Audio Noise Removal Using STFT and CNN

This project implements a deep learning-based approach for removing noise from audio files. It uses a Convolutional Neural Network (CNN) model to process the spectrogram of noisy audio and generate a cleaned version.

Table of Contents

• Features
• Requirements
• Installation
• Usage
• Model Architecture
• Training
• Evaluation
• Results
• Contributing
• License

Features

• Remove background noise from audio files
• Process multiple audio files in batch
• Generate spectrograms and waveforms for input and output audio
• Support for various audio formats (WAV, MP3, etc.)

Requirements

• Python 3.7+
• TensorFlow 2.x
• NumPy
• SciPy
• Matplotlib
• Librosa
• tqdm

Installation

1. Clone this repository:

   git clone https://github.com/yourusername/audio-noise-removal.git
   cd audio-noise-removal
   

2. Install the required packages:

   pip install -r requirements.txt
   

3. Download the pre-trained model:

   # Add instructions for downloading the model file
   

Usage

Single File Processing

To process a single audio file:

from modules import process_audio

input_file = "path/to/your/noisy_audio.wav"
output_dir = "path/to/output/directory"

process_audio(input_file, output_dir)

Batch Processing

To process multiple audio files:

import os
from modules import process_audio

input_dir = "path/to/noisy/audio/files"
output_dir = "path/to/output/directory"

for file_name in os.listdir(input_dir):
    if file_name.endswith(".wav"):
        input_file = os.path.join(input_dir, file_name)
        process_audio(input_file, output_dir)

Model Architecture

The noise removal model is a Convolutional Neural Network (CNN) with the following architecture:

• Input layer: 128x128x1 (spectrogram)
• Multiple convolutional and pooling layers
• Skip connections for better feature preservation
• Output layer: 128x128x1 (cleaned spectrogram)

For more details, refer to the CNNmodel() function in modules.py.

Training

The model is trained on pairs of noisy and clean audio samples. To train the model:

1. Prepare your dataset of noisy and clean audio pairs.
2. Update the data loading and preprocessing steps in modules.py.
3. Run the training script:

   python train.py
   

Training parameters can be adjusted in the train.py file.

Evaluation

To evaluate the model's performance:

1. Prepare a test set of noisy audio files.
2. Run the evaluation script:

   python evaluate.py --test_dir path/to/test/files --output_dir path/to/output
   

This will process the test files and generate cleaned versions along with spectrograms and waveform visualizations.

Results

The model's performance can be assessed by:

1. Listening to the original noisy audio and the processed clean audio.
2. Comparing the spectrograms of the input and output.
3. Calculating objective metrics such as Signal-to-Noise Ratio (SNR) improvement.

Example results and visualizations can be found in the results directory.

Contributing

Contributions to this project are welcome! Please follow these steps:

1. Fork the repository
2. Create a new branch for your feature
3. Commit your changes
4. Push to your fork
5. Submit a pull request

License

This project is licensed under the MIT License - see the LICENSE file for details.