A Binaural Grouping Model for Predicting Speech Intelligibility in Multitalker Environments.

This was a computational model tested in simulated anechoic mixtures, not a clinical study of patients or hearing-aid users. The model did not include internal binaural noise, which the authors say could reduce the benefit of using binaural cues. The model was intended only for anechoic conditions, so it does not address reverberation or ordinary room acoustics. The threshold-setting method that works without knowing target and masker waveforms assumes speech is sparse in time-frequency and roughly equal in long-term power, an idealization. The model uses a binary mask, but the authors state that the brain is unlikely to perform a completely binary operation, so unwanted components may still distract listeners. The predicted speech reception thresholds for separations of 15 degrees or more were 3 to 4 dB lower than measured, so the model did not match the measured benefit exactly. The model uses only binaural cues, while natural speech also has pitch, common onsets and offsets, and other grouping cues.

Do not read this as proof that a hearing-aid algorithm or clinical therapy improves speech understanding. The paper reports a simulated model tested in anechoic mixtures and against published listening data, and its predicted speech reception thresholds were 3 to 4 dB lower than measured for separations of 15 degrees or more.