Single-Channel Speech Enhancement Using Single Dimension Change Accelerated Particle Swarm Optimization for Subspace Partitioning

Speech signal gets contaminated by background noise affecting its quality and intelligibility. There are different sources of additive noise. This additive noise, either stationary or non-stationary, has a distinct distribution of noise energy in the frequency domain. Degraded speech affects the performance of speech-operated systems. Speech enhancement can reduce this additive noise. Here, we propose a subspace-based single-channel speech enhancement method using modified accelerated particle swarm optimization to optimize subspace partitioning. Principal components of noisy speech are partitioned into speech, speech plus noise, and noise only based on the signal-to-noise ratio of principal components. Voice activity detection is implemented to find the variance of additive noise. Modified accelerated particle swarm optimization optimizes the number of principal components in each partition and the weights of the components in each class. The proposed speech enhancement method gives better results for the quality and intelligibility measures of enhanced speech compared with conventional speech enhancement methods. We got 18.8% improvement in STOI for 0 dB restaurant noise, 20.5% improvement for 0 dB train noise, and 11.55% improvement for 0 dB exhibition noise. We got an improvement of 39.15% in PESQ for 0 dB babble noise, 41.57% for 0 dB car noise, and 31.79% increase for 0 dB airport noise. The average improvement in the segmental SNR of the enhanced speech is 8.32 dB for 0 dB noise. There is 4.4 dB improvement in SDR for the airport noise and 5.54 dB improvement for the station noise. We got this improvement with minimum speech distortion.