Speech coherence index with real speech and reverberation

Speech Coherence Index (SCI) is a proposed method of estimating speech intelligibility in real time with program material. The coherence of the complex valued transfer function is used to estimate the signal-to-noise ratio on a per frequency basis. The transfer function is calculated using short time windows at high frequencies and longer time windows at low frequencies to mimic the multi-resolution nature of human hearing. SCI has been shown to produce identical results to Speech Transmission Index (STI) in the case of pure noise interference. Like STI, SCI has been shown to decrease with a single reflection at longer latencies or at greater magnitude. SCI always decreases monotonically with single reflection latency, whereas STI varies up and down at extremely long latencies. For simulated reverberation, both STI and SCI have been shown to decrease with increasing reverberation time and reverberant level. However, SCI is more sensitive to direct-to-reverberant level. This paper will compare SCI to STI under more realistic conditions, such as speech signals and real-world impulse responses. The effect of signal crest factor will also be examined.