Neural Network-based Speech Enhancement Algorithm For Binaural Hearing Aids With FPGA Implementation

Hearing loss is a major public health problem facing the world,and the wearing of appropriate hearing aids is the primary form of treatment at this stage,in addition to medication.Hearing aids are miniature wearable devices with limitations on the size and power consumption of internal algorithms,and many high-performance speech enhancement algorithms cannot be applied to hearing aids due to their large size and high power consumption.In complex sound field environments,it is difficult for existing hearing aid algorithms to achieve satisfactory results.To address these problems,Thesis proposes a recurrent neural network-based speech enhancement algorithm for binaural hearing aids.The algorithm divides binaural speech into 16 frequency bands according to the Mel scale,and then extracts Mel frequency cepstrum coefficients and interaural phase differences as the input features of the speech enhancement network.In order to highlight the amplitude and phase features of binaural speech,a dual-input,dual-output recurrent neural network model consisting of an amplitude feature processing region,a phase feature processing region and a mixed feature processing region is finally used.The network calculates the output gain of each band and applies it to the corresponding band,and then combines the bands to generate the enhanced speech.Experimental results show that the average binaural signal-to-noise ratio of the proposed algorithm improves by an average of 4.68 d B and the short-time speech intelligibility by an average of 4.5% compared to the algorithms commonly used in hearing aids.Compared with the neural network-based algorithm,the average binaural signal-to-noise ratio is improved by 1.63 d B and the short-time speech intelligibility is improved by 4.8%.In order to verify the feasibility of the algorithm for practical application,the hardware circuit design of the network in this thesis was implemented using the Verilog hardware description language.The experimental results show that the hardware processing results and the software simulation results are basically consistent,and their cross-correlation coefficients are above 0.99.When the clock frequency is 10 MHz,the calculation delay of the network is about 4.2ms,which can meet the real-time requirements of hearing aids.