Bengali, Marathi, Kannada Audio to Text (Whisper)

Searchers type “audio to text Bengali”, “Kannada”, or “Marathi” and land on tools that claim fifty languages without saying which model they run. FlashKit’s private transcriber uses Xenova/whisper-base. The language list is copied from that library’s WHISPER_LANGUAGES table — the same codes OpenAI’s Whisper tokenizer publishes.

Languages we actually offer

Indic codes on the selector: Hindi (hi), Marathi (mr), Bengali (bn), Tamil (ta), Telugu (te), Kannada (kn), Gujarati (gu), Punjabi (pa), Urdu (ur), Malayalam (ml). Assamese, Nepali, Sanskrit, and Sindhi exist in the tokenizer too; they are not on the short list because we have not advertised them in the UI. We still will not invent a language the model cannot be asked for.

English remains the strongest. Hindi and other Indic languages give only a rough draft with this in-browser model: expect Latin-script output, repeated phrases and wrong words, so check every line.

Pick the language, then transcribe

The file stays in the tab. First run downloads about 77 MB of model weights.

Open the transcriber

How to get a less messy draft

Record close to the speaker. A phone on the far table will invent more tokens in every language.
One language per file. Mid-sentence code-switching is where whisper-base drops words.
Cut hold music and the “we are waiting for everyone” preamble with the audio cutter.
If the container will not decode, make a WAV with the audio converter.

The no-sign-up video path is transcribe video to text free. Timed exports are in SRT/VTT timestamps. General accuracy habits: better browser transcription accuracy.

What we will not claim

This is not a certified court reporter and not a replacement for a human transcript in a legal or medical matter. It is a local draft so the raw file does not have to hit a random caption SaaS. flashkit.tools is ours; treat the accuracy notes as our own limits, not a third-party lab report.

Frequently asked questions

Does the tool support Marathi and Bengali?

Yes. The selector includes mr, bn, and the other Indic codes in the Whisper tokenizer shipped with @xenova/transformers 2.17.2. We do not list a language that tokenizer rejects.

Is Hindi more accurate than Kannada or Malayalam?

Not in a way you can trust. Hindi, Kannada, Malayalam and the rest of the Indic list all come back as a rough draft with this in-browser model — Latin script, repeated phrases, and invented words. Check every line.

Should I use Auto detect?

Only if you truly do not know the language. A wrong code (or Auto on a mixed Hindi–English clip) inserts more invented words than picking Hindi or English on purpose.