15th International Congress of Phonetic Sciences (ICPhS-15)

Barcelona, Spain
August 3-9, 2003


Continuous Speech Recognition Without Use of High-Level Information

Yuri Kosarev, Andrey Ronzhin, Izolda Lee, Alexey Karpov

Russian Academy of Sciences, Russia

The most of methods for continuous speech recognition and understanding are based on models of a priori generation of phrases hypothesis using semantic-syntactic restrictions in the form of rules or stochastic data. This makes worse the processing of syntactically incorrect phrases that leads to worsening of understanding robustness. In this paper the method for continuous speech recognition without mentioned restrictions is proposed. This method is based on selection of the words hypotheses from the input signal by sliding analysis with the further evaluation of phrases hypotheses taking into account 3 criteria: acoustical probability, probability of inter-word time intervals and probability of total duration of a posteriori word hypothesis. This method is well coordinated with integral idea of speech processing, which is investigated in SPIIRAS. It allows to realize the robust understanding of continuous speech based on this method.

Full Paper

Bibliographic reference.  Kosarev, Yuri / Ronzhin, Andrey / Lee, Izolda / Karpov, Alexey (2003): "Continuous speech recognition without use of high-level information", In ICPhS-15, 1373-1376.