A Convolutional Recurrent Neural Network for Real-Time Speech Enhancement

A minimum unofficial implementation of the A Convolutional Recurrent Neural Network for Real-Time Speech Enhancement (CRN) using PyTorch.

ToDo

• Real-time version
• Update trainer
• Visualization of the spectrogram and the metrics (PESQ, STOI, SI-SDR) in the training
• More docs

Usage

Training:

python train.py -C config/train/baseline_model.json5

Inference:

python inference.py \
    -C config/inference/basic.json5 \
    -cp ~/Experiments/CRN/baseline_model/checkpoints/latest_model.tar \
    -dist ./enhanced

Check out the README of Wave-U-Net for SE to learn more.

Performance

PESQ, STOI, SI-SDR on DEMAND - Voice Bank test dataset, for reference only:

Dependencies

• Python==3.*.*
• torch==1.*
• librosa==0.7.0
• tensorboard
• pesq
• pystoi
• matplotlib
• tqdm

References

• CRNN_mapping_baseline
• A Convolutional Recurrent Neural Network for Real-Time Speech Enhancement
• EHNet
• Convolutional-Recurrent Neural Networks for Speech Enhancement