Normalization and Transformation Techniques for Robust Speaker Recognition

Recognizing a person’s identity by voice is one of intrinsic capabilities for human beings. Automatic speaker recognition (SR) is a computational task for computers to perform a similar task, i.e., to recognize human identity based on voice characteristics. By taking a voice signal as input, automatic speaker recognition systems extract distinctive information from the input, usually using signal processing techniques, and then recognize a speaker’s identity based on the extracted information by comparing it with the knowledge previously learned at a training stage. The extracted distinctive information is encoded in a sequence of feature vectors, which is referred to as frame sequence. In terms of purposes of applications, SR tasks can be classified into two categories: speaker identification and speaker verification. Speaker identification (SI) is an application to recognize a speaker’s identity from a given group of enrolled speakers. If a speaker is assumed to be always in the enrolled speaker group, it is referred to as the closed set speaker identification; Otherwise, it is referred to as the open set speaker identification. On the other hand, speaker verification (SV) is an application to verify a speaker identity by simply making a binary decision, i.e., answering an identity question by either yes or no. SV is one of biometric authentication techniques, along with others, such as fingerprint (Jain et al., 2000) or iris authentication (Daugman, 2004). In the past decades, a variety of techniques for modeling and decision-making have been proposed to speaker recognition and proved to work effectively to some extent. In this chapter, we shall not delve too much into the survey for these techniques, but rather focus on normalization and transformation techniques for robust speaker recognition. For a tutorial of the conventional modeling and recognizing techniques, the reader can refer to (Campbell, 1999; Reynolds, 2002; Bimbot et al., 2004). Here, we just make it explicit that among many techniques the most successful ones are Gaussian mixture model (GMM) and hidden Markov model (HMM). With GMM/HMM, high performance can be achieved in sound working conditions, such as in a quiet environment, and for broadband speech. However, these techniques run into problems in realistic applications, since many realistic applications can not always satisfy the requirements of clean and quiet environments. Instead, the working environments are more adverse, noisy and sometimes in narrow-band width, for instance, telephony speech. Most SR systems degrade their performance substantially in adverse conditions. To deal with the difficulties, robust speaker recognition is such a topic for study. O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg