A Lightweight and Real-Time Binaural Speech Enhancement Model with Spatial Cues Preservation

This is the official implementation of the LBCCN.

⚠️Note: The model is trained in a diffuse noise field with a fixed speaker direction (e.g., 55 degrees to the right of the frontal direction). Each training instance of the model targets only one direction.
The noise data used for training is from the DEMAND dataset.

🔧 Errata

We have identified and corrected the following inaccuracies in the original paper:

1. Speaker Direction:
   The actual speaker direction used in the binaural speech enhancement experiments is 55 degrees, not 45 degrees as previously mentioned.

2. Noise Dataset:
   The noise data used for training is from the DEMAND dataset, not NoiseX92 as may have been previously implied.

We apologize for any confusion this may have caused and appreciate your understanding.