Text-Independent Speaker Verification: State of the Art and Challenges

Abstract

Speech is often the only available modality to recognize the identity of a person (over the telephone, the radio, in the dark,...). Automatic speaker recognition has been studied for several decades. In this chapter the state of the current text-independant speaker verification research is reviewed. Basic principles of speaker recognition are first summarized. The choice of the speech features and speaker models are mostly related to the individual characteristics (variability) of the speakers’ voices. Besides the speaker’s variability, we are faced with other factors, such as microphone or transmission channel variabilities, that degrade the performances of speaker verification algorithms. Some of these issues are illustrated on recent NIST–2005 and 2006 speaker recognition evaluation campaigns.The field of speaker verification is also reviewed in relation to speech recognition, focusing on the usage of this new source of information. This relationship has to be seen as an important issue in the development of new services based on speaker and speech recognition. An overview of recent results in this field is given. More particularly, examples of combining baseline Gaussian Mixture Models (GMM) with high-level information extracted with data-driven speech segmentation are reported.KeywordsSupport Vector MachineGaussian Mixture ModelSpeech DataEqual Error RateSpeaker RecognitionThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.