Performance of Text-Independent Speaker Identification considering In-Car Acoustics

Authors: V. Mildner, S. Goetze, K.-D. Kammeyer
Abstract: Hands free operation of communication and information devices in cars is becoming obliged to greater extents. The operated devices reach from mobile phones to systems permanently installed in the car. A major task for such systems is that of speech recognition, for instance handling a navigation system via voice commands only. Algorithms for speaker identification may be used to provide speaker dependent speech recognition systems with the necessary a-priori information. Furthermore, the retrieved information of who is speaking (and operating the car) may be exploited to enable other systems (air conditioning etc.) adapting to the preferences of the driver. For text-independent speaker identification a system based on Gaussian Mixture Models is used. The models are computed based on clean speech sequences for training. Test-sequences are generated considering the acoustic environment (reverberation and noise) of a car cabin. Post-processing is performed applying singleand multi-channel algorithms to the degraded test sequences. Finally, the test-sequences are compared to the models of all speakers via maximum-likelihood decision and a closed-set speaker identification is performed. The performance of text-independent speaker identification is measured by the error rate of the system comparing the different possibilities of post-processing.
Document type: Conference Paper
Publication: Braunschweig, Germany, 20. - 23. March 2006
Conference: 32nd German Annual Conference on Acoustics (DAGA 06)
Pages: 223-224
Index: 263
DAGA_2006_mildner.pdf130 KB
Last change on 30.04.2008 by S. Goetze
AIT ieee tzi ith Fachbereich 1
© Department of Communications Engineering - University of BremenImprint / Contact