Speaker Identity and Voice Quality: Modeling Human Responses and Automatic Speaker Recognition

Despite recent breakthroughs in automatic speaker recognition (ASpR), system performance still degrades when utterances are short and/or when within-speaker variability is large. This study used short test utterances (2-3sec) to investigate the effect of within-speaker variability on state-of-the-art ASpR system performance. A subset of a newly-developed UCLA database is used, which contains multiple speech tasks per speaker. The short utterances combined with a speaking-style mismatch between read sentences and spontaneous affective speech degraded system performance, for 25 female speakers, by 36%. Because humans are more robust to utterance length or within speaker variability, understanding human perception might benefit ASpR systems. Perception experiments were conducted with recorded read sentences from 3 female speakers, and a model is proposed to predict the perceptual dissimilarity between tokens. Results showed that a set of voice quality features including FO, Fl, F2, F3, H1*-H2*, H2*-H4*, H4*-H2k*, H2le-H5k, and CPP provides information that complements MFCCs. By fusing the feature set with MFCCs, human response prediction RMS error was 12, which represents a 12% relative error reduction compared to using MFCCs alone. In ASpR experiments with short utterances from 50 speakers, the voice quality feature set decreased the error rate by 11% when fused with MFCCs.