Speech Enhancement by Multiple Propagation through the Same Neural Network

Area of Science:

• * Digital Signal Processing
• * Machine Learning
• * Audio Engineering

Background:

• * Monaural speech enhancement aims to improve speech clarity by removing background noise.
• * Deep neural networks are the current state-of-the-art for speech enhancement.
• * Previous research showed multi-forward-pass enhancement with U-Net improved results.

Purpose of the Study:

• * To investigate the effectiveness of multi-forward-pass speech enhancement on different neural network architectures.
• * To determine the performance limits of multi-forward-pass enhancement by testing up to five iterations.
• * To compare the benefits of multi-forward-pass enhancement across U-Net, ResBLSTM, and Transformer-Net models.

Main Methods:

• * Implemented multi-forward-pass speech enhancement for ResBLSTM and Transformer-Net architectures.
• * Tested enhancement iterations from two up to five.
• * Evaluated performance using SI-SDR, STOI, and PESQ metrics on WSJ0, Noisex-92, and DCASE datasets.

Main Results:

• * Multi-forward-pass speech enhancement up to five iterations consistently improved speech intelligibility.
• * Performance gains diminished with each additional iteration.
• * Five iterations yielded an additional 0.6 dB SI-SDR and 4% STOI gain compared to two iterations.
• * U-Net and Transformer-Net architectures benefited more from multi-forward passes than ResBLSTM.

Conclusions:

• * Multi-forward-pass speech enhancement is effective across various architectures, extending beyond U-Net.
• * While gains decrease with more iterations, five passes offer significant improvements over fewer passes.
• * Architecture choice impacts the effectiveness of multi-forward-pass enhancement, with U-Net and Transformer-Net showing greater advantages.