A Low Power Speech Recognition Processor with Precision Recoverable CRNN

Created by W.Langdon from gp-bibliography.bib Revision:1.7630

@InProceedings{Xue:2022:ICSICT,

• author = "AnFeng Xue and Han Yan and RenYuan Zhang and XueTao Wang and Hao Zhang and Hao Cai and Bo Liu2",
• title = "A Low Power Speech Recognition Processor with Precision Recoverable {CRNN}",
• booktitle = "2022 IEEE 16th International Conference on Solid-State \& Integrated Circuit Technology (ICSICT)",
• year = "2022",
• abstract = "In this paper, Convolutional Recurrent Neural Network (CRNN) is deployed in the speech recognition system for recognition of 5 keywords. We propose an 8-bits quantization scheme to quantize the weights and activations. Cartesian Genetic Programming (CGP) to generate approximate multipliers is proposed to drastically reduce hardware power consumption with only 1.5percent loss of accuracy. The proposed partial retraining method compensates for the loss of network accuracy caused by approximate computing, and can basically restore the accuracy to the initial level. Using process library for hardware verification based on synthesis, the proposed speech recognition system can reduce power consumption by 11.percent~24.percent, and reduce area by 10.percent~25.percent with accuracy loss is 0.2percent~ 0.5percent.",
• keywords = "genetic algorithms, genetic programming, Cartesian genetic programming",
• DOI = "doi:10.1109/ICSICT55466.2022.9963226",
• month = oct,
• notes = "Also known as \cite{9963226}",

}

Genetic Programming entries for AnFeng Xue Han Yan RenYuan Zhang XueTao Wang Hao Zhang Hao Cai Bo Liu2

Citations