How hamzaih2k-1108/automatic-voice-pathology-detection-and-classification-systems works
Project type and purpose
• This is a data-driven signal-processing / machine-learning pipeline implemented as a Jupyter notebook.
• Purpose: ingest raw audio (.wav) files, preprocess (pre-emphasis, Fourier transform), train a CNN to detect voice pathologies, and evaluate its performance.
File structure
• Top-level:
• README.md (project description)
• Signal_Processing_Project.ipynb (main implementation)
• data_set/
– healthy/ and pathological/ subfolders full of .wav files
• No separate source, tests or deployment folders—everything lives in the notebook.
• Implies a simple linear pipeline rather than microservices or MVC.
Key insights from the README
• Four processing stages:
1. Pre-emphasis filter
2. Fourier transform (feature extraction)
3. Build/train CNN model
4. Evaluate model
• No external APIs or web UI—core is a batch ML workflow.
Main components and their relationships
a. Data Storage
• Raw .wav files in data_set/healthy and data_set/pathological
b. Preprocessing Pipeline
• Pre-emphasis filter
• Fourier transform (e.g., STFT or FFT)
• Use librosa or scipy.signal
c. Feature Dataset
• Spectrograms, mel-spectrograms or feature matrices
• Stored in memory or on disk as NumPy arrays
d. Model Trainer
• Convolutional Neural Network built with TensorFlow/Keras or PyTorch
• Consumes feature dataset, outputs a trained model artifact
e. Evaluation Module
• Computes metrics (accuracy, confusion matrix, ROC curves)
• Outputs plots and reports in notebook
Data Storage → Preprocessing → Feature Dataset → Model Trainer → Evaluation → Reports
Architectural patterns and design principles
• Pipeline pattern: each stage is a discrete transformation
• Separation of concerns: data ingestion, feature extraction, model training, evaluation
• Reproducibility: version your notebook, data, and model artifacts
Relevant technologies and libraries
• Python 3.x
• Jupyter Notebook (Signal_Processing_Project.ipynb)
• NumPy, SciPy, librosa for audio processing
• TensorFlow/Keras or PyTorch for CNN
• Matplotlib or Seaborn for visualization
Detailed breakdown to capture all elements
1. Data Storage
– Path: data_set/healthy & data_set/pathological
2. Preprocessing
– Pre-emphasis filter (High-pass IIR filter)
– Fourier transform (FFT/STFT)
3. Feature Representation
– Spectrogram / Mel-spectrogram / MFCC arrays
4. Model Architecture
– CNN layers (Conv2D, pooling, dense)
– Loss function, optimizer
5. Model Artifact
– .h5 or .pt file
6. Evaluation
– Metrics (accuracy, precision, recall)
– Plots (confusion matrix, ROC)