Bani STT is the listening half of the voice stack. Training mix weighted conversational and service-desk audio alongside read speech, with explicit coverage of Upper Assam, Lower Assam, Barak Valley, and Guwahati urban registers.

Word error rate alone is a blunt instrument; the chart shows how error moves when the same model meets speakers outside the broadcast norm.

Word error rate by region WER % · Navdyut Assamese speech eval · August 2025
Guwahati urban 7.4%
Upper Assam 9.1%
Lower Assam 10.6%
Barak Valley 12.2%

Deep model documentation stays on the Models page. This post is the release record.