OrchestrAIte: Instrument Classification

Table of Contents

• Overview
• Installation
• Usage
• UI Screenshots
• Dataset
• Performance
• Contributors

Overview

OrchestrAIte processes .wav audio input, extracts features, and identifies multiple instruments using a convolutional neural network (CNN).

Key Features

• Accepts .wav audio files as input
• Uses log-mel spectrograms for feature extraction
• Multi-label CNN for instrument identification
• Web interface built with FastAPI and Streamlit
• Deployable via Docker and Google Cloud Run

Installation

Fork this repository and clone it in your virtual environment.

Install Dependencies

pip install -r requirements.txt

Usage

Run Locally (without Docker)

1. Open a terminal and start the FastAPI server using Uvicorn:

   uvicorn api.fast_api:app --reload

2. In a separate terminal, start the Streamlit application:

   streamlit run interface/app.py

Run Locally with Docker

1. Ensure Docker is running, then build and start:

   docker compose up --build

2. Open the Streamlit application in your browser ➡ http://localhost:8501

3. To stop and remove containers:

   docker compose down

Deploy to Google Cloud Run with Docker

1. Set up a Google Cloud Project and enable Cloud Run.
2. Authenticate with Google Cloud.
3. Build and push the Docker image to Artifact Registry.
4. Deploy to Cloud Run.
5. Update API_URL in interface/app.py with the deployed URL.
6. Test the deployment.

Supported WAV File Format

• 32-bit PCM
• Mono
• 44.1 kHz sample rate

UI Screenshots

User interface screenshots (click to enlarge):

Dataset

The training data comes from the MusicNet dataset on Kaggle, which is pre-split into training and test folders. Although MusicNet contains labels for 11 instruments in the training set, only 7 instruments are labeled in the test set. As a result, the model was trained to identify the following instruments:

1. Piano
2. Violin
3. Viola
4. Cello
5. Bassoon
6. Clarinet
7. Horn

Performance

The model was evaluated on the test set with the following results:

Test Loss: 0.08

Test Accuracy: 76.8%

Precision: 95.8%

Recall: 95.6%

While the model performs well on the test set, real-world performance may vary depending on the quality and complexity of the input audio.

Contributors

OrchestrAIte was developed by a four-person team as part of a project at Le Wagon Tokyo. The project was completed in two weeks and demoed on December 6, 2024.