Low-Complexity DNN-Based End-to-End Automatic Speech Recognition Using Low-Rank Approximation

Targeting the on-device speech-to-text application for streaming inputs, this paper presents an efficient way to reduce the computational complexity of deep neural networks (DNNs) for attention-based speech processing. The proposed technique applies the singular value decomposition (SVD) to the large-sized matrix multiplications, removing less important computations by utilizing the low-rank approximation. The clipping thresholds are carefully adjusted to relax the computing costs as well as the memory overheads while maintaining the recognition accuracy.