Detecting child speaker based on auditory feature vectors for VTL estimation

We introduce novel auditory features in the hidden Markov model (HMM) system for detecting child speakers. The features derived by the gammachirp auditory filterbank (GCFB) have been demonstrated to be suitable for vocal tract length (VTL) estimation, both theoretically and experimentally. We performed numerical experiments to distinguish between child and adult speakers using HMMs trained on 2,360 speech samples collected through a web-based query interface, and we compared the performance of the common mel-frequency cepstral coefficients (MFCC) and the GCFB-based feature vectors. We also introduced the modulation features as the substitution of delta parameters. It has been clearly demonstrated that the error rate distinguishing a child from an adult is reduced by GCFB. To enhance our method for use as a web application, we applied our original voice-enabled web framework to the front-end interface of the proposed system.