This project aims to develop a Noise Cancelation System to be used to remove background noise from the input audio by using Convolutional Neural Network's in Pytorch. This model transforms .wav files into mel spectrograms feeds it into the CNN produces a clean spectrogram and converts it back into waveform and stores it into a .wav file.

• Description
• Table of Contents
• Technologies Used
• Requirements
  • Required
  • Optional
• Usage Instructions
  • Training
    • Example
  • Loading
• Project Roadmap

• Python
• Pytorch
• Torchaudio (set to change)

• Python 3.13+
• Python packages:
  • librosa==0.10.2.post1
  • matplotlib==3.10.0
  • numpy==2.2.3
  • setuptools==75.8.0
  • torch==2.6.0+cu126
  • torchaudio==2.6.0+cu126

• Dataset: Kaggle Dataset

To train a model with the architecture in the src.models.model Module:

• Ensure that the directory of data is created where the clean, noisy data for training and testing will be the directories are specified in the __init__.py file.

ROOT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))

MODEL_DIR = os.path.join(ROOT_DIR, "models")
MODEL_PATH = os.path.join(MODEL_DIR, "model.pt")

MODEL_OUTPUT_DIR = os.path.join(MODEL_DIR, "output")
MODEL_OUPUT_PATH = os.path.join(MODEL_OUTPUT_DIR, "output.wav")

DATA_DIR = os.path.join(ROOT_DIR, "data")

NOISY_TRAIN_DIR = os.path.join(DATA_DIR, "noisy_trainset_wav")
CLEAN_TRAIN_DIR = os.path.join(DATA_DIR, "clean_trainset_wav")

NOISY_TEST_DIR = os.path.join(DATA_DIR, "noisy_testset_wav")
CLEAN_TEST_DIR = os.path.join(DATA_DIR, "clean_testset_wav")

• Load the data using the data loader in the utils.data_loader module
• Instantiate new instance of the model
• Call the train function and enter the relevant parameters

i.e:

model = CNN().to(device)

spectrogram = MelSpectrogram(n_fft=1024)

data_loader = load_data(CLEAN_TRAIN_DIR, NOISY_TRAIN_DIR, spectrogram)

updated_state = train(
    model=model,
    epochs=120,
    optimiser=Adam(model.parameters()),
    loss_func=torch.nn.MSELoss(),
    data_loader=data_loader
)

• Save the updated state to a model.pt file

load_model = src.utils.load_model()
model = src.models.model.CNN().to(device)
model.load_state_dict(load_model)
model.eval()