Voice-Based Speaker Recognition Combining Acoustic and Stylistic Features

We present a survey of the state of the art in voice-based speaker identification research. We describe the general framework of a text-independent speaker verification system, and, as an example, SRI’s voice-based speaker recognition system. This system was ranked among the best-performing systems in NIST textindependent speaker recognition evaluations in the years 2004 and 2005. It consists of six subsystems and a neural network combiner. The subsystems are categorized into two groups: acoustics-based, or low level, and stylistic, or high level. Acoustic subsystems extract short-term spectral features that implicitly capture the anatomy of the vocal apparatus, such as the shape of the vocal tract and its variations. These features are known to be sensitive to microphone and channel variations, and various techniques are used to compensate for these variations. High-level subsystems, on the other hand, capture the stylistic aspects of a person’s voice, such as the speaking rate for particular words, rhythmic and intonation patterns, and idiosyncratic word usage. These features represent behavioral aspects of the person’s identity and are shown to be complementary to spectral acoustic features. By combining all information sources we achieve equal error rate performance of around 3% on the NIST speaker recognition evaluation for two minutes of enrollment and two minutes of test data.