Recent Topics in Speech Recognition Research at NTT Laboratories

RECENT TOPICS IN SPEECH RECOGNITION RESEARCH AT NTT LABORATORIES Sadaoki Furui, Kiyohiro Shikano, Shoichi Matsunaga, Tatsuo Matsuoka, Satoshi Takahashi, and Tomokazu Yamada NTT Human Interface Laboratories 3-9-11, Midori-cho, Musashino-shi, Tokyo, 180 Japan ABSTRACT This paper introduces three recent topics in speech recognition research at NTT (Nippon Telegraph and Telephone) Human In- terface Laboratories. The first topic is a new HMM (hidden Markov model) technique that uses VQ-code bigrams to constrain the output probability distribution of the model according to the VQ-codes of previons frames. The output probability distribution changes depending on the previous frames even in the same state, so this method reduces the overlap of feature distributions with different pho- nemes. The second topic is approaches for adapting a syllable trigram model to a new task in Japanese continuous speech recognition. An approach which uses the most recent input phrases for adap- tation is effective in reducing the perplexity and improving phrase recognition rates. The third topic is stochastic language models for sequences of Japanese characters to be used in a Japanese dictation system with unlimited vocabulary. Japanese characters consist of Kanji (Chinese characters) and Kana (Japanese alphabets), and each Kanji has several readings depending on the context. Our dicta- tion system uses character-trigram probabilities as a source model obtained from a text database consisting of both Kanji and Kana~ and generates Kanji-and-Kana sequences directly from input speech. 1. PHONEME HMM CONSTRAINED BY STATISTICAL VQ-CODE TRANSITION 1.1 Introduction Speaker-independent phoneme models need a large amount of training dam to cover the phonetic features of various speakers and various phoneme environments. However, more training data leads to broader spectral fea- ture distributions of each phoneme. One speaker's spec- tral feature distribution often overlaps the distributions of different phonemes of other speakers. This causes confu- sion and degrades recognition performance. It has widely been confirmed that transitional spectral in- formation, such as that represented by the so-called delta- cepstrum, is effective for decreasing these overlaps and improving the performance of speaker-independent rec- ognition when it is used together with instantaneous spec- tral information [1]. The delta-cepstrum attempts to model the differential spectnma. The second-order differ- ential spectrum [2][3] has also been used to further im- prove the performance. In the vector quantization (VQ) -based recognition, an- other kind of transitional spectral information can be rep- resented by VQ-code sequences. Conditional models of VQ-code transitions have been proposed to obtain accu- rate speech models [4][5]. However, it is very difficult to obtain conditional models from the training data in a real situation, since numerous parameters must be estimated. We have tried to use bigrams of VQ-code sequences to represent statistical transitional information and restrict the feature distributions to a suitable region [6]. This method reduces the overlap of feature distributions be- tween phonemes without requiring a huge amount of training data. 1.2 Bigram-constrained HMM A bigram-constrained HMM is obtained by combining a VQ-code bigram and the conventional HMM. The output probability distribution of the model changes depending on the VQ-code of the …

Content truncated to 500 words