ESCA Tutorial and Research Workshop on
An Intelligibility test has been ran to assess various French synthetisers. Semanticaliy unpredictable sentences were used. The distribution of responses of listeners exhibits a strong relationship between the percentages of correct sentences and correct words. The ratio of their logarithms seems to be a powerful index for the quantification of the complexity of a spoken message. Repiotted data from the litterature confirm the hypothesis that the more the contextual (semantic, syntactic, etc.) cues in a sentence, the lower this ratio of logarithms. This index could be related to the number of perception units a listener must deal with when hearing a sentence ; the sentence being considered as a sequency of more or less related information units. The perception of sentences is here distorted by synthesiser transmissions. Omissions and mistakes do not match the binomial law of their theoretical distribution which could be expected when considering a simple model, where all input units (e.g. words in our case) would be of equal probability to be identified. Discrepancy between observations and theory is here analysed. It provides a fruitful explanation to the fact that linguistic relationships between units do correct "expected" wrong words and distort "expected" correct ones. This correction/distortion essentially depends on the linguistic content of sentences which may thereon be quantified by means of the suggested index. This index also shows smaller variations related to other factors as listeners ability, training and the degradation level of acoustic presentation.
Bibliographic reference. Benoît, Christian (1989): "Towards the perceptual quantification of context redundancy in sentences", In SIOA-1989, Vol.2, 211-214.