Developer builds multi-layer pitch accent checker after TTS API returns all zeros
A developer building a narration pipeline discovered that a text-to-speech engine advertised as VOICEVOX-compatible returned zero values for all pitch recalculation calls, effectively breaking two of four planned quality-check layers. The term 'VOICEVOX-compatible' turned out to guarantee only matching API endpoint shapes, not identical behavior from the underlying engine. Attempts to verify pitch accent placement by analyzing the audio's F0 curve also failed, as isolated word synthesis always produced a natural phrase-final pitch drop regardless of accent nucleus settings. The developer ultimately redesigned the quality gates to rely on transcription accuracy and independent accent authority checks as hard gates, with F0 movement treated as advisory and final naturalness approval left to a human reviewer. A lighter dictionary tool, unidic-lite, was found to carry the required accent type data despite being a fraction of the size of the full UniDic package.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in