Analysis of speech-based Speech Transmission Index methods with implications for nonlinear operations

Area of Science:

• Acoustics
• Speech processing
• Signal analysis

Background:

• The Speech Transmission Index (STI) traditionally measures speech intelligibility degraded by linear distortions like noise and reverberation.
• Existing STI methods using modulated noise are insufficient for assessing nonlinearly processed speech.
• Extending STI to non-linear conditions requires probe signals that mimic speech characteristics.

Purpose of the Study:

• To evaluate existing and novel speech-based Speech Transmission Index (STI) methods.
• To assess the performance of these methods under various degradation conditions, including nonlinear operations.
• To identify reliable methods for predicting speech intelligibility in complex acoustic environments.

Main Methods:

• Compared four established speech-based STI methods against four novel approaches.
• Tested methods under conditions of additive noise, reverberation, envelope thresholding, and spectral subtraction.
• Analyzed intermediate STI metrics to understand method failures with nonlinear distortions.

Main Results:

• No previously proposed speech-based STI method adequately handled all tested conditions.
• Four novel proposed methods demonstrated qualitatively reasonable performance.
• Analysis revealed specific reasons for method failures in the presence of nonlinear operations.

Conclusions:

• Existing speech-based STI methods are insufficient for predicting intelligibility with nonlinear speech processing.
• Novel speech-based STI methods show potential for improved intelligibility prediction.
• Further research on these novel methods is warranted, with implications for cochlear implant speech processing.