Speaker
Description
The short-time objective intelligibility metric (STOI) [Taal et al., 2011, IEEE Trans. Audio, Speech and Lang. Proc., Vol. 19, No. 7] is a widely used predictor of speech intelligibility and is often applied to evaluate the effects of noise-reduction algorithms and other speech-enhancement strategies. However, STOI is not level sensitive and does not account for the effects of hearing loss, which limits its ability to assess the ultimate benefit of such signal manipulations when implemented in hearing-assistive devices.Here we propose a short-time objective audible intelligibility metric (STOAI), derived from STOI, that introduces a level-sensitive processing stage in the form of internal noise in the front end of the model. This stage is adjusted using the audiometric thresholds of individual listeners to account for their loss of sensitivity. Inspired by Plomp’s model of speech intelligibility, we evaluated STOAI on speech reception thresholds (SRTs) in both quiet and noise, such that the impact of the audibility and distortion components of hearing loss could be assessed independently. Overall, individual STOAI correlated well with SRTs in quiet demonstrating that it accounts for the effects of audibility. However, STOAI only partially accounted for elevated SRTs in noise, consistent with the presence of suprathreshold deficits (distortion factor) in the studied population. STOAI extends the STOI framework to account for audibility-related effects and can therefore provide a useful tool for the future development and evaluation of speech-enhancement algorithms targeting hearing-impaired listeners.