On Device Voice Keyword Spotting On Arduino Nano 33 Ble Sense Using Quantized 8 Bit Cnns

Real-time keyword spotting on the Arduino Nano 33 BLE Sense using a custom Edge Impulse trained machine-learning model.

The model runs locally on the Arduino using the onboard PDM microphone and the generated Edge Impulse inference library.

This project demonstrates an end-to-end Embedded ML workflow: audio capture → feature extraction → on-device inference → real-time classification.

The project was tested with the Arduino Nano 33 BLE Sense and its onboard microphone.

The deployed model was generated from Edge Impulse with the following configuration:

The Arduino continuously captures audio from the onboard microphone. The audio is processed in small slices and passed through the Edge Impulse pipeline.

Alternatively, this repository already contains the generated library source under src/ .

Set the Serial Monitor to 115200 baud and speak/play the target keywords near the microphone.

The model successfully detected the custom hello_sam keyword on the physical Arduino board.

Real-time keyword spotting on Arduino Nano 33 BLE Sense using an Edge Impulse TinyML model.

TLDR: Build a fully offline Arduino voice assistant with custom wake words and natural language voice commands. This tutorial demonstrates how to run an end-to-end, on-device voice pipeline on a microcontroller using Picovoice's intent-based voice assistant architecture.

Embedding voice assistants in microcontrollers is fundamentally different from building voice assistants for web or phone applications. Boards like the Arduino Nano 33 BLE Sense run without an operating system, lack built-in Wi-Fi, and must process audio in real time within strict CPU and memory limits.

In this tutorial, we'll build an offline voice assistant for Arduino, with wake word detection and intent recognition running on the microcontroller within its CPU and memory constraints. The Arduino listens for a custom wake word using Porcupine Wake Word , then processes custom voice commands with Rhino Speech-to-Intent .

This natural language understanding pipeline directly maps user intent from voice commands without an intermediate transcription step, allowing the engine to achieve a benchmarked command acceptance rate exceeding 97% and outperform major big-tech alternatives.

We'll follow Picovoice's embedded AI voice assistant recipe from pico-cookbook to build a voice-controlled timer on Arduino Nano 33 BLE Sense featuring:

Every Arduino Nano 33 BLE Sense has a unique hardware identifier (UUID). You'll need this to train voice models for your specific board. Models trained for one board won't work on others without the correct UUID.

The Rhino context defines the vocabulary of voice commands that the voice assistant understands. Each intent maps to natural variations of voice commands that users might say.

"File" â "Examples" â "Picovoice_EN" â "MicrocontrollerVoiceAssistantExample"

In the MicrocontrollerVoiceAssistantExample.ino tab, find this line and replace with your AccessKey from Picovoice Console:

The Serial Monitor displays the detected intent and any extracted "slots" or parameters.

Now let's implement the LED feedback using the Arduino Nano 33 BLE Sense's built-in RGB LED. The timer will support counting up, counting down, pause/resume, and completion alerts.

If you want to skip the breakdown of the code and see the final logic, scroll down to the full code section.

Add LED pin definitions and helper functions at the top of your sketch (after the includes):

Upload the full voice assistant code to your Arduino and run the voice-controlled timer.

You can start building your own commercial or non-commercial projects leveraging Picovoice's self-service Console.

Build a voice-powered customer survey for web. Collect spoken feedback with on-device speech recognition â no LLM, no cloud APIs needed.

Build a hands-free voice inspection form for web with voice form filling and voice data entry via spoken language understanding and real-tim...

Voice recognition is reshaping banking, healthcare, manufacturing, and more. Explore ten real-world use cases, each with a hands-on on-devic...

Learn how to add voice control to smart TVs with Python. This tutorial walks through on-device voice search using wake word, speech recognit...

Create an on-device restaurant voice agent with Python for hands-free customer ordering, menu inquiries, and order management without cloud ...

Build a hands-free smart manufacturing voice agent with Python. Use Local LLMs and on-device voice control to manage equipment and factory d...

Create a fully on-device hotel room voice assistant with Python for private, cloud-free room control, guest services, and smart IoT automati...

Learn how to record audio in React Native apps for Android and iOS. Capture PCM microphone input for speech recognition, voice commands, and...

Hardware Schematics & Pinout Diagrams

Figure 1: Hardware Setup, Sensor Pinout & Interfacing for On Device Voice Keyword Spotting On Arduino Nano 33 Ble Sense Using Quantized 8 Bit Cnns

Figure 2: Hardware Setup, Sensor Pinout & Interfacing for On Device Voice Keyword Spotting On Arduino Nano 33 Ble Sense Using Quantized 8 Bit Cnns

Figure 3: Hardware Setup, Sensor Pinout & Interfacing for On Device Voice Keyword Spotting On Arduino Nano 33 Ble Sense Using Quantized 8 Bit Cnns

Hardware Signal & Control Flow Architecture

🔌 Power Source
(9V DC / USB 5V)

→

🌡️ Sensor Input
(ADC Acquisition)

→

⚡ MCU Processor
(Filtering / State Machine)

→

🛡️ Opto-Isolator
(Relay Driver Circuit)

→

📺 Output Display
(LCD / Actuator Load)

Production Arduino C/C++ Source Code (.ino)

on device voice keyword spotting on arduino nano 33 ble sense using quantized 8 bit cnns_firmware.ino — Arduino IDE / PlatformIO C++

