TinyML Wake-Word Detection on Raspberry Pi Pico

Wake-Word in less than $10!

Update 24th May 2021 We now recommend the Adafruit Auto-Gain microphone instead of adjustable gain. We also now include a version for a PDM microphone, this is the recommended microphone to use due to higher quality.

For a deep dive of the code, please check my Hackster.io post.

This application implements the wake word example from Tensorflow Lite for Microcontrollers on the Raspberry Pi Pico.

The wake word example shows how to run a 20 kB neural network that can detect 2 keywords, "yes" and "no". More information about this example is available on the Tensorflow Lite Micro examples folder.

We use as input an electret microphone to detect the words "yes" or "no" and turn the on-device LED on and off in response.

Acknowledgements

This project was created on behalf of the Arm Software Developers team, follow them on Twitter: @ArmSoftwareDev and YouTube: Arm Software Developers for more resources!

Pico Microphone

This project is dependant on the pico-microphone library by Sandeep Mistry. Take a look here for more information about the library.

Contents

• Overview
• Before You Begin
  • Hardware Requirements
  • Hardware Setup
    • Assembly
    • Wiring
  • Software Setup
• Wake-Word uf2 file
• Build Yourself
• Making Your Own Changes
• Contributions
• License

Overview

The micro_speech app for the Raspberry Pi Pico is an adaptation taken from the "Wake-Word" example on Tensorflow Lite for Microcontrollers. Pete Warden's and Daniel Situnayake's TinyML book gives an in-depth look into how this model works and how to train your own. This repository ports the example to work on the Pico.

The application works by listening to the microphone and processing the data before sending it the model to be analyzed. The application takes advantage of Pico's ADC and DMA for the analog microphone and the Pico's PIO and DMA for the PDM microphone to listen for samples, saving the CPU to perform the complex analysis.

The Pico does not come with an onboard microphone. For this application, we use the Adafruit Auto-gain Electret Microphone Amplifier breakout or the Adafruit PDM MEMS Microphone Breakout.

Before You Begin

We will now go through the setup of the project. This section contains three sub-sections, hardware requirements, hardware setup and software setup.

Hardware Requirements

Analog Microphone

• 1x Raspberry Pi Pico
• 1x Electret Microphone Amplifier - MAX9814 with Auto Gain Control
• 1x Micro USB cable
• 3x Jumper wires
• 2x 1x20 male header pins (for the Pico)
• 1x 1x5 male header pins (for the microphone)

PDM Microphone

• 1x Raspberry Pi Pico
• 1x Adafruit PDM MEMS Microphone Breakout
• 1x Micro USB cable
• 5x Jumper wires
• 2x 1x20 male header pins (for the Pico)
• 1x 1x5 male header pins (for the microphone)

Hardware Setup

Assembly

1. Solder headers onto your Raspberry Pi Pico
2. Solder headers onto your Adafruit Electret Microphone

Wiring

The electret microphone breakout is an analog input, meaning we can connect it to one of the ADC pins on the Raspberry Pi Pico. Make the following connections:

Analog Microphone

Adafruit Electret Microphone	Raspberry Pi Pico
OUT	ADC0 - Pin31
GND	Any ground pin
VDD	3V3(OUT) - Pin36

PDM Microphone

Adafruit PDM Microphone	Raspberry Pi Pico
GND	Any ground pin
SEL	Any ground pin
VCC	3V3(OUT) - Pin36
DAT	GPIO2
CLK	GPIO3

Software Setup

The final step before using this application is to set up the software stack (CMake and compilers). The easiest way to do this is to follow the steps on the Raspberry Pi Pico SDK repository. Once done you can test your tolchain setup by running some of the examples found in the Pico examples repository.

You can now clone this repository.

git clone https://github.com/henriwoodcock/pico-wake-word.git
cd pico-wake-word
git submodule init
git submodule update

Wake-Word uf2 file

The easiest way to get started on this application is to use the provided .uf2 file. Doing this means you will not have to build the application yourself. You can download the .uf2 here. Click on download, and your browser should start downloading a file called micro_speech.uf2.

To install this onto your Pico, you need to do the following:

1. Push and hold the BOOTSEL button and plug your Pico into the USB port of your Raspberry Pi or another computer.
2. It will mount as a Mass Storage Device called RPI-RP2.
3. Drag and drop the micro_speech.uf2 binary onto the RPI-RP2 volume.
4. Pico will reboot, and the built-in LED should start blinking.

You can access the output on your Pico by viewing the serial console. There are many applications for this. Raspberry Pi recommends using minicom. However, if you are on macOS, you can use the in-built screen.

You first need to find the name of your Pico.

On macOS:

ls /dev/cu*

On Linux:

ls /dev/tty*

Copy your device name (hint: if you are unsure which device is your Pico: unplug your Pico, re-run the command and then plug your Pico back in, after re-running the command once more, the new device will be your Pico) and then run the following:

Using screen:

screen {device_name} 115200

Using minicom:

minicom -b 115200 -o -D {device_name}

You should now be able to see an output similar to the following:

Build Yourself

With the Pico-SDK setup on your machine, building the application is the same as building any other Pico application.

1. Change directory into this repository

   cd pico-wake-word
   

2. Initialize the submodules

   git submodule init
   cd lib/pico-sdk
   git submodule init
   cd ../../
   

3. Make a build directory

   mkdir build
   

4. Generate the Makefiles

   cd build
   cmake ..
   

5. Finally run the Makefile

   make -j8
   

Once done, your pico_micro_speech_analog.uf2 and pico_micro_speech_pdm.uf2 file is located in build.

Making Changes

If you would like to use a different microphone, different LED, use other pins on Pico or change the audio quality, you will need to know how to make these changes to the application.

Changing the LED

The LED settings can be found in micro_speech/rp2/command_responder.cc. To change the LED to a different pin (instead of the onboard LED), change the line:

#define LED_PIN 25

To change the functionality of the LED, edit the if-else section:

if (found_command == "yes"){
  //turn led on
  gpio_put(LED_PIN, 1);
}
else if (found_command == "no"){
  //turn led off
  gpio_put(LED_PIN, 0);
}

Changing the ADC

The ADC pin is defined in the src/audio_provider.cc script. To change the pin used in the application, change the lines:

#define ADC_PIN 26
#define CAPTURE_CHANNEL 0

Changing the Audio Quality

You can change the audio quality captured in the application. By default, the Tensorflow model expects a 16kHz quality. 16kHz means 16000 samples every second. Below you can see the analog config in the src/audio_provider.cc file.

const struct analog_microphone_config config = {
    // GPIO to use for input, must be ADC compatible (GPIO 26 - 28)
    .gpio = ADC_PIN + CAPTURE_CHANNEL,

    // bias voltage of microphone in volts
    .bias_voltage = 1.25,

    // sample rate in Hz
    .sample_rate = 16000,

    // number of samples to buffer
    .sample_buffer_size = ADC_BUFFER_SIZE,
};

You can change the different settings to best suit your need.

Contributions

There are possibly many ways to improve this work. Please feel free to make a PR for any improvements made.

License

This repository is licensed under Apache License 2.0. Tensorflow and CMSIS-NN are also both licensed under Apache License 2.0, Pico-SDK is licensed under BSD 3-Clause "New" or "Revised" License. For other libraries and packages used in this repository, please find the licenses in the micro_speech/third_party folder.