Element selection approach to deep-neural-network-based speech enhancement for small devices under complexity constraints

抄録

In this paper, we propose an element selection approach for speech enhancement using deep neural networks (DNNs) targeting small devices with complexity constraints, such as hearing aids. Element selection reduces the input dimensionality by selecting specific elements from the input vector. Unlike other dimensionality reduction algorithms such as principal component analysis, element selection does not require multiplications, making it suitable for low-complexity environments. To optimize which elements are selected, we propose two methods: 1) a linear-regression-based method minimizing the regression error in estimating a target vector from the dimensionality-reduced vector and 2) a pruning-based method that selects elements corresponding to the remaining weight coefficients in the first layer after applying structured pruning to a DNN. We evaluate their performance in a speech enhancement task under complexity constraints, assuming a simple fully-connected network, with no more than 4×10⁵ multiplications per inference and an algorithmic delay below 8 ms. Experiments show that the proposed approach under the complexity constraints achieves a scale-invariant source-to-distortion ratio (SI-SDR) improvement of 5.6 dB on average compared to non-processed noisy speech at signal-to-noise ratios −5, 0, and 5 dB, and 2.54 dB SI-SDR improvement compared to simply using only the latest frames.