On the Identification of Word-Boundaries using Phonological Rules for Speech Recognition and Labeling

In this paper we studied the phonemic structure of the words’ beginnings and endings in standard European Portuguese (hereafter EP). The generativist description of the Portuguese phonology [1] was used as framework basis and the phonetic and acoustic experiments performed by Delgado-Martins [2] served as a model to the phonetic background in EP. We also compared the results between the expected phonological phonemes in word composition (considering the initial and final phonemes) and the phonetic reality in real speech productions affected by dialectal and contextual (also coarticulation) phenomena. To do so, two different corpora were used: one with text output (vide Linguateca) to find grapheme realizations in the beginnings and endings of words; and the other with oral output (vide Teixeira et al. 2001) to have a phonetic perspective of what happens in natural continuous speech production and confirm some of the Portuguese language phonetic tendencies. The most important conclusion obtained from this study is probably the confirmation that the syllabic re-structuring is an ongoing process in modern EP spoken language. In other words, the non-stressed vowel suppression is allowing to form single consonant syllables. Other trends in the distribution of phonemes in word endings and beginnings are also presented and discussed. Our purpose with this work is to contribute with linguistic rule-based information to improve word-boundary segmentation in voice recognition systems. It is well known that the task of word identification is often a difficult task in spontaneous speech, due to the lack of pause boundaries. Therefore, we believe that the stochastic methods usually used in word recognition systems, such as HMM (Hidden Markov Models), Neural Networks and so on, can be strengthened and complemented if some linguistic rules and phonological/ phonetic tendencies are considered.