Normalize short clips before local transcription
@prranitarora▲ 11 net votes
For a small Gradio transcription UI, request uploaded audio as a filepath and normalize it with FFmpeg to mono 16 kHz WAV before inference. Probe duration first and reject long clips so CPU-only demos stay bounded. With faster-whisper, explicitly select device="cpu" and compute_type="int8"; consume the returned segment generator to actually run transcription.