Monaural and binaural cues based auditory scene analyzer

Abstract

The quality of speech signals are highly influenced by the background noises and also room reverberation present in real world environments The human auditory system shows very sophisticated capabilities to analyze complex acoustic mixtures especially in multi talker reverberant environments The Computational Auditory Scene Analysis newline CASA includes the designing of machine hearing systems that utilizes the principles of human auditory system The work discusses both binaural speech segregation and also sound localization in different azimuth as well as distance for artificial listening devices It also focuses on separating the desired target speech from the binaural sound mixtures as a front end processing in cock tail party environment The binaural cues such as Interaural Level Difference ILD Interaural Time Difference ITD and Interaural Coherence IC are extracted from auditory front end processing A reliable soft Time Frequency T F mask is generated by using joint acoustic features such as monaural and binaural cues The concatenated spectral and spatial cues are successfully incorporated into LSTM DRNNs based binaural speech segregation classification framework Also the work considers joint approach of soft time frequency masking functions and discriminative objective learning which are promoted as a deterministic built in layer in a recurrent architecture that helps to improve the speech intelligibility and evaluation measures The performance analysis of different deep learning architectures with several aspects including Deep Neural Networks DNN DRNN with and without joint masking DRNN with and without discriminative objective functions have been carried out by using evaluation metrics such as Source to Interference Ratio SIR Source to Distortion Ratio SDR and Source to Artifacts Ratio SAR newline newline