The Effects of Spontaneous Speech on Disfluencies Assessment of Spanish Speakers with Down Syndrome

Abstract

The aim of this study is to investigate the phonetic and fluency characteristics of spontaneous speech produced by Spanish speakers with Down syndrome (DS) compared to nonspontaneous speech modes (read, elicited and imitation) and assess the impact of these differences both on expert speech quality assessment and on automatic speech recognition (ASR) performance. The PRAUTOCAL corpus includes four different speech generation modes of utterances spoken by people with DS. The results show that there are minor differences in some features between spontaneous speech and other modes, but specific types of disfluencies and phonetic errors are more prevalent in spontaneous speech. The Whisper model showed improved performance on spontaneous speech, achieving a significantly lower Word Error Rate (WER) and fewer substitution errors. The Wav2Vec phoneme recognition model performed significantly worse, showing higher phoneme error rate (PER), more substitutions, and greater total errors, no matter the automatic segmentation tool used (MFA or WebMAUS).