Ninth International Conference on Spoken Language Processing

Pittsburgh, PA, USA
September 17-21, 2006

Highly Noise Robust Text-Dependent Speaker Recognition Based on Hypothesized Wiener Filtering

V. Ramasubramanian (1), Deepak Vijaywargiay (2), Kumar V. Praveen (1)

(1) Siemens Information Systems Ltd., India; (2) International Institute of Information Technology, India

We propose a hypothesized Wiener filtering (HWF) algorithm for noise robust variable-text text-dependent speaker-recognition. The proposed algorithm exploits an important feature of the text - dependent mode of operation of speaker-recognition, namely, the availability of the ‘clean reference templates’ of the words of the ‘password’ text which is supposed to be the text of the input noisy speech. The proposed HWF algorithm is set within the one-pass DP framework proposed by us recently for text-dependent speakerrecognition, which enables use of multiple-templates for each word in the password. We evaluate the proposed HWF algorithm for both speaker-identification and speaker-verification using the TIDIGITS database and show that the proposed HWF algorithm has very high recognition accuracies for both additive white-noise conditions and non-stationary color noise conditions (factory, chopper and babble noises), which are also the typical conditions where conventional spectral subtraction techniques perform poorly.

Full Paper

Bibliographic reference.  Ramasubramanian, V. / Vijaywargiay, Deepak / Praveen, Kumar V. (2006): "Highly noise robust text-dependent speaker recognition based on hypothesized wiener filtering", In INTERSPEECH-2006, paper 1474-Wed1A1O.1.