This paper proposes a dual-path convolutional recurrent network with group attention for ICASSP Signal Processing Grand Challenge: L3DAS23 Challenge. We design a structure based on convolutional encoder-decoder, and frequency-time blocks based on group attention are introduced in the middle. The encoder is used to extract the local representation from the complex spectrum, the correlation along the frequency axis and the time axis are captured through groups of time-frequency processing modules and the key information in the feature flow is extracted by the group attention. As a result, our system ranks the 1st place of the 3D speech enhancement task in L3DAS23 Challenge, and significantly outperforms the baseline, while achieving 0.101 WER and 0.902 STOI on the blind test-set. © 2023 IEEE.

In the face of numerous hearing loss patients around the world, wearing hearing aids is an effective means to improve the hearing of patients with severe hearing, but the traditional inefficient hearing aid fitting method limits the wide use of hearing aids. With the rise of Over The Counter (OTC) hearing aids, how to effectively improve the performance of OTC hearing aids has become the focus of research in recent years. To solve this problem, an Interactive Genetic Algorithm (IGA) fitting algorithm based on a support vector machine (SVM) is proposed. The compensation values at different frequency points are taken as a set of parameter samples. IGA is used to form an interactive system, the fitness function is replaced by subjective five-level evaluation, and the samples are classified and sorted according to the distance from the classification decision plane by using the user evaluation. The high-quality samples are replaced by the low-quality samples for genetic manipulation, and a new generation of voices is generated for the user to evaluate. The experimental results show that compared with the traditional IGA, the algorithm can improve the comprehensiveness and quality of the generated speech, reduce the number of user interactions significantly, and effectively reduce user evaluation fatigue. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

Speech enhancement (SE) algorithms based on deep neural networks (DNNs) often encounter challenges of limited hardware resources or strict latency requirements when deployed in real-world scenarios. However, a strong enhancement effect typically requires a large DNN. In this paper, a knowledge distillation framework for SE is proposed to compress the DNN model. We study the strategy of cross-layer connection paths, which fuses multi-level information from the teacher and transfers it to the student. To adapt to the SE task, we propose a frame-level similarity distillation loss. We apply this method to the deep complex convolution recurrent network (DCCRN) and make targeted adjustments. Experimental results show that the proposed method considerably improves the enhancement effect of the compressed DNN and outperforms other distillation methods. Copyright © 2022 ISCA.

Wearing hearing aids is one of the effective measures to correct the hearing condition of patients with hearing loss. The choice of the prescription formula is the key to hearing aid gain compensation. DSL-v5 is a commonly used fitting prescription formula, which is designed mainly for pediatric patients. It introduces the Real-Ear-to-Coupler Difference (RECD) test, taking into account the patient's age and hearing threshold, adding an understanding of hearing test methods, hearing aid types, microphone position effects, and hearing assessment with plug-in headphones. To aggregate various factors, this paper proposes a DSL-v5 fitting formula through a least squares-based polynomial fitting method, and verifies the accuracy of this fitting formula by testing experiments. The experiments demonstrate that the fitting formula can well meet the gain compensation needs of adult patients and pediatric patients. © 2022 IEEE.

The presence of wind noise will bring bad experiences to hearing aid wearers, impairing the signal quality and intelligibility. In this paper, a novel frequency-domain algorithm is introduced to detect and suppress the wind noise in the hearing aid and improve the user's comfort. Our method takes advantage of the characteristic that the wind noise energy is mainly concentrated in the low-frequency band and the correlation of wind noise collected by two microphones is significantly weaker than that of speech. Firstly, the cross-correlation coefficients of two input signals at low frequencies are calculated and compared with the preset threshold to determine whether there is wind noise. Secondly, the mathematical expression of the de-noised signal is obtained by using the similarity and spatial correlation of the two near-field signals to suppress the wind noise. The experimental results show that the proposed algorithm can effectively reduce the energy of wind noise, avoid speech distortion and improve speech intelligibility. © 2022 IEEE.