Testing acoustic voice quality classification across languages and speech styles

Many studies relate acoustic voice quality measures to perceptual classification. We extend this line of research by training a classifier on a balanced set of perceptually annotated voice quality categories with high inter-rater agreement, and test it on speech samples from a different language and on a different speech style. Annotations were done on continuous speech from different laboratory settings. In Experiment 1, we trained a random forest with Standard Chinese and German recordings labelled as modal, breathy, or glottalized. The model had an accuracy of 78.7% on unseen data from the same sample (most important variables were harmonics-to-noise ratio, cepstral-peak prominence, and H1-A2). This model was then used to classify data from a different language (Icelandic, Experiment 2) and to classify a different speech style (German infant-directed speech (IDS), Experiment 3). Cross-linguistic generalizability was high for Icelandic (78.6% accuracy), but lower for German IDS (71.7% accuracy). Accuracy of recordings of adult-directed speech from the same speakers as in Experiment 3 (77%, Experiment 4) suggests that it is the special speech style of IDS, rather than the recording setting that led to lower performance. Results are discussed in terms of efficiency of coding and generalizability across languages and speech styles.