Requirements for the evaluation of computational speech segregation systems

Area of Science:

• Speech processing
• Computational acoustics
• Machine learning for audio

Background:

• Supervised learning algorithms enhance speech intelligibility in noise via ideal binary masks.
• Robustness to diverse acoustic conditions is crucial for real-world technical applications.

Purpose of the Study:

• To investigate the impact of spectro-temporal noise variations on computational speech segregation performance.
• To understand how these variations affect the identification of acoustical features linked to speech segregation.

Main Methods:

• Utilizing supervised learning algorithms for estimating ideal binary masks.
• Analyzing the influence of spectro-temporal noise variations during both training and testing phases.

Main Results:

• Spectro-temporal noise variations critically determine the achievable speech segregation performance.
• These variations significantly influence the system's ability to identify acoustical features relevant to speech perception.

Conclusions:

• The variability of noise during training and testing is a key factor in speech segregation system performance.
• Establishing a framework for systematic evaluation of future segregation systems based on noise variation is proposed.