Reading vs Natural Speech Classifier

This project is a small end-to-end audio classification system that distinguishes reading-style speech from natural conversational speech using short audio clips.

⚡ Implemented from scratch: This model does not use TensorFlow or PyTorch. All neural network operations—including forward/backward passes and weight updates—are coded manually to demonstrate a deep understanding of audio features and learning dynamics.

The goal is not to build a production-grade model, but to demonstrate:

• Audio feature extraction (MFCCs)
• Basic neural network implementation from scratch
• Model training and evaluation
• A simple web interface with Flask
• Clean separation between data, features, model, and interface

This project was developed as part of a technical portfolio.

---

Overview

The pipeline is:

1. Load short mono WAV files (16kHz, 16-bit)
2. Extract MFCC features (13 coefficients per clip)
3. Normalize features using training statistics
4. Train a small neural network (1 hidden layer, ReLU + Sigmoid)
5. Evaluate classification accuracy
6. Save model weights and feature scaler
7. Serve the model via a Flask web interface for online predictions

---

Dataset

The dataset consists of 23 short audio clips:

• 12 reading-style recordings
• 11 natural speech recordings

Note: Due to the small dataset size, overfitting is expected.
Accuracy may vary across runs, but the focus is on demonstrating a complete ML workflow.

---

Project Structure

reading-vs-natural-online/
├── app.py # Flask application
├── requirements.txt # Python dependencies
├── README.md
├── src/
│ ├── init.py
│ ├── features.py # MFCC extraction
│ ├── model.py # Tiny neural network implementation
│ ├── train.py # Training script
│ └── predict.py # Inference helper
├── data/
│ ├── reading/ # Reading-style audio clips
│ └── natural/ # Natural speech clips
├── uploads/ # Uploaded audio files via Flask
└── models/
  ├── model.npz # Saved NN weights
  └── scaler.npz # Saved StandardScaler statistics

---

Setup

1. Clone the repository:

   git clone https://github.com/yourusername/reading-vs-natural-online.git
   cd reading-vs-natural-online

2. Create a virtual environment and activate it:

   python3 -m venv .venv
   source .venv/bin/activate  # Mac/Linux
   .venv\Scripts\activate     # Windows

3. Install dependencies:

   python3 -m pip install -r requirements.txt

Training the Model

1. Place your audio clips in data/reading and data/natural.

2. Run the training script:

   python3 -m src.train

3. This will:

   • Extract MFCC features
   • Train the TinyNN model
   • Save model weights (models/model.npz)
   • Save scaler statistics (models/scaler.npz)

Using the Web Interface

1. Start the Flask app:

   python3 app.py

2. Open your browser at http://127.0.0.1:5000
3. Upload an audio file (mono WAV, 16kHz recommended)
4. Click Upload and Predict to see the classification
5. The uploaded file name and prediction will remain displayed until a new file is uploaded.

   Tip: Only short clips were used during training. Very long files may yield unreliable predictions.

Notes on Audio Files

• Sampling rate: 16 kHz
• Bit depth: 16-bit PCM
• Mono audio

For best results, match these settings when recording or preprocessing your audio.
I used these settings to keep this prototype at a manageable size.

Development

• src/features.py: MFCC extraction logic
• src/model.py: Tiny neural network with forward, backward, training, and predict functions
• src/train.py: Training pipeline including model/scaler saving
• src/predict.py: Loading saved model and scaler, making predictions
• app.py: Flask app exposing a web interface