An honest view of Assamese accuracy
Assamese has very little training data in the speech model. Expect a rough draft that needs a full read, not a finished document. Results are best with one speaker, a quiet room and a phone held close. Fast conversation, Bihu songs or recordings with traffic noise will need many corrections.
Upload the recording as it is: MP3, M4A, WAV, OGG, OPUS, FLAC, a voice message or a phone recording. You can also record from your microphone in the upload box. Nothing needs converting first.
The transcript opens next to a player that highlights the current line. Click a line to hear it again, correct the word in place, and your changes save automatically.
Assamese script and Bengali
Assamese uses the Eastern Nagari script, shared with Bengali except for a few letters: ৰ (ro) and ৱ (wo) are Assamese, while Bengali uses র and ব. Because the two languages are close and Bengali has far more training data, auto-detect can label an Assamese file as Bengali and write Bengali spellings. Keep Assamese selected, as this page does, and watch for র appearing where ৰ belongs.
Assamese also has sounds that Bengali lacks, such as the velar fricative written with স, শ or ষ in many words. Spelling of these and of vowel signs can vary from the standard, so fix them while you listen.
Mixed speech: English, Hindi and Bengali
Everyday Assamese in Guwahati, on news channels and in offices mixes in English words, and many speakers also switch to Hindi or Bengali. English words may appear in Latin letters or written out in Assamese script. Whole sentences in Hindi or Bengali may be forced into Assamese spelling, so check those passages. If a recording is mostly English with some Assamese, choose English for that file instead.
Typical recordings and next steps
People bring research interviews from Assam, oral-history recordings of grandparents, recorded college lectures, local news clips, podcasts and voice messages from family. Some use the draft as a starting point before a native speaker cleans it up for publication. Long recordings are split into chunks and processed in parallel, so an hour-long interview does not take an hour to come back.
Export TXT, DOCX, PDF, CSV, SRT or VTT, with timestamps and speaker names switched on or off per export. Speaker labels are available on paid plans or with a Minute Pack.
For video files, such as YouTube news segments or recorded events, see the Assamese video to text page.