Application of Speech Recognition to Automatic Intelligibility Testing Procedures

Although most automatic intelligibility testing procedures are based on the articulation index principle, speech recognition techniques offer an attractive alternative to existing procedures. In essence, recognition provides a mechanism by which the communication system can be tested with realistic speech samples rather than sine or square waves. Since formatted test samples can be used, segmentation and time normalization problems are circumvented. A recognition process, which makes use of statistical decision theory to imitate the results of a listening jury, is discussed. Data on the ability of the system to extrapolate learning data automatically to predict the results of new degradations on the same listening jury is presented. Details on the advantages and disadvantages of such techniques are discussed. [This work was supported by the Air Force Systems Command at the Rome Air Development Center, Rome, New York.]