Subjective comparison and evaluation of speech enhancement algorithms

Area of Science:

• Signal Processing
• Acoustics
• Speech Technology

Background:

• Meaningful comparison of speech enhancement algorithms is difficult due to varied databases, noise types, and testing methods.
• A standardized approach is needed to evaluate the effectiveness of different speech enhancement techniques.
• Previous research lacked a common benchmark for performance assessment.

Purpose of the Study:

• To develop a comprehensive noisy speech corpus for evaluating speech enhancement algorithms.
• To conduct a subjective evaluation of 13 distinct speech enhancement methods.
• To assess algorithm performance using the ITU-T P.835 standard for speech quality.

Main Methods:

• Development of a novel noisy speech corpus.
• Subjective evaluation of 13 speech enhancement algorithms (spectral subtractive, subspace, statistical-model based, Wiener-type).
• Utilized the ITU-T P.835 methodology for assessing speech quality.

Main Results:

• The developed corpus facilitates standardized performance comparisons.
• Subjective tests provided detailed quality assessments across algorithm classes.
• Evaluated speech quality dimensions included signal distortion, noise distortion, and overall quality.

Conclusions:

• The created noisy speech corpus serves as a valuable benchmark for speech enhancement research.
• The study provides crucial insights into the comparative performance of various speech enhancement algorithms.
• Standardized evaluation methodologies are essential for advancing speech enhancement technology.