Non-intrusive codebook-based intelligibility prediction

In recent years, there has been an increasing interest in objective measures of speech intelligibility in the speech processing community. Important progress has been made in intrusive measures of intelligibility, where the Short-Time
Objective Intelligibility (STOI) method has become the de facto standard. Online adaptation of signal processing in, for example, hearing aids, in accordance with the listening conditions, requires a non-intrusive measure of intelligibility.
Presently, however, no good non-intrusive measures exist for noisy, nonstationary conditions. In this paper, we propose a novel, non-intrusive method for intelligiblity prediction in noisy conditions. The proposed method is based on STOI, which measures long-term correlations in the clean and degraded speech. Here, we propose to estimate the clean speech using a codebook-based approach that jointly models the speech and noisy spectra, parametrized by auto-regressive parameters, using pre-trained codebooks of both speech and noise. In experiments, the proposed method is demonstrated to be capable of accurately predicting the intelligibility scores obtained with STOI from oracle information. Moreover, the results are validated in listening tests that confirm that the proposed method can estimate intelligibility from noisy speech over a range of signal-to-noise ratios.