An improved method of energy-peak incoherent signal-subspace method based on time-frequency sparseness of speech signal and its application

To ameliorate the situation that the incoherent signal-subspace method (ISM) is computationally expensive, an improved method named speech signal classified energy-peak incoherent signal-subspace method (SSC-EPISM) is proposed in this paper based on time-frequency sparseness of speech signal. According to this method, the amount of the speech targets is firstly estimated based on the microphone array input data. After that, we extract the energy-characterized sub-bands (ECSBs) on which the signal energy is concentrated. Then, with speech signal classification method (SSC), whether or not each ECSB carries information of a near-field source can be checked. By analyzing the estimated amount of speech targets, the result of SSC and the data of ECSBs, the aim of localizing the speech targets is achieved. Simulation results indicate that SSC-EPISM successfully localizes the speech targets and its amount of calculation is smaller than ISM and two improved methods of ISM. Also, as an application of SSC-EPISM, a speech enhancement method based on SSC-EPISM is proposed in this paper. The simulation results indicate that this speech enhancement method can successfully purify the noise-contaminated signal picked by the microphone array.