This is a SoC for neural-network acceleration based on wujian100_open platform, which is a low power open source MCU by T-Head. Specially, the accelerating kernel is designed for Keyword Spotting(KWS) and the preprocess kernel is also integrated. Thus far, the current version has been verified on T-HEAD development board using CDK.
- Integrated the proprecess IP and neural network IP into one IP, further reduce the ultilization of the AXI bus.
- The neural network is quantized by 8 bit to further reduce the usage of BRAM resources and speed up the calculation.
- Add endpoint detection function to furthur improve the recognition accuracy.
- The voice signal is collected by VS1053.
- Realize real-time speech recognition.
This repository includes folders as following:
soc: This folder includes source code of KWS-SoC.
ip: This folder includes preprocess ip and KWS ip.
constraint: This folder includes constraint files.
sdk: This folder includes software files.
cfg.bit: This is a bitstream file generated by Vivado based on project KWS-SOC.
ds_rnn_labels.txt: This file lists the recognizable words.
- Download the
cfg.bitto T-HEAD development board. - Open the cdk project in
sdk\projects\examples\kwsfolder, and download the software to the development board. - Speak words listed in
ds_rnn_labels.txt, and check the results sent back from the board.
For more details about this project, you can visit KWS-SoC——基于Wujian100的音频流关键词检测SoC拓展开发笔记 .