Effect of Within- and Between-Speaker Variability in Voice Quality on Speaker Recognition

The variability in voice quality is a critical factor in most of speech-related applications, but studies regarding this variability are scarce due to the absence of an adequate database. Based on the newly developing database, this study examines the effect of withinand between-speaker variability on speaker recognition systems. The preliminary results with a subset of the database show that voice quality variability can cause a significant degradation in speaker identification performance. It is also found that widely-used features, such as cepstral coefficients, are not robust in the presence of the voice quality change, but that the robustness can be improved by fusing with perceptually important voice quality features. Further studies will include analyzing the whole database, improving voice quality measurement, and developing a model to represent an individual’s vocal identity.