TinyML Language Detector with Edge Impulse on Arduino Nano 33 BLE Sense

Necessary tools and machines

	3D Printer (generic)
	Soldering iron (generic)

Apps and online services

	Edge Impulse Studio
	Arduino IDE
	Autodesk Fusion 360

About this project

Overview

This project demonstrates how to build a TinyML language detector on the Arduino Nano 33 BLE Sense using Edge Impulse Studio. The device continuously listens for three spoken words – “si” (Italian), “oui” (French), and “yes” (English) – and activates the corresponding flag when a word is detected.

The workflow comprises:

• Dataset creation and sampling
• Model design and training
• Performance evaluation
• Hardware‑specific optimization with the EON Tuner
• Deployment and integration with servo‑controlled flags

Sampling & Dataset Creation

The dataset contains 30 minutes of recordings: 10 minutes per target word plus 3 minutes of ambient background noise. Each 1‑second clip captures a single utterance, and all files are labeled accordingly in Edge Impulse.

Background samples help the model differentiate speech from non‑speech.

Model Design

Edge Impulse automates most preprocessing steps. Audio is converted to a time‑series, windowed, and transformed using MFCC features. A lightweight CNN then classifies each window into one of four classes: “si,” “oui,” “yes,” or background.

Sample MFCCs and the network architecture are visualized in the training dashboard.

Training & Arduino Optimization

The EON Tuner automatically explores dozens of model configurations, scoring each by accuracy, inference time, RAM, and flash usage. For this project, the fastest model among the top‑5 accuracies was selected.

Testing

After training, a fresh test set confirmed robust performance. Classification confidence consistently exceeded 80% for the target words.

Deployment

Deployment is streamlined via the Edge Impulse Arduino library. The generated header file is dropped into the Arduino IDE, then customized to control three SG90 servos that display flags.

Code Overview

The source code (available on GitHub) extends the Edge Impulse template with servo control logic. Key additions include:

1. Integration of Adafruit_PWMServoDriver.h for PCA9685 PWM control.
2. A servos_selector function that activates the correct flag based on the model’s confidence.
3. Conditional logic that triggers servos_selector when a classification exceeds 80% confidence.

void servos_selector(int iter){
  unsigned long time_now = millis();
  if (time_now - time_was > 2000){
    time_was = time_now;
    switch (iter){
      case 0: // Italian
        pwm.setPWM(0, 0, 350); delay(500);
        pwm.setPWM(0, 0, 200); pwm.setPWM(1, 0, 200); pwm.setPWM(2, 0, 200);
        break;
      case 1: // French
        pwm.setPWM(1, 0, 350); Serial.println("2222"); delay(500);
        pwm.setPWM(0, 0, 200); pwm.setPWM(1, 0, 200); pwm.setPWM(2, 0, 200);
        break;
      case 2: // English
        pwm.setPWM(2, 0, 350); Serial.println("333"); delay(500);
        pwm.setPWM(0, 0, 200); pwm.setPWM(1, 0, 200); pwm.setPWM(2, 0, 200);
        break;
    }
  }
}

Circuit Design

The microcontroller is powered by a 9V battery and uses a PCA9685 board to drive three SG90 servos. The board’s 12V pin is connected to the battery, while the PWM outputs drive the servos via the PCA9685’s I2C interface.

With the hardware, software, and data pipeline in place, the TinyML language detector is ready for deployment.

Code

code

Custom parts and enclosures

Schematics