5. Automatic Speaker Verification Performance Under Channel Distortions

Voice biometrics are frequently exposed to channel degradations of transmitted speech, which cause speaker recognition systems to perform poorly. Particularly, there may exist a severe mismatch between enrolment and test utterances when each of the transmissions presents different characteristics, causing an undesired increase of within-speaker variability. In this chapter, three typical automatic speaker verification systems are built and evaluated with speech data presenting different channel degradations in NB and in WB, in matched and in mismatched conditions. The three systems are GMM-UBM [218], JFA [151], and i-vectors [53]. The latter has been proven to offer state-of-the-art performance under the most recent NIST SRE conditions and is commonly employed in today’s speaker recognition applications [194, 229]. These systems have been generally tested employing extensive data released by the Linguistic Data Consortium (LDC), however limited to NB [149, 218, 229], and with mostly each speaker recorded over only one phone channel, which makes it difficult to choose appropriate training data in some NIST evaluations[148]. Unlike the purpose of the NIST challenges, the intention of this chapter is not to outperform existing speaker recognisers. Its objective is instead to offer a comparison among text-independent speaker verification results under different telephone channel distortions. The bandwidth, codec, and packet loss rate of speech transmissions varied among experiments. Commonly used speaker verification systems are employed and the benefits of extending the bandwidth evaluated. It is also examined in the second part of this chapter which distortions should be contained within the development data for an improved system performance.