Why AI Models Sometimes Elongate Their Greetings Like 'Kon'nichiwa~'
๐ Originally published (in Japanese) at forge.workstyle.tech . When I had the trained voice model read "ใใใซใกใฏ" (Hello), it stretched the phrase to "ใใใซใกใใ." There was no instruction to stretch it in the script. The feedback was as follows: For "ใใใซใกใฏ," it's pronounced as "ใใใซใกใใ" with an accent on the last syllable. It feels like something is mixed in. "Something mixed in" was accurate, andโฆ
When the trained voice model pronounced the Japanese greeting ใใใซใกใฏ, it elongated the phrase to ใใใซใกใใผ. The transcription was performed using Whisper, which transcribed the audio and compared it with a script, but there was no instruction in the script to stretch the greeting. The feedback indicated that the transcription read ใใใซใกใใผ with an accent on the last syllable, suggesting something was mixed in.
The training corpus contained clips with stretched endings, but the mechanism to detect this was fundamentally non-functional by design. Script matching was done by converting both the script and transcription to kana (hiragana) and comparing them. However, the normalization process removed prolongation marks (the ใผ character), making it impossible to detect stretched endings even if they were present in the raw transcription from Whisper.
The normalization was intended to absorb variations in notation and check for content consistency, but it also discarded information about stretched endings. The same issue occurred with consecutive vowels. The problem arose when trying to separate the judgment of content consistency from the judgment of stretched endings, as the raw transcription needed to be passed to the latter function for accurate detection.
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.
