Multichannel speech separation using hybrid GOMF and enthalpy-based deep neural networks

Abstract

Listen

Translate article

Speech signal is commonly debased by room reverberation and included noises in genuine climates. This paper focuses on disengaging objective speech signals in reverberant conditions from multichannel input signals. To overcome all the existing drawbacks, this work proposes an efficient technique like, multichannel speech signal separation using a new hybrid method that combines grasshopper optimization-based matrix factorization (GOMF) and enthalpy-based DNN (EDNN). To predict and remove the unwanted noise into the multichannel input signal, this paper presents a narrative classification framework in the manner of following steps namely, STFT, GOMF-based rank estimation, identify signal eigenvalues, noise removal, feature extraction well as classification. At first, STFT is utilized to plan the multichannel blend waveforms to complex spectrograms. Then, GOMF is used to estimates the obvious speech signals and noise. After the estimation, important features are extracted. Feature extraction is based on spatial feature, spectral feature, and directional features. To attain the enhanced classification outcomes, the spectrogram is reconstructed based on enthalpy-based deep neural network (EDNN). At last, convert the resultant speech spectrogram back to the extracted output signal based on inverse STFT. Experimental results show that our proposed approach accomplishes the most extreme SNR outcome of − 6dB of 24.0523. Comparable to the DNN-JAT, which achieves 18.50032. The RNN and NMF-DNN had the worst SNR 13.45434 and 12.29991. The proposed outcome is compared with various algorithms and some existing works. Compared with other existing works our proposed methodology achieves higher outcomes.