PeterMS123 / KWS-DS-CNN-for-embedded

Tensorflow training scripts for depthwise separable convolutional neural networks for keyword spotting, and C++ code for deployment.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

KWS-DS-CNN-for-embedded

This repository contains modified python scripts, based on the Speech Commands Tensorflow example, for training tensorflow models based on depthwise separable convolutional neural networks for keyword spotting. It also contains scripts for quantizing trained networks and evaluating performance in single-inference tests and on continuous audio streams. C++ source code for implementation of a pretrained network on the Cortex M4 based FRDM K66F development board is also included.

Training and deployment

To train networks and deploying them to ARM Cortex-M boards, the guides from ARM's should be followed.

Note: After cloning the CMSIS-NN library, the "CMSIS\NN\Source\ConvolutionFunctions\arm_depthwise_separable_conv_HWC_q7_nonsquare.c" function must be replaced with the slightly modified version found in this repository.

About

Tensorflow training scripts for depthwise separable convolutional neural networks for keyword spotting, and C++ code for deployment.

License:Apache License 2.0


Languages

Language:C 47.0%Language:C++ 26.6%Language:Python 17.1%Language:Objective-C 9.3%