stoI

STOI, the Short-Time Objective Intelligibility metric, is an objective measure used to predict human speech intelligibility from processed audio signals. It is widely employed in evaluating speech enhancement, coding, and cochlear implant processing, where distortions such as noise, reverberation, or nonlinear processing can affect intelligibility. STOI focuses on predicting how understandable speech would be to a listener, rather than assessing overall sound quality.

The method compares a reference clean speech signal to a degraded version. Both signals are analyzed with an auditory-inspired filterbank to produce multiple frequency-band representations. In each band, short-time envelopes are extracted and aligned in time. The core step is computing correlations between the envelopes of the reference and degraded signals over short temporal windows and across bands. These correlations are averaged to yield a single score, typically in the range from 0 to 1, where higher values indicate better predicted intelligibility. The approach emphasizes distortions in the temporal envelope, which are closely tied to intelligibility.

STOI is valued for its computational efficiency and its strong correlation with human intelligibility under a range of distortions, especially additive noise and channel distortions. It requires a clean reference signal, making it a relative measure rather than an absolute quality score. It may be less sensitive to distortions that do not substantially affect temporal envelopes and does not directly measure perceived listening quality.

Variants and extensions exist, such as ESTOI (Extended STOI), which aim to improve prediction by incorporating more temporal context. STOI remains a standard benchmark in evaluating speech processing algorithms and algorithms for intelligibility assessment.