A production-ready deep learning pipeline that transforms raw audio waveforms into Log-Mel Spectrograms to classify urban environmental noises using a Convolutional Neural Network (CNN).
This ecosystem was co-developed in collaboration with Layla Abdikarim.
- Core Responsibilities: I engineered the modular backend architecture (
src/), implemented the PyTorch 2D CNN model, and built the Streamlit deployment interface. - Collaborative Focus: Jointly iterated on hyperparameter tuning, analysis and preprocessing and analyzed the classification report to diagnose model blind spots.
- Objective: Automatically identify and classify short urban audio clips into 10 distinct environmental categories (e.g., sirens, street music, drilling).
- The Challenge: Raw audio data is irregular in duration and complex to interpret directly. This project leverages computer vision techniques by translating audio frequencies over time into a spatial image format that a neural network can easily learn from.
- Core Stack: Python, PyTorch, Librosa, Matplotlib, Pandas, Streamlit.
This project utilizes the UrbanSound8K dataset, which contains 8,732 labeled sound excerpts (
- Air Conditioner
- Car Horn
- Children Playing
- Dog Bark
- Drilling
- Engine Idling
- Gun Shot
- Jackhammer
- Siren
- Street Music
The dataset is pre-sorted into 10 structural folds to facilitate strict stratified cross-validation and prevent data leakage.
The core technical strength of this project is the audio preprocessing pipeline. Raw signal data is passed through a multi-stage transformation:
- Resampling & Normalization: All files are downsampled to a uniform 22,050 Hz to balance audio fidelity with computational efficiency.
- Short-Time Fourier Transform (STFT): Raw waveforms are broken down into time-frequency frames.
- Log-Mel Conversion: Frequencies are mapped to the human-centric Mel scale and converted to decibels (
power_to_db), yielding a 2D matrix (image) of shape(128, fixed_time_steps). - Padding & Trimming: Clips shorter than 4 seconds are zero-padded, and longer clips are trimmed to ensure every input matrix matches exact dimensions before entering the network.
Below is an example of a raw audio sample transformed into a Log-Mel Spectrogram. This 2D representation allows our vision model to detect frequency patterns over time.
I evaluated two distinct approaches to tackle the classification task:
A lightweight, 4-layer Convolutional Neural Network built from scratch. It uses alternating Conv2D, BatchNorm2D, ReLU, and MaxPooling2D blocks, capped with a Dropout(0.4) layer to minimize overfitting on smaller folds.
To optimize performance, I adapted a pre-trained ResNet-18 architecture. The input layer was modified to accept our single-channel spectrograms, and the final classification head was replaced with a Dense layer outputting 10 logits.
To prevent data leakage, evaluation was performed using Stratified 10-Fold Cross-Validation matching the dataset's native fold design. This ensures chunks of the same original source recording never appear in both the training and testing sets simultaneously.
- Baseline CNN Accuracy: ~XX.X%
- Transfer Learning (ResNet-18) Accuracy: ~XX.X%
Below is the final evaluation confusion matrix.
Key Takeaways:
- The model achieved near-perfect accuracy on distinct, continuous sounds like Sirens and Jackhammers.
- Minor confusion occurred between Dog Barking and Children Playing due to overlapping transient frequency spikes.
This project includes an interactive web dashboard built with Streamlit that allows users to upload their own .wav files and see live inference results alongside the generated spectrogram.
# Clone the repository
git clone [https://github.com/](https://github.com/)[Your-Username]/urban-sound-classifier.git
cd urban-sound-classifier
# Install dependencies
pip install -r requirements.txt
# Run the Streamlit interface
streamlit run app.py