The Culprit Behind the 'Slow Speech' Bug in Voice Conversion Was Whisper's 30-Second Limit

A slow-sounding voice conversion bug in long recordings was traced not to the model, but to Whisper’s 30-second input limit.

Audio longer than 30 seconds was truncated during feature extraction, and the shortened semantic features were stretched across the full output, causing unnatural timing.

The fix was to split long audio into overlapping 30-second chunks and merge the results afterward.

Anyone using Whisper as a feature extractor for voice conversion, TTS, lip-sync, or subtitles can run into the same truncation issue.