🎵 papermusic ✏️

draw an instrument, and play it! (fun with PaliGemma and SCAMP)

video walkthrough

(click to play)

(📺 Original implementation)

architecture

screenshot

how to run

Note - this is a prototype, not a production-grade app. (The OpenCV streaming setup is a bit buggy and occasionally crashes. You may need to restart.)

You will need:

• A local machine with a webcam, and Python 3.11+. (Client has been tested on MacOS Sonoma / M1 Mac).
• A Google Cloud project, with access to Google Compute Engine.
• A HuggingFace account with an Access Token

Setup (Server)

1. Create a Google Compute Engine instance with at least one GPU. I used two NVIDIA T4 GPUs, but adjust to whatever your quota allows. Set "Allow HTTP/HTTPS" traffic to true.
2. SSH into the instance.
3. Install Python packages for the server workloads.

git clone https://github.com/askmeegs/papermusic 
cd papermusic 
python3 -m venv . 
source ./bin/activate 
pip install -r requirements.txt

4. Set your HuggingFace Access Token as an environment variable.

export HUGGINGFACE_USER_ACCESS_TOKEN=your_token_here

Setup (Client)

1. Clone the repo on your local machine.

git clone https://github.com/askmeegs/papermusic 
cd papermusic 

2. Install client packages.

python3 -m venv . 
source ./bin/activate 
pip install -r requirements.txt

Run all components

Place a hand-drawn musical instrument in front of your webcam, like the screenshot shown above. Make sure the notes are written clearly on the instrument.

1. Start server/handlestream.py on GCE, to listen for the webcam stream.
2. Start client/webcamclient.py on your local machine, to send your webcam stream to GCE.
3. Start server/server.py on GCE, to process the webcam stream and identify the notes.
4. Start client/audioclient.py on your local machine, to poll the Server and play the notes over local audio.

📚 sources

• PaliGemma (vision-language model)
• HuggingFace Transformers
• HuggingFace Inference Optimization
• SCAMP (Suite for Computer-Assisted Music in Python)
• ASCII Art Generator