Speech Separation System Based On Trinicon Algorithm

With the rapid development of communications,intelligent speech interaction devices and the Internet of Things(Io T),speech enhancement algorithms based on microphone arrays have been widely used on a lot of audio terminal equipment.When the noise is the speech from interfering speakers,it will have similar statistical features to the desired speech.Speech separation is an effective strategy dealing with this scenario.Both beamforming and blind source separation(BSS)can be utilized as speech separation methods.Comparatively,the BSS algorithm requires fewer array elements and is insensitive to the consistency between microphones.Besides,BSS algorithms are capable of separating speech signals without the information of accurate location of the desired speaker,resulting in better performance in adverse environments.This thesis focuses on the blind source separation algorithm based on TRINICON(Triple-N ICA for convolutive mixtures),aiming at improving the separation performance of the system when the speech signals are discontinuously mixed.An appropriate initialization of the dual-channel TRINICON algorithm is proposed to guarantee fast convergence and stable performance under different spatial distribution.A multi-source activity detection method is also proposed to locate the active period of each source,based on which the filter updating strategy for offline TRINICON algorithm is regularized to improve the separation performance.A source counting method based on DBSCAN(Density-Based Spatial Clustering of Applications with Noise)is proposed to detect speech status of every segment efficiently.Combined with the TRINICON algorithm,this status detection can effectively improve performance of the online TRINICON algorithm especially when speech signals are sparsely mixed.The objective metric provided by the BSSEVAL toolkit is utilized to evaluate the performance of the proposed schemes in simulations and experiments.For speech separation in automobile environments,a specific initialization is designed based on the distributed microphone array to ensure the accurate mapping between speakers and output channels.A feature vector-based multi-source activity detection method according to the amplitude-frequency response of the distributed microphone array is proposed.Then it is introduced into the TRINICON algorithm,contributing to extracting the desired speech under complicated circumstances in automobiles more effectively.Simulations with recordings on the actual car demonstrate the efficacy of the proposed method.