Microscopic and Blind Prediction of Speech Intelligibility: Theory and Practice

Area of Science:

• Speech Processing
• Acoustics
• Signal Processing

Background:

• Estimating speech intelligibility objectively, without listening tests, is crucial for speech processing applications.
• Existing objective speech intelligibility prediction (SIP) methods often average over longer time spans (macroscopic view).
• There is a need for microscopic evaluation of SIP methods for more granular assessment.

Purpose of the Study:

• To develop a theoretical framework for the microscopic evaluation of speech intelligibility prediction (SIP) methods.
• To introduce a novel, blind SIP method called DIscriminative Speech intelligibility Predictor (DISP).
• To evaluate the performance of microscopic SIP methods, including the new DISP, against theoretical limitations.

Main Methods:

• Developed a theoretical framework for microscopic SIP evaluation, deriving a Statistically estimated Accuracy based on Theory (StAT) metric.
• Evaluated a state-of-the-art microscopic SIP approach using automatic speech recognition (ASR) within the StAT framework.
• Introduced and evaluated a novel, fully blind DIscriminative Speech intelligibility Predictor (DISP) using the StAT framework.

Main Results:

• The StAT framework quantifies inherent statistical limitations in microscopic SIP.
• The ASR-based microscopic SIP approach showed results consistent with theoretical predictions.
• The novel blind DISP method achieved comparable or superior accuracy to the non-blind ASR-based approach.

Conclusions:

• The developed StAT framework provides a theoretical basis for evaluating microscopic SIP methods.
• The blind DISP estimator offers a highly accurate and practical solution for objective speech intelligibility prediction.
• Microscopic SIP evaluation, particularly with blind methods like DISP, holds significant promise for speech technology.