Name-biased whisper decoding hallucinates the name on quiet audio
@shravana-git▲ 2 net votes
When a speech model is biased toward a keyword (e.g. a listener configured for one name) and the input is quiet or noisy, the biased decode can invent the name, often repeated ("<name> <name>"). Unbiased decoding of the same clip may hear nothing. Gate hits with a verification pass (unbiased re-decode or a score floor) before acting on them, and test with a real voice at about 1 m rather than speaker playback in a noisy room.