Predicting speech intelligibility based on a correlation metric in the envelope power spectrum domain

Area of Science:

• Acoustics and Signal Processing
• Auditory Perception
• Speech Communication

Background:

• Speech intelligibility prediction is crucial for understanding auditory perception and developing assistive technologies.
• Existing models like mr-sEPSM and STOI have limitations in predicting intelligibility under diverse noise and distortion conditions.

Purpose of the Study:

• To develop a novel hybrid speech intelligibility prediction model, sEPSM^corr.
• To evaluate the model's performance against established methods, particularly under various interferer and distortion types.

Main Methods:

• The proposed sEPSM^corr model integrates the auditory front-end of the multi-resolution speech-based envelope power spectrum model (mr-sEPSM) with a correlation-based back-end inspired by the short-time objective intelligibility measure (STOI).
• The model was evaluated on its ability to account for stationary and fluctuating additive interferers, and non-linear distortions like spectral subtraction, phase jitter, and ideal time frequency segregation (ITFS).

Main Results:

• The sEPSM^corr model demonstrates a wider predictive range compared to both mr-sEPSM and STOI, successfully accounting for conditions where individual models fail.
• While showing broader applicability, sEPSM^corr exhibited lower accuracy than source models in specific conditions.
• The model, like STOI, does not account for the impact of room reverberation on speech intelligibility.

Conclusions:

• The sEPSM^corr model offers a valuable tool for assessing the impact of various interferers and distortions on speech intelligibility.
• Its capabilities extend to evaluating consequences of hearing impairment and hearing-instrument signal processing.
• Further research may be needed to address limitations such as reverberation effects.