Spotting Prosodic Boundaries in Continuous Speech in French

Two kinds of marks have been set by the listeners (frontiers and accents), which are attached to the syllable nucleus. The MLP fed with any of the previously described values (F0,duration...), no matter the size of the temporal window, is not capable of reproducing the accent marking with a good score. Thus we consider that listeners' accent marks are not consistent, at least from a local point of view. But for the frontier marks, the MLP fed with the duration, on a 5 vowel context , achieves the task with 11% insertion and 43% omissions. Phonetician marks At this stage, we use the auditory marks to select a significative subset of marks set by the expert. Considering the given number of mark types obtained, we found it necessary to gather them in generic classes to achieve a correct training of the MLP : R for initial rise (129 occurrences), P for peaks (128), B for baseline (105), C for continuation rise (50), Nil for no marking at all (1287). After several tests, we kept vowel duration, F0 values, and pseudo-syllable duration on a 7 vocalic nucleus window to feed a MLP with 10 neurons in its hidden layer. The MLP has 5 outputs: one for each class mentioned above. Nil 227 6 0 4 3 B 7 20 0 0 0 C 0 5 7 1 1 P 5 0 1 25 5 R 0 0 0 5 34 The MLP gives no answer for 44 configurations (concurrent answers). Surprisingly, no nasality tag is required to draw the MLP attention on the fact that nasal vowels are much longer than vocalic ones. RESULTS AND CONCLUSION The main result is that this experience validates both the expert prosodic marking and the automatic spotting system. Furthermore, the confusion rate between P and R marks is rather low, which agrees with the results of [4]: lengthening is a more important correlate of F0 peak for P than for R. R marks recognized as P, are accented monosyllabics words. The recognition rate for C is enhanced when we add F0 regression parameters, as involved vowels bear a long upward F0 move. However this adds a slight confusion in the identification of P marks. Future work will aim at incorporating long term prosodic variations in the modelling of our prosodic marks. REFERENCES [1] J. Vaissière (1982), «A supra-seg-mental component in a French speech recognition system: …