Prosody contains information that is lost when utterances are transcribed into letters or characters. This paper is concerned with exploiting such information for syntactic analysis of read Japanese sentences. In our previous work, we employed 12 prosodic features, and made a statistical model to represent the relationship between those features and dependency distances. Then, by incorporating the model in our parser, which allows the use of numerical information as linguistic knowledge, we showed that prosodic information is in fact effective for syntactic analysis. In the present work, we took up 24 prosodic features, and conducted an extensive search for effective ones. Also the statistical model was modified to account for the actual distribution of the feature values. In open experiments using an ATR 503-sentence database, parsing accuracy was improved by 21.2% compared with the case where no prosodic information was used. The duration of pauses at phrase boundaries was consistent...

In this paper, we present the results that our n-gram based word string language model, combined with speaker and noise adaptation of the acoustic model, improves recognition performance of noisy broadcast news speech. The focus was brought into a remedy against recognition errors of short words. The word string language models based on POS and n-gram frequency reduced deletion errors by 17%, insertion errors by 20%, and substitution errors by 3% in Japanese TV broadcast news speech recognition. 1 Introduction Recently large vocabulary continuous speech recognition systems have been challenged with increasingly difficult tasks, as the research focus has shifted from read speech data to speech data found in the real world. The Japanese counterpart of HUB-4 [1] broadcast news recognition projects have been pursued by NHK (Japan Broadcasting Corporation) and other research institutes since 1996 [2, 3]. The research has now stepped in the stage where real time captioning of news speech [...

This paper reports our work to improve a bigram language model for Japanese TV broadcast news speech recognition. First, frequent word strings were grouped into phrases in order that the phrases were added to the lexicon as new units of recognition. The test set perplexity was improved when frequent function word strings were used as additional recognition units. The speech recognition performance was improved both by grouping function word strings and by grouping compound nouns that were selected by word association ratio. Secondly, in order to alleviate the OOV problem related with nouns, we built and tested a language model that allows switching its noun lexicon according to the domain of the article to be recognized next. 1 INTRODUCTION The DARPA Hub-4 launched in 1995 is evaluating LVCSR systems to transcribe audio recordings of broadcast news [1]. The evaluation results have been updated annually [2]. The Japanese counterpart of HUB-4 broadcast news materials has been being deve...

This paper is concerned with a system to recognize continuously uttered "bunsetsu", which is a syntactic unit in the Japanese language. This system has two main features. One is that its top-down processor is directly controlled by bottom-up information: in the first stage, the bottom-up processor recognizes relatively unambiguous parts of the input speech signal using only local acoustical information, and then the recognition result is used to control the top-down processor for restriction of the number of hypotheses. The other feature is an unsupervised learning algorithm for vowel templates. This algorithm generates new vowel templates automatically by cutting out appropriate portions from the input speech, thus acquiring new knowledge for phonemes such as devocalized and nasalized vowels. A recognition experiment shows that the "bunsetsu" recognition rate for 1022 "bunsetsu"s uttered by 1 male speaker at 3 mora/sec rate is 86 %, and 65 % at 4 mora/sec rate.