A New Framework for CNN-Based Speech Enhancement in the Time Domain

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Speech enhancement is crucial for improving the intelligibility of noisy audio signals.
• Traditional methods often struggle with complex noise types and preserving speech quality.
• Fully convolutional neural networks (CNNs) show promise but require effective training strategies.

Purpose of the Study:

• To propose a novel learning mechanism for time-domain speech enhancement using CNNs.
• To leverage frequency-domain analysis within a time-domain CNN training framework.
• To address limitations of existing speech enhancement techniques.

Main Methods:

• A fully convolutional neural network (CNN) is employed for time-domain speech enhancement.
• A differentiable operation converts time-domain signals to the frequency domain during training.
• Mean absolute error loss is applied to the Short-Time Fourier Transform (STFT) magnitude for training.

Main Results:

• The proposed method significantly outperforms existing speech enhancement techniques.
• The CNN operates in the time domain, avoiding the invalid STFT problem.
• The approach effectively utilizes frequency-domain knowledge for enhanced speech quality.

Conclusions:

• The novel learning mechanism enables effective time-domain speech enhancement with CNNs.
• This method offers a robust and implementable solution for speech processing tasks.
• The approach demonstrates superior performance and applicability to related signal processing challenges.