🎧 Cats & Dogs Audio Classification — Overview

A binary audio classification project that distinguishes cat meowing from dog barking using raw .wav files. Demonstrates end-to-end ML pipeline construction — from unstructured audio data through feature engineering to trained deep learning and classical ML classifiers.

---

ML Pipeline

flowchart LR
    A[Raw .wav Files<br/>277 samples] --> B[EDA<br/>Signal lengths, sample rates]
    B --> C[Resampling<br/>Normalise to 22,050 Hz]
    C --> D[Length Normalisation<br/>Pad/truncate → 160,000 samples]
    D --> E[MFCC Extraction<br/>13 coefficients per frame]
    E --> F[CNN Training<br/>4-layer Conv2d + 5-Fold CV]
    E --> G[Logistic Regression<br/>Baseline comparison]
    F --> H[Inference<br/>Per-sample prediction]
    G --> H

---

Dataset

Property	Value
Total samples	~277 .wav files
Classes	Cat (164 samples), Dog (113 samples)
Sample rate	22,050 Hz
Fixed length	160,000 samples (~7.3 seconds)

---

Technology Stack

Layer	Technology	Purpose
Language	Python 3.11	Application runtime
Deep Learning	PyTorch (torch.nn, torch.optim)	CNN architecture, training loops
Audio Processing	torchaudio, librosa	MFCC, mel spectrograms, STFT, resampling
Traditional ML	scikit-learn	Logistic Regression baseline, K-Fold CV
Data Handling	Pandas, NumPy	DataFrames, array operations
Visualisation	Matplotlib, librosa.display	Spectrograms, waveforms, histograms
Environment	Google Colab, Jupyter Notebooks	Collaborative development

---

Methodology

1. Exploratory Data Analysis

• Catalogued all audio files and inspected class distribution
• Analysed signal length distribution (histogram) to determine the optimal fixed-length target
• Verified consistent sample rates across the dataset

2. Audio Feature Engineering

Feature	What It Captures
MFCC (13 coefficients)	Timbral/tonal characteristics weighted to human auditory perception
Mel Spectrogram	Time-frequency energy distribution on perceptual scale
STFT	Raw frequency decomposition per time window
Spectral Centroid	"Centre of gravity" of frequency content (overall pitch)
Spectral Rolloff	Frequency threshold containing most energy (high-freq characterisation)

3. Custom PyTorch Data Pipeline

A production-quality CatsAndDogsDataset class extending torch.utils.data.Dataset:

• Resampling — normalises all audio to 22,050 Hz
• Length normalisation — zero-pads short clips, truncates long clips to 160,000 samples
• On-the-fly MFCC transform — torchaudio.transforms.MFCC(n_mfcc=13, n_fft=400, hop_length=160, n_mels=23)
• Label encoding — binary classification (0 = cat, 1 = dog) parsed from filename
• Integrated with DataLoader for batched training (batch_size=32)

4. CNN Architecture

A 4-layer Convolutional Neural Network built from scratch:

Input (1 × 13 × T MFCC)
  → Conv2d(1→16, 3×3) + ReLU + MaxPool2d
  → Conv2d(16→32, 3×3) + ReLU + MaxPool2d
  → Conv2d(32→64, 3×3) + ReLU + MaxPool2d
  → Conv2d(64→128, 3×3) + ReLU + MaxPool2d
  → Flatten
  → Linear → 2 classes
  → Softmax → Prediction

Training configuration:

Parameter	Value	Rationale
Optimizer	Adam	Adaptive learning rates per-parameter; robust default
Learning rate	0.001	Adam's recommended default; proven stable starting point
Loss function	CrossEntropyLoss	Standard for multi-class classification
Epochs	10	Small dataset overfits quickly; sufficient for convergence
Validation	5-Fold Cross-Validation	Robust evaluation using all samples; reliable with only 277 data points

5. Classical ML Baseline

A Logistic Regression model using scikit-learn:

• 20 MFCCs extracted per file via librosa.feature.mfcc
• Additional features: spectral centroid, spectral rolloff
• Standard train/test split evaluation
• Demonstrates methodological rigour — always compare deep learning against simpler approaches

---

Design Decisions

Why MFCC over raw audio?

Raw audio is 160,000 numbers per sample — high-dimensional and semantically sparse. MFCCs compress this into a 13 × T matrix that captures perceptually meaningful frequency content, weighted to match human hearing (mel scale). This gives the CNN a compact, information-dense input.

Why ReLU activation?

• No vanishing gradient — sigmoid/tanh squash gradients toward zero in deep networks, causing earlier layers to stop learning. ReLU's gradient is 0 or 1.
• Computational efficiency — simple threshold comparison vs. exponential functions
• Sparsity — neurons with negative input output zero, creating efficient sparse representations

Why MaxPool2d?

• Translation invariance — detects patterns regardless of their exact temporal position (a meow at t=1s and t=3s are the same pattern)
• Progressive abstraction — each layer captures increasingly high-level features (from "energy spike at 500Hz" to "repeated harmonic pattern typical of barking")
• Max over Average — preserves the strongest feature activation in a region, avoiding dilution of distinctive signals

Why double channels at each layer? (16 → 32 → 64 → 128)

MaxPool halves spatial dimensions at each stage, reducing available positional information. Doubling channels compensates by providing more "vocabulary" to represent increasingly abstract pattern combinations. This follows the balanced-computation principle established by VGGNet — each layer performs roughly equal work.

Why 160,000 samples as fixed length?

Neural networks require fixed-size inputs. The average signal length in the dataset (determined via EDA) was ~160,000 samples (~7.3 seconds at 22,050 Hz). This minimises both information loss from truncation and wasted computation from excessive zero-padding.

Why 5-Fold Cross-Validation over a single train/test split?

With only 277 samples, a single 80/20 split gives a test set of ~55 samples — too small for reliable accuracy estimation. K-Fold ensures every sample is used for testing exactly once, providing a more statistically robust performance metric and revealing model stability across different data partitions.

Why compare CNN against Logistic Regression?

The simplest solution should always be evaluated first. If Logistic Regression achieves 95% accuracy, a complex CNN isn't justified. This comparison demonstrates awareness that model complexity must be earned — a principle critical in production ML where simpler models are cheaper to deploy and maintain.

---

Skills Demonstrated

Competency	Evidence
Deep Learning (PyTorch)	Custom CNN architecture, training loops, loss functions, optimizers
Audio/Signal Processing	MFCC, mel spectrograms, STFT, spectral features, resampling
Feature Engineering	Domain-specific extraction from unstructured audio data
Data Pipeline Design	Custom PyTorch Dataset + DataLoader with on-the-fly transforms
Model Evaluation	K-Fold cross-validation for robust performance estimation
Classical ML	Logistic Regression baseline for methodological comparison
EDA	Statistical analysis of signal properties, distribution visualisation
Unstructured Data	Working with raw .wav audio — not tabular or image data
End-to-End ML	Data loading → EDA → preprocessing → feature extraction → training → inference