Missing feature mask generation in BSS outputs using pitch frequency

Abstract

This paper proposes a new method to generate missing feature mask based on pitch frequency in Blind Source Separation (BSS) outputs. Missing feature theory is a promising approach to improve noise-robustness of automatic speech recognition. The most critical issue in the missing feature theory is automatic generation of the mask. Since frequency of BSS output remains fixed during the mixing and the separating procedures, the proposed method relies on mask generation based on the pitch frequency in BSS outputs to determine unreliable time-frequency components which are destroyed due to crosstalk. Simulation results show that the proposed method outperforms the state-of-the-art algorithms in terms of word accuracy.