Discriminative Universal Background Model Training for Speaker Recognition

Abstract

Speaker recognition (SRE), also called as voiceprint recognition, is the problem of determining the identity of the speaker from a sample of speech signal. It is an important branch of speech signal processing and has many potential applications such as in telephone banking, access control, information security, law enforcement and other forensic applications (Bimbot et al., 2004; Campbell Jr., 1997; Cole et al., 1997; Kinnunen & Li, 2010; Reynolds, 2002). Compared with other biometrics techniques, speaker recognition has its own advantages: (1) It is very convenient, natural and low-cost to acquire the speech sample: it does not need the special devices; the telephone, mobile phone or ordinary microphone is adequate. (2) It can be used remotely: with the ubiquitous telecommunications networks and the Internet, the speech sample can be easily transferred through telephone or VoIP, which makes the remote recognition possible. (3) The speech sample contains many inborn characters: from the speech, we can extract some information about vocal tract, mouth, tongue, soft palate, nasal cavity, and etc. (4) The speech sample also contains some acquired characters, such as tone, volume, pace, rhythm, rhetoric, which reflect speaker’s place of living, education level, and some personal habits information. In speaker recognition, the Gaussian mixture model universal background model (GMM-UBM) is a classical yet widely used method for text-independent speaker verification (Reynolds et al., 2000). In this method, the target speaker is modeled as a GMM and the imposters are modeled as a UBM. When testing, the speech sample is scored as likelihood by the GMM and UBM respectively, and then the likelihood ratio hypothesis test is used for speaker verification. Besides the GMM-UBM, several other methods are developed recently. The most successful ones include the support vector machine using GMM supper vector (GSV-SVM) (Campbell et al., 2006), which concatenate the GMM mean vectors as the input for SVM training and test, and joint factor analysis (JFA) (Kenny et al., 2007), which jointly models the channel subspace and the speaker subspace. Although other methods achieve rapid progress, GMM-UBM is still the basis for their developments. As the meanwhile, the discriminative technologies, such as minimum classification error (MCE), maximum mutual information (MMI), minimum phone error (MPE), feature domain MPE (fMPE), have been achieved great success in speech recognition and language recognition (Burget et al., 2006; Juang & Katagiri, 1992; Povey & Kingsbury, 2007; Woodland & Povey, 2002). 12