Bengali audio and video to text

Bangladesh, West Bengal and regional speech

Standard colloquial Bengali (cholito bhasha), as spoken in Dhaka and Kolkata, is the target. Bangladeshi and West Bengali speech differ in vocabulary and some sounds, for example the common word choices জল and পানি for water. Sylheti, Chittagonian, Noakhali and Rangpuri are far from the standard and are often treated as separate languages, so expect them to come out badly. Formal sadhu bhasha is a literary register and rarely spoken.

Script, conjuncts and punctuation

Bengali is written left to right in its own script, with spaces between words and the danda (।) as a full stop. Conjunct letters such as ক্ষ and vowel signs placed around a consonant need a font and renderer with Bengali support. If you see boxes or broken clusters in a player, that is a font problem, not the transcript. Some letters can be encoded in more than one way in Unicode, so a word that looks right can occasionally fail to match a search until you retype it. Check numbers and English words, which appear inconsistently.

What speakers of Bengali use it for

University and coaching-class students record lectures to revise from. Journalists and researchers in Dhaka, Kolkata and the diaspora transcribe interviews. Families and community groups keep voice notes and talks as text. Creators draft captions for vlogs and short videos. Others check literary or religious talks for exact quotes. Every result needs a careful proof-read before it is shared.

Which model works best for Bengali

Bengali is among the languages Whisper-family models have seen the least speech in, and the small model is often only a rough guide. Use short, clean, single-speaker clips and correct the draft against the audio. The larger models help when you self-host Scriptorium, but even they make many errors in Bengali.

On tryscriptorium.com the no-account preview and regular accounts run the small model (the larger medium and large-v3 models run when you self-host Scriptorium). In an account, pick Small in the model menu rather than Auto: Auto prefers the fastest, least accurate model for long files or a busy queue.

Bengali is a weak language on the small model, so the quality of the recording decides a lot: record near the microphone, avoid background music and overlapping voices, and split long recordings. For important material, self-host with large-v3 and still read through the text.

Subtitles and captions in Bengali

Bengali conjuncts make characters visually wide, so keep subtitle lines shorter than for Latin scripts, around 35 to 40 characters, with two lines at most. SRT and VTT exports are UTF-8 and display in VLC and YouTube, but a device without a Bengali font may show boxes. Proof-read every line before publishing. Subtitles stay Bengali.

The uploader is set to Bengali

The upload box on this page transcribes as Bengali, so the recogniser does not have to guess the language from the first seconds. If the recording is in another language or switches languages, use Audio to text, which detects the language automatically. Automatic transcription is a draft and no accuracy figure is promised: check names, numbers and anything you will rely on.