Predicting speech intelligibility based on the signal-to-noise envelope power ratio after modulation-frequency...

Area of Science:

• Acoustics
• Speech Processing
• Psychoacoustics

Background:

• Accurate prediction of speech intelligibility in noisy conditions is crucial for developing effective hearing aids and communication systems.
• Existing models, like the speech transmission index, may not fully capture intelligibility in complex noise scenarios such as spectral subtraction.
• Understanding the impact of noise characteristics on speech perception is an ongoing area of research.

Purpose of the Study:

• To propose and validate a novel model for predicting the intelligibility of processed noisy speech.
• To assess the model's performance across different noise types, including stationary noise, reverberation, and spectral subtraction.
• To investigate the underlying mechanisms contributing to intelligibility changes, particularly in spectral subtraction conditions.

Main Methods:

• Developed a speech-based envelope power spectrum model, drawing parallels with existing modulation detection models.
• Estimated the speech-to-noise envelope power ratio (SNRenv) at the output of a modulation filterbank.
• Related the SNRenv metric to speech intelligibility using an ideal observer framework.
• Compared model predictions with empirical intelligibility data for speech in various noise conditions.

Main Results:

• The proposed model demonstrated good agreement with intelligibility data for speech in stationary noise, reverberation, and spectral subtraction.
• Analysis revealed that for spectral subtraction, intelligibility reduction was linked to the estimated noise envelope power exceeding speech envelope power.
• The model successfully predicted intelligibility where the classical speech transmission index failed, particularly in spectral subtraction scenarios.

Conclusions:

• The speech-to-noise envelope power ratio (SNRenv) derived from a modulation filterbank is a robust predictor of speech intelligibility.
• The proposed model offers improved accuracy over traditional methods for processed noisy speech.
• Modulation-based signal-to-noise ratio measures are key to understanding and predicting speech intelligibility in challenging acoustic environments.