ChirpApp

A machine learning-powered audio analysis system designed to detect chirp signals in acoustic environments. ChirpApp combines mobile data collection with advanced signal processing and ML models to classify audio as containing chirp signals or not.

---

🎯 Project Overview

ChirpApp addresses the problem of detecting ultrasonic chirp signals in noisy environments. The project uses a data-driven approach to:

1. Collect audio samples via a mobile app
2. Process signals with bandpass filtering and feature extraction
3. Train multiple ML models to classify chirp presence
4. Evaluate performance across different chirp period durations (1s, 2s, 4s)
5. Select optimal parameters based on model prediction accuracy

The system is designed to be robust, generalizable, and suitable for deployment in real-world acoustic monitoring applications.

---

📁 Project Structure

mobileApp/ — Android Data Collection

Contains the Android mobile application for recording and collecting audio samples.

Key Components:

• src/main/ — Source code for the Android app
  • MainActivity.kt — Main application activity and UI
  • ui/theme/ — Custom Material Design theme configuration
• AndroidManifest.xml — App permissions and configurations
• build.gradle.kts — Gradle build configuration
• res/ — Resource files (layouts, strings, drawables, icons)

Purpose: Records audio clips in natural environments and labels them as positive (contains chirp) or negative (no chirp) for training data collection.

---

analysis/ — Signal Processing & Machine Learning

Core data analysis and model training pipeline.

Key Files:

audio-filter.py (Main Pipeline)

The comprehensive analysis script that:

1. Loads & Cleans Audio

   • Bandpass filters signals (17.5 kHz - 20.5 kHz)
   • Converts .wav and .pcm files
   • Normalizes amplitude

2. Segments Signals

   • Divides long audio into fixed-duration chunks based on chirp period T
   • Tests three periods: T ∈ {1s, 2s, 4s}
   • Each segment becomes one training/test sample

3. Extracts Features (4 modalities)

   • Statistical Features: Mean, std, max, min, power, zero-crossing rate
   • Wavelet Features: Discrete wavelet coefficients (db4, level 4)
   • MFCC (Mel-Frequency Cepstral Coefficients): 40 coefficients, 64 frames
   • Spectrograms: Mel-spectrogram and STFT (64×64 normalized images)

4. Trains Models

   • Classical ML (on statistical/wavelet features):
     • Support Vector Machine (SVM)
     • Random Forest (100 estimators)
     • k-Nearest Neighbors (k=5)
   • Deep Learning (on spectral features):
     • Simple 2D CNN with Conv2D → MaxPool → Dense layers
     • Trained separately on mel-spec, MFCC, wavelet scalogram, STFT

5. Evaluates Performance

   • Cross-validation: 5-fold stratified K-fold (or fewer with small datasets)
   • Train/Test Split: 80% train, 20% test (when sample size permits)
   • Metrics Recorded:
     • Accuracy, Precision, Recall, F1-score
     • Mean ± std for all metrics

6. Saves Results

   • Outputs model_results.csv with comprehensive results:
     • T value, model type, feature modality, evaluation type
     • Full performance metrics for every model×modality×T combination

Data Directories

• raw_data/ — Positive samples (contain chirp signals)
• negative_data/ — Negative samples (no chirp signals)
• analyzed_data/ — Output folder for processed results

Requirements

numpy, pandas, librosa, PyWavelets, opencv-python, scipy, scikit-learn, tensorflow

---

🚀 Quick Start

Prerequisites

• Python 3.10+
• Virtual environment (recommended)

Installation

# Clone the repository
git clone https://github.com/DaveTron4/ChirpApp.git
cd ChirpApp/analysis

# Create and activate virtual environment
python -m venv .venv
.\.venv\Scripts\activate  # Windows
source .venv/bin/activate  # macOS/Linux

# Install dependencies
pip install numpy pandas librosa PyWavelets opencv-python scipy scikit-learn tensorflow

Running the Analysis

python audio-filter.py

Output:

• Console logs showing progress for each T value
• model_results.csv — Full results table with all metrics

---

📊 Results & Interpretation

The pipeline tests three chirp periods and evaluates each model across all feature modalities.

Key Metrics per Model:

• Accuracy: Overall correctness
• Precision: True positive rate among predicted positives
• Recall: Detection rate of actual positives
• F1-Score: Harmonic mean of precision & recall

Expected Performance:

• T=1s: Best results (23 samples) — Models typically achieve 95-100% accuracy
• T=2s: Good results (11 samples) — Models typically achieve 90-100% accuracy
• T=4s: Limited by data (5 samples) — Less reliable; ~58% accuracy

Choosing the Optimal T:

Based on results in model_results.csv, T=1s is recommended for production because:

• Most training samples
• Highest and most stable accuracy
• Best generalization to test data
• Consistent performance across all model types

---

🔧 Configuration

Edit audio-filter.py constants to customize the pipeline:

SAMPLE_RATE = 44100        # Audio sample rate (Hz)
LOWCUT = 17500.0           # Bandpass filter lower cutoff (Hz)
HIGHCUT = 20500.0          # Bandpass filter upper cutoff (Hz)
FILTER_ORDER = 6           # Butterworth filter order
T_VALUES = [1, 2, 4]       # Chirp periods to test (seconds)
LOAD_POSITIVE = True       # Include positive samples
LOAD_NEGATIVE = True       # Include negative samples

---

📈 Model Details

Classical ML Pipeline

Raw Features → StandardScaler → SVM/RF/kNN → Predictions

Uses 5-fold stratified cross-validation for robust evaluation.

CNN Architecture

Input (64×64×1 image) 
  → Conv2D(16, 3×3) + ReLU
  → MaxPooling2D(2×2)
  → Conv2D(32, 3×3) + ReLU
  → MaxPooling2D(2×2)
  → Flatten
  → Dense(64) + ReLU + Dropout(0.4)
  → Dense(1) + Sigmoid

Simple, lightweight design to avoid overfitting on small datasets.

---

📝 Output Format

model_results.csv contains:

Column	Description
T	Chirp period (1, 2, or 4 seconds)
model	Model type (SVM, RF, kNN, CNN)
modality	Feature type (stat, wavelet, mel_spec, mfcc, wavelet_scalo, stft)
eval_type	Evaluation method (cross_validation or 80_20_split)
accuracy_mean	Mean accuracy score
accuracy_std	Standard deviation of accuracy
precision_mean	Mean precision
recall_mean	Mean recall
f1_mean	Mean F1-score

---

⚙️ Troubleshooting

"No samples found" error

• Check that raw_data/ and negative_data/ folders contain .wav or .pcm files
• Ensure files are not corrupted

NaN values in results

• Occurs when dataset is too small (T=4s with <5 samples)
• kNN in particular struggles with minimal data
• Recommendation: Collect more audio samples

TensorFlow warnings

• "oneDNN custom operations" and "function retracing" warnings are informational
• They don't affect results; suppress with TF_ENABLE_ONEDNN_OPTS=0

---

🎓 Future Improvements

• Add data augmentation techniques (time-stretch, pitch-shift)
• Implement ensemble methods combining multiple modalities
• Deploy optimal model to mobile app
• Support real-time inference on device
• Add confusion matrices and ROC curves to analysis

---

📄 License

This project is part of Georgia State University coursework.

---

👤 Author

David (DaveTron4)

Last Updated: December 9, 2025