A5007482521- Network Packet Processing and Optimization
- Blind Source Separation Techniques
- Neural Networks and Applications
- Neural Networks and Reservoir Computing
- Advanced Memory and Neural Computing
- Brain Tumor Detection and Classification
- Ferroelectric and Negative Capacitance Devices
- Advanced Adaptive Filtering Techniques
- Image and Signal Denoising Methods
- Machine Learning and ELM
- Speech and Audio Processing
- Speech Recognition and Synthesis
Southeast University
2019-2020
An ultra-low power always-on keyword spotting (KWS) accelerator is implemented in 22nm CMOS technology, which based on an optimized convolutional neural network (CNN). To reduce the consumption while maintaining system recognition accuracy, we first perform a bit-width quantization method proposed CNN to data/weight bit width required by hardware computing unit without reducing accuracy. Then, propose approximate architecture for quantized using voltage-domain analog switching multiplication...
This paper proposed an energy-efficient reconfigurable accelerator for keyword spotting (EERA-KWS) based on binary weight network (BWN) and fabricated in 28-nm CMOS technology. system consists of two parts: the feature extraction melscale frequency cepstral coefficients (MFCC) keywords classification a BWN model, which is trained through Google's Speech Commands database deployed our custom. To reduce power consumption while maintaining recognition accuracy, we first optimize MFCC...
A low-power high-accuracy reconfigurable processor is proposed for noise-robust keywords recognition and evaluated in 22nm technology, which based on an optimized one-dimensional convolutional recurrent neural network (1D-CRNN). In traditional DNN-based system, the speech feature extraction algorithms DNN classification are two independent modules. Compared to architecture, both processed by 1D-CRNN with weight/data bit width quantized 8/8 bits. Therefore unified training optimization...
This paper proposed a background-noise self-adaptive voice activity detection (VAD) accelerator using SNR prediction based precision dynamic reconfigurable approximate computing. To improve the energy efficiency while maintaining high recognition accuracy for different background noises, two optimization techniques are proposed. Firstly, we module to analyze and pre-classify back-ground noise into levels, binarized weight network (BWN) with data bit width implement feature classification of...
This paper proposed an ultra-low power keyword-spotting (KWS) accelerator using circuit-architecture-system co-design and precision self-adaptive approximate computing based binarized weight network (BWN). To reduce the consumption while maintaining system recognition accuracy for different background noise, we first a bit-by-bit layer-by-layer quantization method to quantize deep neural (DNN) BWN. Then, addition unit further BWN energy consumption. Evaluated under TSMC22nm ULL process...
In this paper, an ultra-low power always-on keyword spotting (KWS) accelerator is implemented based on optimized convolutional neural network (CNN). To reduce the consumption while maintaining system recognition accuracy, we proposed a deep-shift (DSNN) to hardware resources requirements without accuracy loss. This new form of conventional CNNs and fully connected networks can almost waive multiplication operations greatly consumption. Implementation results show that DSNN support four...