Phone recognition in critical bands using sub-band temporal modulations

This study investigates a multistream phone recognition system, which consists of 21 parrallel sub-systems, each covers two critical bands, and fused by a multi-layer perceptron (MLP) system. Within each band, speech information is encoded by the frequency-domain linear prediction (FDLP) feature, which characterizes the temporal modulation of subband envelope. Two experiments are conducted to determine the optimal parameters for speech features, the maximum temporal modulation Fm and the context window length T , followed by an experiment to evaluate the robustness of the fused system in noise. Results show that the phone accuracies of subsystems reach the maximum point at about 500–600ms; they keep increasing monotonically as the maximum frequency of temporal modulation changes from 4 to 40 Hz, where it saturates. Tests of the fused system in babble and subway noise at 15 dB SNR indicate that the multi-stream system is more robust to noise than the single-steam baseline system.