Sprint 2: Augmentation Benchmarking

README_SPRINT2_AUGMENTATION.md

@@ -0,0 +1,58 @@
+# Sprint 2: Augmentation Benchmarking
+
+## Overview
+
+This task benchmarks selected audio augmentation techniques identified in Sprint 1. The aim is to evaluate how each method affects CNN model performance.
+
+## Methods Tested
+
+- Baseline (no augmentation)
+- Time Stretching
+- Pitch Shifting
+- Gaussian Noise Injection
+- SpecAugment
+- Combined Noise + SpecAugment
+
+## Approach
+
+A controlled experiment was completed using:
+
+- A fixed subset of the dataset with 5 categories
+- Mel spectrogram feature extraction
+- A CNN model with the same architecture for all methods
+- The same training settings for each benchmark run
+
+The dataset was split into training, validation, and test sets. Augmentation was applied only to the training data, while validation and test data remained unchanged.
+
+## Evaluation Metrics
+
+- Accuracy
+- Weighted F1 Score
+
+## Results Summary
+
+| Method | Accuracy | F1 Score | Decision |
+|---|---:|---:|---|
+| Noise Injection | 0.393 | 0.328 | Keep |
+| Baseline | 0.286 | 0.167 | Reference only |
+| Pitch Shift | 0.286 | 0.221 | Defer |
+| Noise + SpecAugment | 0.214 | 0.129 | Defer |
+| SpecAugment | 0.179 | 0.077 | Drop |
+| Time Stretch | 0.143 | 0.145 | Drop |
+
+## Key Findings
+
+- Noise injection showed the best performance in this benchmark.
+- Pitch shifting produced the same accuracy as the baseline, with a slight improvement in F1 score.
+- SpecAugment, time stretching, and the combined Noise + SpecAugment method performed below the baseline.
+
+## Final Recommendation
+
+- Keep: Noise Injection
+- Defer: Pitch Shift, Noise + SpecAugment
+- Drop: SpecAugment, Time Stretch
+- Reference only: Baseline
+
+## Files Included
+
+- `Sprint2_Augmentation_KD.ipynb` – main benchmarking notebook