The Research And Implementation Of Parallelization Method For Identification And Localization Of Sound Source

With the advent of the "Internet +" and the era of big data, voice interaction of intelligent terminal has drawn wide attention and the three-dimensional sound has played a vital role in achieving real-time human-computer interaction. The identification and localization technology of sound source, which has wide developing prospect, is widely used in civil and military field. However, complicated calculation and the huge data scale of auditory localization have restricted the real-time capability of the application program.Therefore, this paper would propose a parallelization method for sound identification and positioning. The following work has been carried out based on pickup array and the parallel calculation of GPU.(1) Based on the existing methods for sound source identification and localization, the paper studies the technologies for the pre-processing technique of the speech signal, and introduces common methods for the voiceprint identification and sound source localization.It also analyses the sound source identification and localization model which based on the pickup array.(2) On the basis of the traditional voiceprint recognition neural network, the deep learning technology is applied to the neural network of voiceprint identification. Voiceprint identification model based on DBN has overcome the shortcoming of traditional neural network, and has given the methods to improve the identification of the voice source in the target scope.(3) In the original sound source localization model of TDOA, the technology of signal enveloped analysis is combined with the time delay estimation algorithm, the generalized cross-correlation time delay estimation algorithm based on enveloped matching method.The location compute method and performance of different sound source localization models are studied and compared. Methods of target speech separation and enhancement are studied based on time-frequency masking by using the location feature of pickup array.(4) For the strong independence and method's consistency of voice signal treatment,the paper adopts the parallel computing method based on CUDA architecture to improve the training process of DBN model, the analysis of time delay estimation and time-frequency masking algorithm of signal fusion, as well as, to improve the speed of identification positioning.The experiment shows that the parallelization methods for the sound source identification and localization model based on pickup array can effectively identity and position determination of the target, with better anti-noise performance and high efficiency,so as to meet the real-time requirements. This method provides an effective method for the high performance treatment of voice signal in big data environment.