Speaker Characteristic-based Acoustic Model Adaptation Method for Speaker Recognition Systems

Speaker recognition systems are used to identify persons by their voice. Traditional applications include speaker verification for security purposes (Are you person X?), as well as speaker identification (Who are you?). These systems, although accurate, are not trusted for verification by themselves. The human voice can reveal not only who the speaker is, but their characteristics as well. Information such as the gender, age, and emotional state of a speaker can accurately be identified from voice. Most computer applications of this characteristic information consist of separating speakers into clusters for speaker recognition (to divide recognition effort), adjusting speech models for speech recognition systems, and presenting targeted information to a user base (e.g., teenage male vs. senior advertisements). This thesis presents a method for applying gender and age speaker characteristic classifiers to the traditional speaker recognition process. Typically, a speaker's speaker model is derived from a general model composed of all speakers. This work derives speaker models from characteristic models (e.g., all males, or all females, etc) with an unsupervised classification method. Empirical analysis shows moderate improvement in recognition accuracy