The Extended Speech Transmission Index Predicting speech intelligibility in non-stationary noise and reverberation

The Speech Transmission Index (STI) was designed for the evaluation of the quality of speech transmission. A limitation is its usability when background noise is non-stationary. The goal of the current thesis was to increase the accuracy of the STI in non-stationary background noises.
In chapter 2, the measurement conditions were investigated. It was shown that a minimum impulse-to-noise ratio of +25 dB in was needed to accurately measure the STI. The Extended STI (ESTI) was introduced in chapter 3. The primary adaptation was the calculation of the STI per time window. The ESTI predicted intelligibility better than the STI. However, inaccuracies remained when noise had speech-like characteristics. In chapter 4 the hypothesis was tested that these inaccuracies were caused by context effects. Prediction accuracy of the context-based ESTI-model (cESTI) improved for interruption frequencies below 5 Hz. The cESTI-model was evaluated further in chapter 5. The model outperformed the ESTI-model. However, intelligibility at several interruption rates was still underestimated. In chapter 6, the performance of the cESTI-model using the data from chapter 3 was discussed. Inaccuracies with regard to speech-like noises remained. It is likely that these were still caused by modulation masking and informational masking.
The ESTI proved to be a valuable extension of the STI, although performance decreased for speech-like noises. The addition of context improved the model further, but at the cost of higher complexity. It did show how a context model can be successfully used in combination with traditional methods for estimating speech intelligibility.