Challenges and Limits in Explaining and Acoustic Modeling of Voice Characteristics

To this day, the assessment of human voices remains a challenge due to (i) inconsistencies in subjective ratings and (ii) the lack of objective measurements for the perceptual impressions of voice characteristics. This can lead to significant consequences in applied fields such as speech therapy, where the assessment of voices is crucial for a successful treatment. In this paper, we address the explanation of voice and its characteristics from two different angles: In a first study, 22 speech therapists in training assessed a set of 20 non-pathological voices regarding 20 voice characteristics before and after receiving an expert explanation. Although the expert explanation did not lead to an improvement in overall rating performance, the analysis still yielded valuable insights into the particular challenges for novice voice practitioners in their characterization of voices. A second study aimed at a better understanding of the link between perceived voice characteristics and acoustic features. A data set of 295 voice samples of the same corpus was labeled by an expert with regard to the same 20 voice characteristics as in the first study. Afterwards, we analyzed the speech samples using a set of acoustic features, which were then used as predictors in statistical models of the annotated characteristics. This analysis yielded a unique set of significant acoustic features as main effects predicting each individual voice characteristic, although the model fits were overall modest. Furthermore, all of the voice characteristic models showed interactions with the speakers' gender. These results suggest a necessity for paying special attention to gender differences when assessing voice. Interestingly, we obtained a tendency for a higher model accuracy for those voice characteristics that have also shown to be rated more accurately and consistently by human listeners.