Improvement in robustness of speech feature extraction method using sub-band based periodicity and aperiodicity decomposition

This posters show improvements in the robustness of a speech feature extraction method using Sub-band based Periodicity and Aperiodicity DEcomposition (SPADE). With SPADE, the speech signal is divided into sub-band signals through bandpass filter banks, after which the subband signals are decomposed into their periodic and aperiodic features by comb filters. The comb filters are designed individually based on the estimated periodicities of each sub-band signal. Both the periodic and aperiodic features are used as speech feature parameters. An evaluation experiment conducted with AURORA-2J (Japanese AURORA-2) showed that SPADE certainly reduces the average word error rate (WER) under open noise condition. However, SPADE degrades the performance under open channel condition. To cope with this problem, we apply cepstral mean normalization (CMN) to SPADE. The result shows that CMN greatly improves the performance not only for test data under the open-channel condition but also for data under the closedchannel condition. SPADE with CMN achieves an average word accuracy of 89.96 %, and an average WER reduction of 28.61 %. This word accuracy is better than that achieved when using MFCC with CMN.