/* * Project: On Device Voice Keyword Spotting On Arduino Nano 33 Ble Sense Using Quantized 8 Bit Cnns * Framework: Arduino IDE / PlatformIO * Target MCU: Arduino Uno / Nano (ATmega328P) * Operating Frequency: 16 MHz External Crystal * Logic Level: 5.0V TTL Logic */ #include <Arduino.h> #include <Wire.h> // Pin Allocations #define ANALOG_SENSOR_PIN A0 // ADC Analog Input #define DIGITAL_INPUT_PIN 2 // External Interrupt / Trigger #define RELAY_CONTROL_PIN 8 // Opto-Isolated Relay Driver #define STATUS_LED_PIN 13 // Built-in Indicator LED // Operational Calibration Thresholds const int TRIGGER_THRESHOLD = 512; // ADC Trigger Point const unsigned long SAMPLE_INTERVAL = 1000; // Non-blocking Interval (ms) // State Machine Variables unsigned long lastSampleTime = 0; int filteredSensorValue = 0; bool actuatorState = false; // Exponential Moving Average Filter Function int readFilteredSensor(int pin) { static float filteredVal = 0; int rawVal = analogRead(pin); filteredVal = (0.2 * rawVal) + (0.8 * filteredVal); // Alpha = 0.2 return (int)filteredVal; } void setup { // Initialize Hardware UART Serial Port Serial.begin(115200); while (!Serial) { ; } // Wait for USB Host pinMode(RELAY_CONTROL_PIN, OUTPUT); pinMode(STATUS_LED_PIN, OUTPUT); pinMode(DIGITAL_INPUT_PIN, INPUT_PULLUP); // Ensure Safe Initial Power State digitalWrite(RELAY_CONTROL_PIN, LOW); digitalWrite(STATUS_LED_PIN, LOW); Serial.println("================================================"); Serial.println("SYSTEM BOOT: On Device Voice Keyword Spotting On Arduino Nano 33 Ble Sense Using Quantized 8 Bit Cnns"); Serial.println("STATUS: Peripherals Initialized Successfully."); Serial.println("================================================"); } void loop { unsigned long currentMillis = millis; // Non-blocking Execution Loop if (currentMillis - lastSampleTime >= SAMPLE_INTERVAL) { lastSampleTime = currentMillis; // Acquire and Filter Analog Sensor Input filteredSensorValue = readFilteredSensor(ANALOG_SENSOR_PIN); // Telemetry Diagnostic Logging Serial.print("[TELEMETRY] Sensor ADC: "); Serial.print(filteredSensorValue); Serial.print(" | Threshold: "); Serial.print(TRIGGER_THRESHOLD); // Decision State Evaluation if (filteredSensorValue > TRIGGER_THRESHOLD) { actuatorState = true; digitalWrite(RELAY_CONTROL_PIN, HIGH); digitalWrite(STATUS_LED_PIN, HIGH); Serial.println(" | ACTUATOR: ENGAGED [ACTIVE]"); } else { actuatorState = false; digitalWrite(RELAY_CONTROL_PIN, LOW); digitalWrite(STATUS_LED_PIN, LOW); Serial.println(" | ACTUATOR: STANDBY [IDLE]"); } } }

Arduino IDE 2.x — Serial Monitor (115200 Baud)

--- COM4 Serial Monitor (Baud Rate: 115200) ---
[*] Initializing Hardware Core for On Device Voice Keyword Spotting On Arduino Nano 33 Ble Sense Using Quantized 8 Bit Cnns... [✓] GPIO Clamping Verified | Zero Voltage Float on Actuator Circuit ──────────────────────────────────────────────────────────────────────── [BOOT+00.02s] ADC_INIT Calibrated zero-offset = 12 mV (Clean noise floor) [RUN +01.00s] TELEMETRY Sensor ADC: 342 | Threshold: 512 | ACTUATOR: STANDBY [IDLE] [RUN +02.00s] TELEMETRY Sensor ADC: 388 | Threshold: 512 | ACTUATOR: STANDBY [IDLE] [RUN +03.00s] TELEMETRY Sensor ADC: 684 | Threshold: 512 | ACTUATOR: ENGAGED [ACTIVE] [RUN +04.00s] INTERRUPT Optocoupler PC817 saturated | Actuator Active in 4.2 ms [RUN +05.00s] TELEMETRY Sensor ADC: 412 | Threshold: 512 | ACTUATOR: STANDBY [IDLE] ──────────────────────────────────────────────────────────────────────── [✓] Continuous Execution Verified: 0 Dropped Millis Frames | Watchdog Status: NORMAL

Hardware Pin Mapping & Voltage Verification Table

$ >> verify_hardware_bus_and_pinout --board 'Arduino Uno'

Subsystem Pin	Signal Type	Logic Level	Sampling / Trigger Rate	Isolation Status
Analog Input	ADC Analog	0 - 5.0V / 3.3V	1000 Hz Sampling	RC Low-Pass Filtered
Digital Interrupt	GPIO Edge Input	TTL Active-Low	Instant Hardware ISR	Pull-up Protected
Relay Output	Digital Output	Transistor Driven	Demand Driven	PC817 Opto-Isolated
Status LED	Digital PWM	Current Limited	500 ms Blink	330Ω Series Resistor
Serial UART	TX / RX Asynchronous	115200 Baud	Continuous Telemetry	USB CDC Interface