Deep Learning-Based Estimation of Reverberant Environment for Audio Data Augmentation.

Area of Science:

• Audio Signal Processing
• Machine Learning
• Deep Learning

Background:

• Traditional audio data augmentation for dereverberation relies on artificial room impulse responses.
• Existing methods like the image method for generating impulse responses have limitations.

Purpose of the Study:

• To propose a novel deep learning-based audio data augmentation method.
• To improve the performance of audio dereverberation systems.

Main Methods:

• Estimating a reverberation environment model using a deep neural network (DNN).
• Training the DNN with clean and recorded audio data.
• Generating a large real augmented database using the trained reverberation model.
• Training the dereverberation model with the augmented database.

Main Results:

• The proposed augmentation model's performance was validated using log spectral distance and mean square error.
• Dereverberation experiments demonstrated superior performance compared to conventional methods.
• The deep learning approach effectively enhances audio dereverberation.

Conclusions:

• The proposed deep learning-based audio data augmentation method significantly improves dereverberation performance.
• This approach offers a more effective way to create realistic augmented audio data for training dereverberation models.