Capturing time-frequency patterns along the frequency axis, is crucial for the precision of sound event detection systems. Frequency dynamic convolution (FDY) and a series of its variants which incorporate frequency-adaptive kernels in standard 2D convolutions, have demonstrated remarkable performance, yet also suffered from high computational costs. To address the issue, we propose an efficient and light-weighted frequency-aware Fourier filter (FAF-Filt), which performs a 2D Fourier transform on features to the frequency domain and employs a learnable frequency-aware filter to process the transformed features, thereby integrating global information more effectively to extract decisive frequency components. In addition, frequency-adaptive convolution (FA-Conv) is adopted to further strengthen the representative ability of convolution, which incorporates the frequency-aware attention mechanism into the inputs and outputs of the convolutions. Experimental results exhibit superiority of the proposed method, achieving comparable performance with FDY-CRNN in terms of polyphonic sound event scores (PSDS) with a significantly 56% reduction in parameters.

Acoustic echo cancellation is a key issue in hand-free communication systems. In this paper, we proposed a hybrid signal processing and deep echo cancellation method, where a two-stage neural network is designed to remove residual echo progressively. For the personalized acoustic echo cancellation, we proposed to decouple the tasks of echo cancellation and target speech extraction, and introduced a speaker attentive module for personalized separation, where the ECAPA-TDNN is used for speaker embedding generation. The proposed method (ByteAudio-18) ranked first on both Track 1 and Track 2 in ICASSP 2023 AEC Challenge.

Recent studies in deep learning based acoustic echo cancellation proves the benefits of introducing a linear echo cancellation module. However, the convergence problem and potential target speech distortion impose an additional learning burden for the neural network. In this paper, we propose a two-stage progressive neural network consisting of a coarse-stage and a fine-stage module. For the coarse-stage, a light-weighted network module is designed to suppress partial echo and potential noise, where a voice activity detection path is used to enhance the learned features. For the fine-stage, a larger network is employed to deal with the more complex echo path and restore the near-end speech. We have conducted extensive experiments to verify the proposed method, and the results show that the proposed two-stage method provides a superior performance to other state-of-the-art methods.

Recent studies in deep learning based speech enhancement have seen great progress. However, it remains a challenging problem to balance between high accuracy and complexity of the speech enhancement models. To address this issue, we propose a novel speech enhancement framework that consists of a two-stage training module and a co-worker based speech enhancement network (Co-worker-SENet). In the training phase, we first train a teacher model to extract basic features. Then, a student model learns from the teacher at some early steps and further refines the features. Both the teacher and student models are Co-worker-SENet where a stack of feature extraction (FE) blocks is used to learn and refine the features. The FE block consists of multiple cheap workers, which extract features independently. Those features are then fused as the output of the FE block. We conduct extensive experiments on the commonly used VoiceBank-Demand dataset, and the experimental results show that the two-stage training framework can effectively improve the performance of the one-stage method and performs comparably to other state-of-the-art approaches with simple and cheap operations.