Abstract
We apply the ETSI’s DSR standard to speaker verification over telephone networks and investigate the effect of extracting spectral features from different stages of the ETSI’s front-end on speaker verification performance. We also evaluate two approaches to creating speaker models, namely maximum likelihood (ML) and maximum a posteriori (MAP), in the context of distributed speaker verification. In the former, random vectors with variances depending on the distance between unquantized training vectors and their closest code vector are added to the vector-quantized feature vectors extracted from client speech. The resulting vectors are then used for creating speaker-dependent GMMs based on ML techniques. For the latter, vector quantized vectors extracted from client speech are used for adapting a universal background model to speaker-dependent GMMs. Experimental results based on 145 speakers from the SPIDRE corpus show that quantized feature vectors extracted from the server side can be directly used for MAP adaptation. Results also show that the best performing system is based on the ML approach. However, the ML approach is sensitive to the number of input dimensions of the training data.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Preview
Unable to display preview. Download preview PDF.
References
Euler, S., Zinke, J.: The influence of speech coding algorithms on automatic speech recognition. In: Proc. ICASSP, pp. 621–624 (1994)
Lilly, B.T., Paliwal, K.K.: Effect of speech coders on speech recognition performance. In: Proc. ICSLP, October 1996, vol. 4, pp. 2344–2347 (1996)
Pearce, D.: Enabling new speech driven services for mobile devices: An overview of the ETSI standards activities for distributed speech recognition front-ends. In: AVIOS 2000: The Speech Application Conference (2000)
ETSI ES 202 050 V1.1.1 (2002-10), Speech Processing, Transmission and Quality Aspects (STQ); Distributed Speech Recognition; Advanced Front-end Feature Extraction Algorithm; Compression Algorithms (October 2002)
Kelleher, H., Pearce, D., Ealey, D., Mauuary, L.: Speech recognition performance comparison between DSR and AMR transcoded speech. In: Proc. ICSLP 2002, pp. 1873–1876 (2002)
Dempster, P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the EM algorithm. J. of Royal Statistical Soc., Ser. B. 39(1), 1–38 (1977)
Reynolds, D.A., Quatieri, T.F., Dunn, R.B.: Speaker verification using adapted gaussian mixture models. Digital Signal Processing 10, 19–41 (2000)
Martin, A., Doddington, G., Kamm, T., Ordowski, M., Przybocki, M.: The DET curve in assessment of detection task performance. In: Proc. Eurospeech 1997, Rhodes, Greece, pp. 1895–1898 (1997)
Author information
Authors and Affiliations
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2004 Springer-Verlag Berlin Heidelberg
About this paper
Cite this paper
Sit, CH., Mak, MW., Kung, SY. (2004). Maximum Likelihood and Maximum a Posteriori Adaptation for Distributed Speaker Recognition Systems. In: Zhang, D., Jain, A.K. (eds) Biometric Authentication. ICBA 2004. Lecture Notes in Computer Science, vol 3072. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-540-25948-0_87
Download citation
DOI: https://doi.org/10.1007/978-3-540-25948-0_87
Publisher Name: Springer, Berlin, Heidelberg
Print ISBN: 978-3-540-22146-3
Online ISBN: 978-3-540-25948-0
eBook Packages: Springer Book Archive