karolpiczak / EARS

EARS: Environmental Audio Recognition System

EARS is a proof of concept implementation of a convolutional neural network for live environmental audio processing & recognition on low-power SoC devices (at this time it has been developed and tested on a Raspberry Pi 3 Model B).

EARS features a background thread for audio capture & classification and a Bokeh server based dashboard providing live visualization and audio streaming from the device to the browser.

Caveats:

EARS is quite taxing on the CPU, so some proper cooling solution (heatsink) is advisable. Nevertheless, when not using the Bokeh app too much, it should work fine even without one.

The live audio stream can get choppy or out-of-sync, especially when using the mute/unmute button.

Actual production deployments would profit from a server-node architecture where SoC devices are only pushing predictions, status updates and audio feeds to a central server handling all end user interaction, material browsing and visualization. This may be implemented in future versions, but no promises here.

Quick look

Installation

EARS has been developed and tested on a Raspberry Pi 3 Model B device. To recreate the environment used for developing this demo:

Step 1 - prepare a Raspberry Pi device

• Get a spare Raspberry Pi 3 Model B with a blank SD card.
• Install a Raspbian Jessie Lite distribution (tested on version April 2017):
  • Download a Raspbian Jessie Lite image from RaspberryPi.org.
  • Use Etcher to flash the SD card (see Raspberry Pi docs for details).
• Boot the device with the new card.
• Attach some input & display devices for configuration.
• Login using default credentials (user: pi, password: raspberry).
• Setup Wi-Fi access (see Wi-Fi config on Raspberry Pi).
• Use sudo raspi-config to enable SSH.
• Recreate SSH host keys:

sudo rm /etc/ssh/ssh_host_*
sudo dpkg-reconfigure openssh-server

Step 2 - install Python 3.6 using Berry Conda

• Install conda for armv7l to /opt/conda:

wget http://repo.continuum.io/miniconda/Miniconda3-latest-Linux-armv7l.sh
chmod +x Miniconda3-latest-Linux-armv7l.sh
sudo ./Miniconda3-latest-Linux-armv7l.sh

• Add export PATH="/opt/conda/bin:$PATH" to the end of /home/pi/.bashrc. Then reload with source /home/pi/.bashrc.

• Install Python with required packages:

conda config --add channels rpi
conda create -n ears python=3.6
source activate ears
conda install cython numpy pandas scikit-learn cffi h5py

• Make sure PortAudio headers are available. If not, installing pyaudio will complain later on:

sudo apt-get install portaudio19-dev

Step 3 - download EARS and install requirements

• Download EARS source code and unpack it to /home/pi/ears. Then install the required packages by issuing:

pip install -r /home/pi/ears/requirements.txt

• Plug a Zoom H1 microphone into the USB port (or some other audio device, but that's the one I used for initial testing), switch it into an audio interface mode (44.1 kHz/16 bit), and verify it's listed by python -m sounddevice.
• Update the --allow-websocket-origin option inside /home/pi/ears/run.sh file with the IP address of the Raspberry Pi device.
• Finally, run the app with:

chmod +x /home/pi/ears/run.sh
cd /home/pi/ears
./run.sh

• Point the web browser to: http://RASPBERRY_PI_IP:5006/

Training new models

For the time being, EARS comes preloaded with a very rudimentary model trained on the ESC-50 dataset (convnet consisting of 3 layers, 3x3 square filters), so it's recognition capabilities are limited for actual live scenarios.

If you want to train the same model on a different dataset:

• Download the source code to a workstation/server with a GPU card.
• Put all audio files (WAV) into ears/dataset/audio.
• Replace the ears/dataset/dataset.csv file with new CSV:

filename,category

• Run python train.py - this should result in the following files being generated on the server:

• Upload the new model files to the Raspberry Pi device and restart the app.

If you want to train a completely different model, then you can have a look at train.py. In this case you probably know what to do either way.

Photos from my development field:

License

MIT © Karol J. Piczak