Feature normalization based on non-extensive statistics for speech recognition

► We propose a feature normalization method for robust speech recognition. ► It operates in a spectral domain intermediate between log and linear. ► We name our method q-logarithmic Spectral Mean Normalization (q-LSMN). ► It is based on non-extensive statistics in which additivity does not hold. ► It was better than CMN, MVN, and ETSI AFE in our experiments. Most compensation methods to improve the robustness of speech recognition systems in noisy environments such as spectral subtraction, CMN, and MVN, rely on the fact that noise and speech spectra are independent. However, the use of limited window in signal processing may introduce a cross-term between them, which deteriorates the speech recognition accuracy. To tackle this problem, we introduce the q-logarithmic (q-log) spectral domain of non-extensive statistics and propose q-log spectral mean normalization (q-LSMN) which is an extension of log spectral mean normalization (LSMN) to this domain. The recognition experiments on a synthesized noisy speech database, the Aurora-2 database, showed that q-LSMN was consistently better than the conventional normalization methods, CMN, LSMN, and MVN. Furthermore, q-LSMN was even more effective when applied to a real noisy environment in the CENSREC-2 database. It significantly outperformed ETSI AFE front-end.