A syllabic-filler-based continuous speech recognizer for unlimited vocabulary

Most continuous speech recognition systems make the assumption that the input speech to process is limited to correspond to a text from a given dictionary, while fluent speech is, in fact, characterized by the addition of out-of-dictionary words, hesitations and interrupted words. The authors describe an unlimited vocabulary continuous speech recognizer based on new-word detection and transcription. In the system the words from a pre-defined dictionary as well as the extraneous speech (unknown words) use simply the same context-dependent phoneme hidden Markov models (HMMs) trained on data including only known words. The distinction between dictionary words and unknown words is made during the search by the addition of lexical fillers in the lexical tree used by the two-pass Viterbi type algorithm of the scoring method as well as by modifying the language models. A performance comparison is given for phonemic as well as syllabic fillers