.3 in 100 wrong. The service named the language correctly on 240 of 240 clips, and answered in about half a second a clip.
On the device, Hindi got 10.6 words wrong in 100; on the service, 9.7. Bengali is the same story, 15.8 against 14.9. If you speak Hindi or Bengali, keeping everything on your own device costs you almost nothing in accuracy.
The service halved the words wrong in Malayalam, 40.0 to 20.5 in 100, and cut Nepali from 32.3 to 20.6 and Gujarati from 24.8 to 14.6. For these three, the service is the way to the closest first draft.
Tamil, Telugu and Malayalam words are long: one word often carries what English says in three, so one wrong letter makes a whole word wrong. On the service Telugu had 20.6 words wrong in 100 but only 3.1 letters. That is why the text usually reads better than the word figure suggests.
Captions need each word at the moment it is said. On the service the first word started a median of 0.05 seconds from where the voice started, and 170 of the 240 clips were within a tenth of a second. Two weak spots: in Punjabi 54 of 508 words were placed inside a silence, and in Odia the last word ended a median 0.33 seconds after the voice stopped. In those two languages, play the end of each line back before you export.
Google FLEURS, a public set of sentences read aloud by native speakers, test split, licensed CC-BY-4.0. From each language we took the first 20 clips that are complete, 5 to 25 seconds long, loud enough (peak above minus 30 dBFS) and each a different sentence. The rule picks the same 240 clips every time it runs.
Each heard text was compared with FLEURS's own sentence after both were lower-cased and stripped of brackets and punctuation. The number of edits needed to turn one into the other, counted in words and in letters, divided by the length of the sentence, gives the wrong in 100. A number written as a word, a compound joined or split, or a second accepted spelling all count as wrong, so these figures are harsher than a reader's impression.
Every clip's result, with the FLEURS file it came from, the reference sentence, the text the service heard and its errors: speech-service-fleurs-indic-2026-10-16.json (JSON, about 230 KB). The reference sentences are FLEURS's, under its CC-BY-4.0 licence.
FLEURS is clear read speech by one person at a time. Casual talk, music under the voice, two people speaking at once or a busy street will do worse in every language. Twenty clips a language is a small sample: treat a difference of one or two points between languages as a tie. FLEURS is public, so some speech models may have been trained on it. Use the table to compare languages and the two ways of listening, not as a promise for one recording.
Drop a talking video in one of these twelve languages and Burncaps asks you once which way you want it heard.
| Our speech service | The model on your device | |
|---|---|---|
| Accuracy in our test | The closer of the two in every language | As good in Hindi and Bengali, behind in the rest |
| What leaves your device | The sound only, in pieces of up to 30 seconds, never the picture; nothing is kept | Nothing |
| First time | Starts at once | A 138 MB model downloads once for the language |
| Your video | Never uploaded | Never uploaded |
The words open as a script under the video. Click any word to correct it, and one click fixes the same word everywhere it appears. Then export the video with animated word-by-word captions burned in, or download SRT, VTT, ASS or text files. Free to use: two videos every 30 days with no mark, or $49 once for unlimited videos. Subtitle files are free without limit for recordings up to ten minutes.
Hindi, in our test: the speech service got 9.7 words wrong in 100 and the on-device model 10.6. Punjabi (13.3), Gujarati (14.6), Bengali (14.9) and Urdu (15.0) follow on the service. Tamil, Telugu, Nepali and Malayalam had the most words wrong, about 20 in 100, but only 3 to 9 letters in 100, because one wrong letter in a long word counts as a whole wrong word.
In all twelve languages we measured, yes, but by very different amounts. In Hindi and Bengali the gap is under one word in 100. In Malayalam the on-device model got 40.0 words wrong in 100 against 20.5 on the service, and in Nepali 32.3 against 20.6. The device keeps everything private; the service hears better.
On Google FLEURS, a public set of read sentences (test split, CC-BY-4.0): 20 clips a language, 240 in all, picked by a fixed rule. Each heard text was compared with FLEURS's own sentence after lower-casing and removing punctuation, and the edits needed were counted as words and letters wrong. The service run was made on 9 October 2026 and repeated on 16 October 2026 with the same numbers. Every clip's result is downloadable.
Not exactly. FLEURS is clear read speech; casual talk, music under the voice, several people speaking or a noisy street will do worse, and FLEURS may also be in some models' training data. Use the numbers to compare languages and the two ways of listening, not as a promise for one recording.
No. The video never leaves your device. Burncaps asks once whether you want our speech service or the model on your device. With the service, only the sound goes, in pieces of up to 30 seconds, to our server and on to Microsoft's MAI-Transcribe in the EU, and nothing is kept. With the device model, nothing leaves at all.
What it is: a browser tool that burns animated word-by-word captions into a talking video, or cuts it to its best 30 or 60 seconds.
Price: free with every feature, for two videos every 30 days with no mark; Pro, $49 once, gives unlimited videos with no mark. Pro is not a subscription.
Privacy: the video never leaves your browser. The sound leaves only if you choose our speech service.
Files: MP4 or MOV up to ten minutes in; an MP4 with the captions burned in out. SRT, VTT, ASS and text files are free without limit for recordings up to ten minutes. A longer recording, up to 90 minutes on a computer, gives its script and those files: the first ten minutes free, all of it with Pro.
Languages and styles: captions in the language spoken, twelve Indian languages heard on your device, translation into 37 languages, 20 caption styles.
After it listens: correct any word (one click changes it everywhere), add or cut words, one to six words at a time, position, size, text and highlight colours, and names this browser remembers, all free and without listening again.
Made by: Jordan Labs Ltd, UK company no. 17358646.
Against other tools: sixteen caption tools compared, from the vendors' own pages.