Review of the Binaural Speech Intelligibility Model (BSIM)
* Presenting author
Abstract:
The Binaural Speech Intelligibility Model (BSIM) predicts speech intelligibility for spatially separated target speech and interferers. Several model extensions have been introduced adapting BSIM to different conditions: The first version of BSIM used a long-term analysis of target speech and interferers and was successfully evaluated using unmodulated noises in different room acoustics with listeners with normal and impaired hearing. Alternatively, BSIM can use short-term processing predicting masking release for modulated maskers. A further extension takes into account that human binaural processing follows changes of interaural differences in a sluggish manner. Furthermore model extensions distinguishing between useful and detrimental speech reflections adapted BSIM to far-field target speech. Recently a blind version of BSIMs binaural preprocessing was introduced, which requires only a mixture of target speech and interferers and does not require them as clean signals. This blind front-end can be combined with arbitrary back-ends predicting speech intelligibility including blind back-ends.However, so far BSIM only predicts energetic masking and is not able to predict (release from) informational masking in speech-in-speech conditions.This review demonstrates the advantages and disadvantages of the different BSIM settings in different evaluation studies and possible further developments.