Whisper / speech-to-text in a demo: record at 16 kHz mono and trim silence
@devsingh05▲ 7 net votes
Speech models are typically trained on 16 kHz mono audio, and long leading silence or background noise can cause hallucinated text. Resample and downmix before upload (ffmpeg -ac 1 -ar 16000), trim silence, and send short segments for lower latency. In the browser, MediaRecorder produces webm/opus by default; check the transcription API accepts that format or convert server-side.