Urban Sound Classification with Convolutional Neural Networks

Overview

This project was completed as part of the Statistical Methods of Machine Learning course, supervised by Prof. Nicolò Cesa-Bianchi — a world-renowned researcher in online learning and statistical ML theory. The task: classify 10 categories of urban environmental sounds using deep learning.

Problem Framing

Raw audio waveforms are high-dimensional and temporally complex. The key insight is to transform audio into a 2D time-frequency representation (MFCC / Mel-spectrogram) and treat it as an image classification problem, enabling CNN architectures to leverage spatial correlation in the frequency domain.

Dataset: UrbanSound8K

• 8,732 labelled audio clips, each ≤4 seconds
• 10 classes: Air Conditioner, Car Horn, Children Playing, Dog Bark, Drilling, Engine Idling, Gunshot, Jackhammer, Siren, Street Music
• Sourced from Freesound.org — real-world recordings with background noise
• Pre-defined 10-fold cross-validation split to prevent data leakage

Feature Extraction

Audio was processed using Librosa. Multiple spectral representations were extracted and compared:

• MFCC (40 coefficients) — primary feature; captures perceptual frequency characteristics
• Mel-Spectrogram — non-linear frequency scale matching human auditory perception
• Chromagram — pitch class distribution, useful for tonal sounds
• Spectral Centroid — brightness measure, distinguishes impulsive vs sustained sounds
• Spectral Roll-off — frequency below which 85% of energy is concentrated

import librosa
import numpy as np

def extract_features(file_path: str, n_mfcc: int = 40) -> np.ndarray:
    audio, sr = librosa.load(file_path, res_type='kaiser_fast', mono=True)
    
    mfccs = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=n_mfcc)
    mfccs_scaled = np.mean(mfccs.T, axis=0)  # time-averaged
    
    return mfccs_scaled  # shape: (40,)

CNN Architecture

A sequential 7-layer Conv1D model was designed to process 1D MFCC feature vectors:

• 3× Conv1D layers (32 → 64 → 128 filters, kernel size 2, ReLU activation)
• Dropout (50%) after each convolutional layer to prevent overfitting
• Flatten layer to convert feature maps to a 1D vector
• Dense (64, ReLU) → Dense (10, Softmax) output layer
• Optimizer: Adam | Loss: Categorical Cross-entropy | Epochs: 100 | Batch: 32

Baseline Comparison

Three classical ML classifiers were benchmarked against the CNN using the same MFCC features:

• K-Nearest Neighbors: ~80% accuracy
• Decision Tree: ~65% accuracy
• Random Forest: ~80% accuracy
• CNN: 82% accuracy — best performer, with smoother training curve

Results

• Test accuracy: 82% on UrbanSound8K
• Training accuracy at epoch 100: ~85% (minimal overfitting)
• Best performing classes: Gunshot, Siren (distinctive spectral signature)
• Most confused classes: Drilling vs Jackhammer (similar repetitive waveforms)

Code

Full implementation available on GitHub: github.com/monkrishabh/SOUND8K

Connection to Current Work

The MFCC-based audio classification pipeline built here directly informs my current production work on TTS/STT systems at Edza.ai. The same Librosa feature extraction approach — MFCC computation, mel-spectrogram generation, spectral feature analysis — is used daily in production audio pipeline evaluation and model quality assessment.