Build and run a letter recognition NN model on an STM32L4 Discovery board
Introduction
Prepare development environment
Collect training data
Train the model
Feature extraction
Run the model on development board
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

About this Learning Path

Skill level:  Advanced
Reading time:  1 hr
Last updated:  06 Oct 2026
Skill level:
  Advanced
Reading time:
  1 hr
Last updated:
  06 Oct 2026
Author:Pareena Verma, Arm
Arm IP:Cortex-M
Tags: ML Baremetal TensorFlow STM32
Author:
  • Pareena Verma, Arm
Arm IP:
Cortex-M
Tags:
ML Baremetal TensorFlow STM32

Who is this for?

This is an advanced topic for software developers interested in building network models for microcontrollers.

What will you learn?

After completing this Learning Path, you'll be able to perform the following tasks:

  • Build a letter recognition neural network(NN) model using TensorFlow framework
  • Run the NN model on an STM32 B-L475E-IOT01A2 board using STM32CubeAI

Prerequisites

Before starting, you'll need the following prerequisites:

  • Familiarity with ML concepts
  • Familiarity with C programming on microcontrollers
  • STM32 B-L475E-IOT01A2 board

Summary

AI-assisted

This summary was drafted with an approved AI-assisted workflow and reviewed by Arm contributors before publication. Human technical review remains part of the process so the final page reflects engineering rigor, accuracy, and Arm editorial standards.

Close
?
You’ll build a TensorFlow model that recognizes letters from motion data on an STM32 B-L475E-IOT01A2. First, you’ll use Anaconda and a Jupyter notebook to collect accelerometer samples and train a multilayer perceptron. Then, you’ll extract statistical features, retrain the model, import it with STM32Cube.AI, and run predictions with the board’s onboard accelerometer.

Frequently asked questions

AI-assisted

These FAQs were drafted with an approved AI-assisted workflow and reviewed by Arm contributors before publication. Human technical review remains part of the process so the final page reflects engineering rigor, accuracy, and Arm editorial standards.

Close
?
How do I know when a Jupyter cell has finished running?
Jupyter shows In[*] while a cell is running. When complete, it changes to In[N], where N is the execution count.
Where do the training samples come from and what files should I expect?
You’ll collect accelerometer data from the STM32 board and save it as CSV files. The notebook discovers these CSV files from the samples_dir location and loads them for training.
What input shape does the neural network expect and how are labels prepared?
The example model uses an input shape of (3, stride), representing three accelerometer axes over a time window. Labels are one-hot encoded using tf.keras.utils.to_categorical with the class count derived from the unique labels.
What features are extracted before retraining the model?
The process computes the mean and standard deviation for each accelerometer axis and saves these as features. You then train a model on these feature vectors instead of the raw time-series samples.
Which options should I use in STM32CubeMX to set up the project and import the model?
Use the Board Selector to choose the B-L475E-IOT01A board, then set the project name and location in Project Manager. Import the trained model using STM32Cube.AI within STM32CubeMX, and continue with Pinout & Configuration.
Next