In this paper, we will describe a new method for small vocabulary connected word recognition. This method consists of three steps: vocalic nuclei detection, vowel identification and word identification. We report on its evaluation for the first two steps on two databases: a database of connected digits, NOISEX-92, and a subset of connected letters of the French BDSON database. It is shown that using neural networks is a robust method for automatically segmenting speech into vocalic/non-vocalic classes in connected speech. Moreover, for the vowel identification, preliminary results are given. Word identification is currently underway.
Keywords: connected word recognition, noise robustness, artificial neural networks