A data-driven speech intelligibility assessment method using sum-sorted spectrogram feature

A novel data-driven non-intrusive method to assess speech intelligibility is proposed. The approach uses a new segment-based feature called Sum-Sorted Spectrogram (SSS) and a logistic regression network to predict the intelligibility score of degraded speech. Experiment results show that this approach predicts speech intelligibility with an RMS error of 0.07 against short time objective intelligibility (STOI) index on a test database of noisy speech, and a Spearman Correlation Coefficient (SCC) of 0.98.