Experimental comparisons of speech transmission index prediction methods

Good speech intelligibility is the primary goal of acoustic environment control in lecture rooms and auditoria. The speech transmission index (STI) is the primary objective evaluation metric for speech intelligibility. Therefore, accurate prediction of the STI during the design stage is important for acoustic environment control in lecture rooms and auditoria. Two types of STI prediction methods based on simulated impulse responses and statistical are recommended by the IEC standard. In this study, the prediction accuracy and influencing factors of these two types of prediction methods were systematically analysed using STI measurements collected at 25 receiver positions in six rooms. The results reveal the following. (1) When the influence of the signal-to-noise ratio (SNR) is not considered, for the STI prediction method based on simulated impulse responses, the prediction accuracy is in an acceptable range with the average predicted STI discrepancy of the 25 receiver positions being −0.004 and the maximum discrepancy being 0.046. For the STI prediction method based on statistical, the average predicted STI discrepancy of the 25 receiver positions was −0.015, and the maximum discrepancy was −0.091. STI prediction based on statistical exhibited a larger prediction error at the receiver positions close to the sound source and smaller prediction error at the receiver positions farther away from the sound source. (2) When the influence of the SNR was considered, the errors of both prediction methods increased significantly. For the STI prediction method based on simulated impulse responses, the average predicted STI discrepancy of the 25 receiver positions was −0.035 and the maximum discrepancy was −0.088. For the STI prediction method based on statistical, the average predicted STI discrepancy of the 25 receiver positions was −0.030 and the maximum discrepancy was −0.112. (3) Using a sound source that is different from the onsite measurement sound source can induce noticeable errors in both methods. (4) For the STI prediction method based on statistical using the actual Q value of the sound source, the prediction accuracy of receiver positions close to the sound source can be significantly improved and the effect on receiver positions farther away from the sound source is relatively small. (5) Both methods require significant technical expertise from users, rendering it difficult to obtain accurate prediction results during the design stage.