A continuous-speech recognition system for Japanese text input is described. The system is speaker dependent and recognizes continuous phrasal speech. A consonant-vowel (CV) concatenation syllable lattice and a word lattice are generated from input speech using a vector-quantization-based continuous DP matching technique. These lattices are converted into Japanese written form using a word dictionary and grammatical information. Two methods for acoustic processing are proposed: a large vocabulary word spotting technique using high-frequency words and an optimization method for adjusting CV templates to a speaker by maximizing recognition accuracy. In recognition experiments, 500 words were selected as word templates, covering more than 90% of the technical words in reports written about X-ray computer tomography scanning. 1000 CV syllables were generated using these words. A word spotting accuracy of more than 70% and a CV spotting accuracy of 72% were obtained after iterative training. The CV spotting accuracy was improved by about 10% through iterative training.>

The performance of several algorithms for the quantization of the line spectrum pair (LSP) parameters is studied. An adaptive method which utilizes the ordering property of the LSP parameters is presented. The performance of the different quantization schemes is studied on a long sequence of speech samples. For the spectral distortion measure, appropriate performance comparisons between the different quantization schemes are rendered. Experimental results indicate that high-quality synthesized speech can be obtained using the LSP parameters at relatively low rates.>

The performance of several algorithms for the quantization of the LSP (line spectrum polar) parameters is studied. An adaptive method that utilizes the ordering property of the LSP parameters is introduced. A combination of this adaptive algorithm with nonuniform-step-size quantization is shown to be very effective for encoding the LSP parameters. The performance of the different quantization schemes is studied on several sequences of speech samples. For the spectra distortion measure, appropriate performance comparisons between the different quantization schemes are rendered.>

A Japanese text input system using speech, which is now under development, is described. This system recognizes continuous phrase speech and translates it into Japanese text. This system is composed of two major parts, acoustic processing and linguistic processing. In the acoustic processing CV (consonant-vowel) syllables and words in continuous speech are detected by continuous dynamic programming (DP). As the result a CV lattice and a word lattice are obtained. In the linguistic processing, the CV lattice and word lattice are converted into written form. In the conversion process the word lattice is used in combination with the CV lattice to improve translation accuracy. A CV recognition accuracy of 63-83 % and a translation accuracy of 38-62 % was obtained as an experimental result.

A Japanese text input system using continuous speech recognition is described. This system is composed of two major parts, acoustic processing and linguistic processing. Consonant-Vowel (CV) lattice is generated by acoustic processing from input speech uttered phrase by phrase. CV syllables in continuous speech are detected by continuous dynamic programming (DP) based on SPLIT (Strings of Phoneme-LIke Templates) method. Multiple CV templates extracted from training speech data are used to improve detection accuracy. In linguistic processing, CV lattice is converted into written form using a word dictionary. CV recognition accuracy of 67 % and Japanese translation rate of 53 % are obtained as an experimental result.

This paper describes new technique for use in a word recognition system. This recognition system is especially efffective in speaker-dependent large vocabulary word recognition based on multiple reference templates. In this system, word templates are represented as sequences of descrete phoneme-like (pseudo-phoneme) templates which are automatically determined from a training set of word utterances by a clustering technique. In speaker-dependent 641 city names word recognition experiments, 96.3% recognition accuracy was obtained using 256 phoneme-like templates.