We have developed LITHAN (LIsten-THink-ANswer) speech understanding system which automatically recognizes continuously uttered speech utilizing higher linguistic information such as syntactic, semantic and pragmatic information. This system predicts possible words utilizing linguistic information at the unrecognized portion of the input utterance, and identifies each predicted word by the use of the optimum matching algorithm between a recognized phoneme string and the representative one in the word dictionary. We propose an effective tree search method of parsing when the results of phoneme recognition and word identification are not error free. LITHAN uses many types of a priori information; the statistic of each phoneme; the similarity matrix between phonemes; the word dictionary; the spoken grammar with the additional information as regards the spoken grammar; the semantic and pragmatic information. We have applied this efficient, flexible system to restricted utterances with vocabulary of about 100 words which concerned with operational commands and queries of the status of a computer network. According to the results tested on a sample 200 sentences spoken by 10 male speakers at a normal speed, 64% of the sentences and 93% of the output words were correctly recognized.

In order to evaluate the speaker-factor in uttered words, multivariate analysis of variance for four-factor design with repeated measurements has been applied to the analysis of the spectral vectors obtained from Japanese V1 CV2 utterances by 10 adult male speakers, where V1, V2/a, i, u, e, o/ and Cm, n, η/. The speaker, the vowel V1, the consonant C, and the vowel V2 have been assigned to the four factors, respectively, and the interaction between speaker-factor and V1, C or V2-factor has been especially interesting in the analysis. As the results, what we should take into account as to the co-articulation of a certain phoneme are (1) the main effect of the phoneme, (2) the main effect of the just preceding (or following) phoneme, (3) the interaction between the phoneme and the just preceding (or following) phoneme, and (4) the main effect of the speaker-factor. The interactions between the speaker-factor and these phonemes are relatively small, and the influence of the speaker-factor to the co-articulation is not so complicated. These facts imply possibility of the speaker-independent description about the phoneme and the rule of co-articulation.

A new depth measurement method is described in this paper. This method uses the parallel planes of light each of which flickers in the time domain according to the binary code uniquely assigned to it. When the code length is n bits, we input n pictures projected on a certain object in which we locate and "identify" each slit image. Then the depth to the points on the slit images can be calculated by triangulation. The experimental results show that this method is faster and stronger for noise than the conventional methods. We aim to investigate the importance of knowledge about the task domain being used in signal level processing. Picture processing systems are usually task dependent, so the knowledge about the domain can be applied even in signal level processing. The knowledge is, we believe, more powerful to be used at earlier stage in picture processing than to be used at the latter one. A result about the transformation from knowledge to signal-to-noise ratio is shown as an example.