Water Bottle Challenge

Purpose:

Build an unsupervised classifier to distinguish between knife-strikes on the top vs bottom of a stainless steel water bottle, using only two labeled examples and 24 unlabeled spectrogram CSVs.

Results:

95% accuracy on labeled data!

---

✅ Project Overview

Most code in this challenge was developed with help of ChatGPT. I gave it a set of requirements for my code and had it create the scaffolding. Given 3 hour limit, I wanted to honor it as best I could, and using ChatGPT helped me iterate quickly. My exploration.ipynb took about 2 hours of work. The actual water_bottle_challenge.py, the tests, and readme were created in just over 1 hour - this required more manual coding versus ChatGPT assist.

1. Data

   • Preprocessed spectrogram CSVs in data/:
     • top.csv, bottom.csv (one clear example each).
     • unlabeled_XX.csv files (unknown strike location).
   • Each CSV: rows = frequencies (Hz), columns = timepoints (ms), values = magnitude.

2. Goal

   • Implement classify_preprocessed_audio(fpath: str) -> int:
     • 0 = top strike
     • 1 = bottom strike
     • None = unsure or invalid
   • Must not error on any valid CSV.

3. Constraints

   • Only two labeled samples.
   • No external labels for evaluation.
   • Code readability, concise explanations, and reproducibility matter.

---

🛠 Steps & Rationale

1. Initial Feature Engineering

These features were created with help from ChatGPT, I used it to research typical features in audio analysis of spectograms

• Spectral Centroid: average frequency weighted by energy → indicates brightness.
• Spectral Bandwidth: RMS spread around centroid → harmonic richness.
• Rolloff 85%: cutoff where 85% of energy resides → high-frequency content.
• Peak Frequency: single frequency with maximum total energy.

Rationale: Top strikes produce sharper, tinny high-freq bursts; bottom strikes ring with lower, sustained tones as tested on a waterbottle in my room!

2. Visual Inspection

• Plotted log-scaled spectrograms with librosa.display.specshow.
• Noted distinct color-dominant zones for top vs bottom.
• Tried to just seperate by where the longest and most colorful streak was

3. Cosine angle similarity Prototyping

• Take cosine sim which is inverse cosine of (dot product of vectors divided by their (magnitudes multiplied))
• Only looks at angle between vectors, which loses magnitude information
• With this data I believe it is not appropriate discard magnitude of the vector when classifying, magnitude should be important in classifying top and bottom
• Though I did utilize these results to validate if K-means was at least on track, and found about 70% agreement between this method and K-means by hand
• Visually looking at cosine results after labeling I felt there were different labels that were wrong after PCA was applied.

4. K-means

• Extract features for all files.
• Standardize via StandardScaler.
• KMeans with 3 clusters (top, bottom, unsure).
• Map clusters to labels via top.csv/bottom.csv.
• Predict new files’ cluster and translate to 0/1/None.

Why 3 clusters? Some files are ambiguous—allow an “unsure” group. K-means always makes a best guess, so we don't want it guessing top or bottom when really it should be "unsure".

5. Feature Expansion & Evaluation

• Added extras: spectral flatness, time-to-peak, decay rate, low/high energy ratio, spectral entropy.
• Automated silhouette score evaluation.
• Visualized PCA+KMeans to inspect separation.
• These features were derived with help from ChatGPT, I wanted to quickly evaluate if there were other features I could test and see if it helped our groupings

6. Unit Testing & Robustness

• Created pytest tests verifying top.csv → 0, bottom.csv → 1, and unlabeled handling.

---

▶️ How to Run

1. Install:

   pip install -r requirements.txt
   

2. Test:

   pytest -v -s test_water_bottle_challenge.py
   

3. Run:

   from water_bottle_challenge import classify_preprocessed_audio
   print(classify_preprocessed_audio('data/unlabeled_00.csv'))