Defects Missed in Transcription — AI Speaks After 0.5-Second Silence
📝 Originally published (in Japanese) at forge.workstyle.tech . Quality Control for TTS Models: Why Transcription Isn’t Enough I used to perform quality control (QC) for TTS models using this process: Have the model read probe sentences Transcribe with Whisper Compare against the script to check accuracy and trailing elongation Analyze the waveform for utterance duration, sound pressure, and F0…
On a TTS model quality check, transcription with Whisper proved ineffective for catching defects. Despite achieving 4 out of 4 accuracy, multiple voices exhibited silent segments followed by brief, unexpected utterances. The root cause was corrupted training data, with three allowed characters becoming verbal tics. Whisper dropped these brief sounds, leading to a failure to detect the defects.
Analysis of the waveform revealed the issue, but the transcription still appeared flawless. The problem stemmed from the STT model's inability to capture short, irrelevant sounds. By examining voiced blocks in the waveform, defects such as silent segments and brief utterances could be identified. However, this required extracting voiced regions from the RMS envelope and comparing the sequence of voiced blocks.
The detection criteria included a minimum gap before the final block and a minimum duration for the last block to identify potential artifacts.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
