Phase-aware deep speech enhancement: It's all about the frame length.

Area of Science:

• Speech Processing
• Deep Learning
• Signal Analysis

Background:

• Algorithmic latency in speech processing is primarily determined by the frame length used in Fourier analysis.
• Magnitude-centric speech enhancement approaches face performance limitations due to fixed frame lengths.
• Prior research indicates that phase information becomes increasingly critical as frame length decreases.

Purpose of the Study:

• To systematically investigate the roles of phase and magnitude in deep neural network (DNN)-based speech enhancement.
• To evaluate performance across various frame lengths, focusing on the impact of reduced frame sizes.

Main Methods:

• Utilized modern deep neural networks (DNNs) for speech enhancement.
• Conducted experiments across a range of frame lengths to analyze phase and magnitude contributions.
• Compared performance of DNNs employing short versus long frames.

Main Results:

• DNNs demonstrated the capability to accurately estimate phase information even with short frames.
• Speech enhancement performance using short frames was comparable or superior to that achieved with longer frames.
• Phase-aware DNNs achieved high-quality results at reduced latency.

Conclusions:

• Short frame lengths are viable for high-quality speech enhancement when using advanced DNNs.
• DNNs effectively leverage phase information in short-frame scenarios, enabling low-latency processing.
• This research enables practical, high-performance, low-latency speech enhancement solutions.