Numerous compression and acceleration techniques achieved state-of-the-art results for classification tasks in speech processing. However, the same techniques produce unsatisfactory performance for regression tasks, because of the different natures of classification and regression tasks. This paper presents a novel integer-adder deep neural network (IA-Net), which compresses model size and accelerates the inference process in speech enhancement, an important task in speech-signal processing, by replacing the floating-point multiplier with an integer-adder. The experimental results show that the inference time of IA-Net can be significantly reduced by 20% and the model size can be compressed by 71.9% without any performance degradation. To the best of our knowledge, this is the first study that decreases the inference time and compresses the model size, simultaneously, while producing good performance for speech enhancement. Based on the promising results, we believe that the proposed framework can be deployed in various mobile and edge-computing devices.
Cite as: Lin, Y.-C., Hsu, Y.-T., Fu, S.-W., Tsao, Y., Kuo, T.-W. (2019) IA-NET: Acceleration and Compression of Speech Enhancement Using Integer-Adder Deep Neural Network. Proc. Interspeech 2019, 1801-1805, doi: 10.21437/Interspeech.2019-1207
@inproceedings{lin19b_interspeech,
author={Yu-Chen Lin and Yi-Te Hsu and Szu-Wei Fu and Yu Tsao and Tei-Wei Kuo},
title={{IA-NET: Acceleration and Compression of Speech Enhancement Using Integer-Adder Deep Neural Network}},
year=2019,
booktitle={Proc. Interspeech 2019},
pages={1801--1805},
doi={10.21437/Interspeech.2019-1207}
}