Towards Robust Speech Super-resolution.

Area of Science:

• Signal Processing
• Machine Learning
• Audio Engineering

Background:

• Speech super-resolution (SR) aims to reconstruct high-frequency components of degraded speech signals.
• Existing deep neural network (DNN) models for SR face challenges with robustness to variations in microphone channels and downsampling methods.

Purpose of the Study:

• To propose a novel convolutional neural network (CNN) based speech super-resolution model.
• To leverage both time and frequency domain information for improved SR performance.
• To investigate and enhance the robustness of SR models against diverse real-world conditions.

Main Methods:

• A time-domain CNN is developed, accepting raw low-resolution speech waveforms as input.
• A cross-domain loss function is employed during network training for optimization.
• The proposed CNN model is compared against several established DNN-based SR approaches.

Main Results:

• The proposed CNN-based SR model demonstrates superior performance compared to existing DNN models.
• The study investigates the robustness of DNN-based SR models concerning microphone channels and downsampling schemes.
• Improved generalization capabilities are achieved for untrained microphone channels and unknown downsampling schemes through proper training and preprocessing.

Conclusions:

• The developed CNN model offers an effective solution for speech super-resolution, outperforming current DNN methods.
• Addressing robustness issues is crucial for practical deployment of SR systems.
• Strategic training and preprocessing enhance the adaptability of SR models to varied audio acquisition conditions.