Frequency-aware convolution for sound event detection

Abstract

In sound event detection (SED), convolutional neural networks (CNNs) are widely employed to extract time-frequency (TF) patterns from spectrograms. However, the ability of CNNs to recognize different sound events is limited by their insensitivity to shifts of TF patterns along the frequency dimension, caused by translation equivariance. To address this issue, a model called frequency dynamic convolution (FDY) has been proposed, which involves applying specific convolution kernels to different frequency components. However, FDY requires a significantly larger number of parameters and computational resources compared to a standard CNN. This paper proposes a more efficient solution called frequency-aware convolution (FAC). FAC incorporates frequency positional information by encoding it in a vector, which is then explicitly added to the input spectrogram. To ensure that the amplitude of the encoding vector matches that of the input spectrogram, the encoding vector is adaptively and channel-dependently scaled using self-attention. To evaluate the effectiveness of FAC, we conducted experiments within the context of the DCASE 2023 task 4. The results show that FAC achieves comparable performance to FDY while requiring only an additional 515 parameters, whereas FDY necessitates an additional 8.02 million parameters. Furthermore, an ablation study confirms that the adaptive and channel-dependent scaling of the encoding vector is critical to the performance of FAC.