Evaluation Board

The evaluation board used this time is the one shown on the right, which is equipped with Infineon's PSOC™ Edge, an MCU for edge AI. PSOC™ Edge has the following features:

・ ARM Ethos-U55: Efficient neural network calculations with NPU

​-Helium DSP: Efficient signal processing

・ NNLite: Neural network calculations possible with low power consumption

(Infineon's proprietary hardware accelerator)

*This evaluation board will not be available for general sale as of March 2025.
If you are interested, please contact us below.

development tools

In this development, we used "DEEPCRAFT™ Studio" to create the learning model and "ModusToolbox™" for PSOC development. The versions of each are as follows.

ModusToolbox™: 5.2.2135.0 BETA
DEEPCRAFT™ Studio: 3.3.0

Learn more aboutModusToolbox ™

Click here for more information aboutDEEPCRAFT™ Studio

The application to be implemented

We will implement a voice recognition application that identifies keywords from voice collected by a microphone.

[Keywords to be judged]
・Red ("aka")
・Green ("midori")

Implementation Flow Overview

The overall overview is shown in the diagram below. We will explain each step in the following order.

Data collection and labeling → Data management → Data preprocessing → Model construction → Model evaluation (statistics) → Model evaluation (real-time) → C code generation

1. Data collection and labeling

Use DEEPCRAFT™ Studio's Graph UX to collect and label data.
This time we used the Infineon MEMS Mic that came with the evaluation board, but it is also possible to collect data from a PC's microphone.
You can also record while selecting the labels you've created, eliminating the need to label data.

Once data collection and labeling is complete, save it.

Streaming sample program for PSOC is here

2. Data Management

Import the labeled dataset. The data will be divided into Train, Validation, and Test, as shown below.

• Train (training data)
  → Data to "study" the model
• Validation
  → Data equivalent to a "mock test" for studying (practice questions)
• Test (evaluation data)
  → Data for "production testing" or "final testing"
  
  

3. Data Preprocessing

Build a functional block of the function to be used for signal processing. Since it is built according to the data to be detected, know-how is required for each learning model, but there are also functional blocks available as packages, so use them according to the purpose.

⇒ Supported functions

The results of data preprocessing can be checked graphically as shown below.
You can output the results of data preprocessing by clicking the Create Track from Preprocessor button at the bottom left of the image above.

4. Model Building

Train using the neural network provided by the tool.
In this demo, we chose Conv1D (Convolution-1D: a lightweight model effective for time series data).

*You can also define the model network (layer arrangement) yourself.

​⇒Usable layers

5. Model evaluation (statistics)

You can check the statistics for the trained model (.h5 format) downloaded in "4. Model Building".

6. Model evaluation (real-time)

You can use Graph UX to perform model evaluation in real time.
As shown below, the inference results for the actual input data are displayed in real time.

7. C Code Generation

Operation check on actual device

Deploy the model.

The inference results are displayed on a PC using Teraterm.