A new two-stage method for speech enhancement based on perceptual enhancement, wavelet packet and nonstationary noise detection and QPSO-based optimization of its hyperparameters

Nowadays, speech enhancement has become one of the most important issues in signal processing. Noise reduction such that it does not disturb the original signal is an important challenge in speech enhancement. In this paper, we have proposed a new hybrid two-stage method for speech enhancement. In the proposed method, noisy speech signal is enhanced using perceptually-motivated Bayesian approach in the first stage. Then, the signal is decomposed into sub-bands using wavelet packet decomposition. In the second stage, each sub-band signal is enhanced using NNESE method. Hyperparameters of NNESE is optimized using QPSO metaheuristic. Finally, all enhanced sub-band signals are combined together using wavelet packet reconstruction. Proposed method is better than others in terms of PESQ and Segmental SNR criteria over a subset of TIMIT speech database polluted by Noisex-92 noise database.