A prototype of the dialog system developed in the Department of Computer Science carries continously spoken human machine dialogs utilizing speech input and output techniques (Kleckova 1995). The main components of the system prototype are: the speech input/output interface, the acoustic phonetic recognizer, the linguistic processor, and the dialog manager.

The pitch movement in a spoken utterance that is not related to differences in meaning of the words is called the intonation. The intonation often serves an information of a broad meaning nature. For example, the falling pitch we hear at the end of a statement in Czech such as " Vlak uz odjel. (The train has already gone.)" indicates that the utterance is complete. And on the contrary, the question " Vlak uz odjel? (Has the train already gone?)" in Czech equivalent has rising intonation. For this reason, falling intonation at the end of an utterance is called a terminal intonation contour. On the other hand, rising or unvaried level intonation, often indicates incompleteness (Palkova 1994). However, Czech sentences that contain question words like kdy, co, kdo, jak (when, what, who) usually do not have any rising intonation. It is like that the words in the question suffice to indicate that the answer is expected. The fact that rising or level intonations are correlated with incompleteness and falling intonation with completeness admits other utilizations of the intonation. One of them helps to make clear the interpretation of potentially ambiguous utterances (Noth 1993). The prosody is a very complex subject. Besides the intonation the hierachy of pauses is very important. Pauses of standard length in the places of punctuation marks between syntactic units are felt as bizzare in the spontaneous speech. After several experiments have been treid out, a three-tier pause hierarchy seems acceptable in Czech.

Table 1. Three-tier pause hierarchy

To make finer distinction of pauses would require to respect semantic relations of units in the dialog.

The design of the prosody module is based on the partitioning the speech into sentences. The sentences are processed using the following method:

• each sentence is divided into n-windows
• in each of the window is assigned with thevoice energy (the first feature) and the fundamental frequency (the second feature)

The whole sentence is represented by 2n features. The quality of the pattern recognition depends on the choice of the type and of the number of features. To classify sentences according to prosody, the prosodic characteristics must be computed and then considered as features. However, the features must be normalized and their number reduced to simplify recognition which then follows. Taking into account properties of the neural network employed in the recognition the number associated with a simple sentence must be the same for each one. The optimal number of features is proposed to be set 40 (in particular we consider 20 features energy and 20 features of frequency). As to our experience, a greater number than that proposed above does not improve the recognition. The attributes determined by this procedure form another input to the linguistic module, see Fig.3.1. The structure of the modules for the prosodic information processing is introduced in Fig. 3.2.

• All the program equipment including the code SNNS (SNNS 1995) has been used in a number of experiments focused on the sentence evaluation. The list now follows:`
• assignment of sentences with the types with respect to the speaker's standpoint
• differentiating a question from other sentences
• differentiating an investigation question from other sentences
• differentiating a sentence construction from a single sentence

The experiments reported in this paper were performed on a subset of Czech sentences. The sentences were generated using the ERBA templates (Kleckova 1998). The summarize the results is introduced in the tables Tab.2, Tab.3 we can state that:

1. we are able to detect the types of the sentences.
2. the set of 80 features is has been probed (40 features energy and 40 features of frequency), but the results are unconvincing.
3. a greater number than that proposed above does not improve the recognition.

In the near future we shall use a neuronal network with different sets of prosodic features like duration of words, syllables and syllables nucleus, etc.

Table 2. Assignement of sentence with respect to the speaker’s standpoint

Table 3. Differentiating a complex sentence from a single sentence

The subsystem uses prosodic information takes the form the fundamental voice frequency and energy. The prosodic information processed using a special arranged neural network forms an important attribute of an n-element array which, being taken as an image of the sentence analyzed, enters the linguistic analyzer to be further processed (Kleckova 1997). The system described above operates over a flexible data basis by means of components of the database system ORACLE and the language SQL. Some other alternatives considered for further development of the system have been experimentally tested .