Machine Listening in Multisource Environments (CHiME) 2011

Florence, Italy
September 1, 2011

CHiME Data Separation Based on Target Signal Cancellation and Noise Masking

Zbynĕk Koldovský (1), Jiří Málek (1), Jan Nouza (1), Miroslav Balík (2)

(1) Technical University of Liberec, Czech Republic
(2) Brno University of Technology, Czech Republic

The task of the CHiME challenge is to separate distant speech from noise and recognize the commands being spoken. We propose a separation approach suitable for the CHiME data that relies on the existence of a speech-only recording and assumes fixed positions of the speaker and hearer and stationary reverberant conditions. Using the clean speech segment, a filter that suppresses the target speech is designed. Then, its output provides an estimate of the noise, when the noise is present and comes from different positions. The estimated noise is suppressed from original recordings by an adaptive filter, which outputs the enhanced target signal. The experiments with CHiME data show that this approach improves the keyword recognition accuracy by 7.27%.

Full Paper

Bibliographic reference.  Koldovský, Zbyn.ek / Málek, Jirí / Nouza, Jan / Balík, Miroslav (2011): "CHiME data separation based on target signal cancellation and noise masking", In CHiME-2011, 47-50.