msalhab96 / SNR-Estimation-Using-Deep-Learning

An implementation for Frame-level Speech Signal-to-Noise Ratio Estimation using deep learning

SNR-Estimation-Using-Deep-Learning

PyTorch implementation of Frame-level Signal-to-Noise Ratio Estimation using Deep Learning.

This implementation includes distributed training and trained on LibriSpeech -train-clean-100.tar.gz- dataset and the noise collected from different sources.

The below image is taken from the training on LibriSpeech

The below images are samples shows the results

Setup

1. Download and extract LibriSpeech
2. Clone this repo: git clone https://github.com/msalhab96/SNR-Estimation-Using-Deep-Learning
3. CD into this repo: cd SNR-Estimation-Using-Deep-Learning
4. Install the requirements: pip install -r requiremnts.txt

Training

To train the model follow the steps below:

1. Preprocess all the audio files and make sure all of them are single channeled audios
2. Change the configuration in the config/configs.yaml file
3. Run python train.py to train from scratch or python train.py checkpoint=path/to/checkpoint to train the model from a checkpoint
4. Run tensorboard --logdir=logdir/ to monitor the training (optional)

Pre-trained model

You can download the pretrained model from here