Hindi speech to text: what a free browser model can and cannot do
The tool is set to Hindi, but read this first: the small model that runs in your browser is not accurate for Hindi. Here is the evidence and the better route.
Public-domain LibriVox recording. Load it, then press Transcribe to try the full tool.
Whisper base, the model this tool runs, scored a 101.1% word error rate on Hindi in the Whisper paper's FLEURS test (lower is better). Most sentences will be wrong. Use a larger model or a human transcriber for anything you publish.
First use downloads the speech model and engine (about 97 MB, of which 72.5 MB is the model); later visits load it from your browser cache. Runs on your CPU; nothing is uploaded. Up to 100 MB and 2 hours per file; under 30 minutes is the comfortable range on a laptop.
A free in-browser model does not transcribe Hindi reliably. Whisper base, which runs here, scores 101.1% word errors on Hindi in the Whisper paper (FLEURS), meaning more errors than words, and in our own test it could not produce Devanagari at all. Larger Whisper models are far better: large-v2 measures 21.5%.
Checked by Koldflux · updated 2026-09-25 · tested on a public-domain Hindi recording
How accurate is Hindi transcription?
Word error rate from the Whisper paper (lower is better): FLEURS, read sentences recorded under the hi_in (India) locale, and Common Voice 9, volunteers with varied microphones. This tool runs Whisper base. Koldflux Studio runs Whisper large-v3-turbo, which the paper does not measure; OpenAI reports large-v3 cuts errors by 10 to 20% against large-v2, and the turbo version trades a little of that accuracy for speed.
| Model | FLEURS WER | Common Voice WER | Verdict (FLEURS) |
|---|---|---|---|
| Whisper base (this tool) | 101.1% | 106.5% | Not reliable |
| Whisper small | 38.4% | 43.6% | Rough draft only |
| Whisper medium | 26.8% | 31.5% | Usable, needs a proofread |
| Whisper large-v2 | 21.5% | 21.9% | Usable, needs a proofread |
What we saw on a real Hindi recording
We ran the first 30 seconds of a Hindi reading of the Panchatantra (LibriVox, public domain) through the same files this page loads, with the language set to Hindi. One 30-second clip is an illustration, not a benchmark. Processing took 66.8 seconds on one CPU thread of our test server.
First pass: ھاہہہہہہہہ… (one Urdu-script letter repeated until the length limit). After all four fallback passes the text was still unusable.- Detected as Hindi with 81.8% confidence (Marathi 6.1%, Punjabi 3.2%).
- With Hindi selected, the model never wrote Devanagari: it produced Urdu-script letters and got stuck repeating them, even after the tool’s fallback decoding.
- With Urdu selected, the same clip came out as a readable phonetic transcript in Urdu script (“… لبروعوکس کی پیشکش ہے … ساروجنیک دومین میں ہے …”, i.e. “LibriVox ki peshkash hai … sarvajanik domain mein hai”). The model hears Hindi; it just cannot spell it in Devanagari.
Why Hindi fails here: Devanagari, Urdu and model size
Spoken Hindi and Urdu are very close, and the small Whisper model writes Urdu script far more readily than Devanagari. In our test it transcribed the sounds reasonably in Urdu script and failed completely in Devanagari. Converting Urdu script to Devanagari is not something this tool does.
Model size is the fix. In the paper, Hindi goes from 101.1% word errors with base to 38.4% with small and 21.5% with large-v2 (FLEURS). The browser can only carry the base model at a sensible download size.
Hindi subtitles and translation
| Value | |
|---|---|
| Characters per line (Netflix) | 42 characters per line |
| Reading speed (Netflix) | up to 22 characters per second (adult programs) |
| Tool default line length | 42 characters |
| Speech to English, Whisper base (BLEU, higher is better) | 3.4 |
| Speech to English, Whisper large-v2 (BLEU) | 22 |
Tips for Hindi
- Mostly English with Hindi words (Hinglish)?If the talk is mainly English, choose English. Hindi phrases will not be written in Devanagari: with English selected, Hindi-heavy passages were translated or garbled in our test.
- If you read Urdu scriptSelecting Urdu gave a phonetic transcript of our Hindi clip in Urdu script. That can help as a listening aid, not as a Hindi transcript.
- Hindi subtitlesNetflix’s Hindi guide allows 42 characters per line and about 22 characters per second for adults, faster than its English limit.
Limits and when to use something else
- Do not publish Hindi subtitles made with this free model; the error rate is higher than the word count.
- Hindi-to-English translation with this model is also weak (BLEU 3.4; large-v2 reaches 22.0).
- Other Indian languages written in Devanagari or their own scripts, such as Marathi, Bengali, Telugu, Kannada, Malayalam and Gujarati, score 100% or worse with Whisper base in the paper.
- Speed depends on your device. A laptop handles a 30-minute file comfortably; phones may run out of memory on long files. Hard limits: 100 MB and 2 hours per file.
- No speaker labels: Whisper writes one stream of text. It also does not mark music, laughter or background sounds reliably.
Frequently asked questions
Is there a free Hindi speech to text that runs in the browser?
Not one that is accurate with a small download. Whisper base, the model small enough for a browser tab, scores 101.1% word errors on Hindi. Larger Whisper models are accurate enough to edit (large-v2: 21.5%) but are several gigabytes and run on servers.
Why did the transcript come out in Urdu letters?
Hindi and Urdu sound very similar, and the small model writes Urdu script more readily than Devanagari. Choosing Hindi does not force Devanagari reliably with this model.
Can it do Hinglish?
Pick English for speech that is mostly English; the Hindi words will not come out in Devanagari. Mostly-Hindi speech did not work either way in our test.
What about Hindi to English subtitles?
With this model the direct translation is weak (BLEU 3.4 in the paper). A larger model or a human translator is the realistic route.
Turn one idea into a week of posts
For Hindi, model size is the whole difference: the paper measures 101.1% word errors for the base model that fits in a browser and 21.5% for large-v2. Koldflux Studio (paid, from $19/month) transcribes uploads on our servers with Whisper large-v3-turbo, then turns the recording into Shorts and carousel posts you review, approve and export. It does not publish or schedule for you.
- 1 · Recording, idea or script
- 2 · Pick the angles
- 3 · Review and approve
- 4 · Export Shorts and carousels
Language pages
Each page has the tool preset to that language, measured accuracy, our own test on a real recording and subtitle rules.