Digital Speech Processing

The paper investigates reversible neural network layers for constructing a duplex speech chain model that utilizes bidirectional supervision signals. It introduces a novel approach to address speech synthesis and recognition challenges while analyzing the impact of bidirectional supervision on performance.

Randomly weighted neural networks have potential in computer vision and demonstrate a positive correlation in performance with fully trained models. This paper confirms the continued positive correlation and explores the value of randomly weighted networks in audio source separation, a challenging generative audio task.