With the increasing number of AI-based voice recognition devices in the smart city applications, the devices require instant response speed, privacy protection for user and ultra-low power consumption, thus Keyword Spotting (KWS) system is the solution to guarantee three points.
Dr. YK Li, Chief Technology Officer of Simbury said, “Low computing complexity and simplifying hardware implementation are two essential elements for a long-lasting smart device, our team invented a new feature extraction method which is ensuring over 93% accuracy with less convolution layer”.
Audio features are usually quantized with multiple-bit binary code to maintain high classification accuracy, which requires addition operations in the first convolutional layer of the binary neural network (BNN) model. Simbury team invented a new feature extraction method called ‘Thermometer Codes of MFSC (MFSC-TC)’, which directly generates thermometer-code features after Mel filtering by comparing Mel spectrum with exponential thresholds.
Dr. Li further explained, “Having several experiments based on the Google Speech Commands dataset, MFSC-TC technique has advantages on both implementation complexity and test accuracy for BNN accelerator-based KWS systems. The team did a great job!”
The innovation for audio feature extraction has been protected by the patent. Furthermore, the Paper co-authored by Dr. Yuzhong Jiao, Dr. YK Li, Dr. Chi Hong Chan, Dr. Yun Li and Ms. Zhilin Ai of Simbury has been accepted by Institute of Electrical and Electronics Engineers (IEEE) and will also be introduced at 2022 IEEE the 18th Asia Pacific Conference on Circuits and Systems (APCCAS) held in Shenzhen, China during 11 to 13 November 2022 while published on IEEE Transaction on Circuits and Systems I (TCAS-I), IEEE Transaction on Circuits and Systems II (TCAS-II) and IEEE Open Journal of Circuits and Systems (OJCAS).
Simbury is dedicated to exchange the latest findings in circuits and systems and contribute to technological innovation and professional development of the semiconductor industry.