Method for Asr Performance Prediction Based on Temporal Properties of Speech Signal

Extending previous work on prediction of phoneme recognition error from unlabelled data, corrupted by unpredictable factors, the current work investigates a simple but effective method of estimating ASR performance by computing Mean Temporal Distance (MTD), which is the mean distance between speech feature vectors, determined as a function of temporal distance between the vectors. It is shown that MTD is a function of the signal-to-noise ratio of the speech signal. Comparing MTD curves, derived on data used for training of the classifier, and on test utterances, allows for predicting error on the test data. Another interesting observation from the proposed technique is that the Mean Temporal Distance remains approximately constant, as temporal separation exceeds certain critical interval (about 200 ms), corresponding to the extent of coarticulation in speech sounds. This lends further support to the notion that speech message is coded in overlapping speech sound units, lasting approximately 200 ms.