ASR Errors Bias Text-to-Speech Evaluation via Rare-Word Omissions
October 23, 2026
Evaluation recognizers miss 33-67% of very rare words in human speech, creating a false floor for TTS metrics. Using multiple recognizer families can reduce apparent errors on rare words from 10-13% to 3-5%, but significant bias remains.
HOW THIS AFFECTS YOU
●
builderYour TTS evaluation metrics may be significantly underreporting errors caused by ASR limitations.
●
researcherYou must account for recognizer-induced bias when benchmarking TTS performance on rare vocabulary.