The '3-Character' Catchphrase That the Quality Gate Allowed Became the Model's Habit
๐ Originally published (in Japanese) at forge.workstyle.tech . A trained speech model was asked to read a short sentence, and this is what it returned: Input: ใคใพใใใใใใใใจใงใใ Transcribed output: ใใคใพใใใใใใใใจใงใใใดใงใผใ Input: ๅคงไธๅคซใงใใใๆฐใซใใชใใงใใ ใใใ Transcribed output: ใๅคงไธๅคซใงใใใๆฐใซใใชใใงใใ ใใใใใใใ There was an extra short sound at the end. It was a meaningless sound, 1-2 characters long. The cause was theโฆ
A 3-character catchphrase that became the model's habit emerged due to the quality gate in the corpus. Whisper transcribed a short sentence, appending an extra sound of 1-2 characters at the end. This extra sound matched the size of the hallucination that passed through the quality gate. The model learned to add a little sound at the end of short sentences, with the mechanism triggered when speech ended and silence began.
Written by urgent.news from Dev.to's reporting โ not their text. Machine-written โ may contain errors; check the original before relying on it.
