Kazuyuki Takagi

Segmentation of spoken dialogue by interjections, disfluent utterances and pauses

The paper attempts to segment spontaneous speech of human to human spoken dialogues into a relatively large unit of speech, that is, a sub phrasal unit segmented by interjections, disfluent utterances and pauses. A spontaneous speech model incorporating prosody was developed, in which three kinds of speech segment models and the transition probabilities among them were specified. The segmentation experiments showed that 87.6% of the segment boundaries were located correctly within 50 msec, 81.2% within 30 msec, which showed 10.1 point increase in performance comparing with the initial model without prosodic information.

An Efficient Algorithm for Japanese Sentence Compaction Based on Phrase Importance and Inter-Phrase Dependency

This paper describes an efficient algorithm for Japanese sentence compaction. First, a measure of grammatical goodness of phrase sequences is defined on the basis of a Japanese dependency grammar. Also a measure of topical importance of phrase sequences is given. Then the problem of sentence compaction is formulated as an optimisation problem of selecting a subsequence of phrases from the original sentence that maximises the sum of the grammatical goodness and the topical importance. A recurrence equation is derived by using the principle of dynamic programming, which is then translated into an algorithm to solve the problem. The algorithm is of polynomial-time with respect to the original sentence length. Finally, an example of sentence compaction is presented.

Evaluation of a Japanese Sentence Compression Method Based on Phrase Significance and Inter-Phrase Dependency

Sentence compression is a method of text summarisation, where each sentence in a text is shortened in such a way as to retain the original information and grammatical correctness as much as possible. In a previous paper, we formulated the problem of sentence compression as an optimisation problem of extracting a subsequence of phrases from the original sentence that maximises the sum of topical importance and grammatical correctness. Based on this formulation an efficient sentence compression algorithm was derived. This paper reports a result of subjective evaluation for the quality of sentences compressed by using the algorithm.

The Use of Overlapped Sub-Bands in Multi-Band, Multi-SNR, Multi-Path Recognition of Noisy Word Utterances

A solution to the problem of improving robustness to noise in automatic speech recognition is presented in the framework of multi-band, multi-SNR, and multi-path approaches. In our word recognizer, the whole frequency band is divided into seven-overlapped sub-bands, and then sub-band noisy phoneme HMMs are trained on speech data mixed with the filtered white Gaussian noise at multiple SNRs. The acoustic model of a word is built as a set of concatenations of clean and noisy sub-band phoneme HMMs arranged in parallel. A Viterbi decoder allows a search path to transit to another SNR condition at a phoneme boundary. The recognition scores of the sub-bands are then recombined to give the score for a word. Experiments show that the overlapped seven-band system yields the best performance under nonstationary ambient noises. It is also shown that the use of filtered white Gaussian noise is advantageous for training noisy phoneme HMMs.

Japanese Dependency Structure Analysis Using Information about Multiple Pauses and F0

Syntax and prosody are closely related to each other. This paper is concerned with the problem of exploiting pause information for recovering dependency structures of read Japanese sentences. Our parser can handle both symbolic information such as dependency rule and numerical information such as the probability of dependency distance of a phrase in a unified way as linguistic information. In our past work, post-phrase pause that immediately succeeds a phrase in question was employed as prosodic information. In this paper, we employed two kinds of pauses in addition to the post-phrase pause: post-post-phrase pause that immediately succeeds the phrase that follows a phrase in question, and pre-phrase pause that immediately precedes a phrase in question. By combining the three kinds of pause information linearly with the optimal combination weights that were determined experimentally, the parsing accuracy was improved compared to the case where only the post-phrase pause was used as in our previous work. Linear combination of pause and fundamental frequency information yielded further improvement of parsing accuracy.

Formant frequency estimation by moment calculation of the speech spectrum

Moment calculation is applied to extract the formant frequencies of a speech spectrum. Three kinds of first‐order moments divide a spectrum into four frequency regions. The centers of gravity of the first three regions are calculated to give the 0th order estimation of the 1st, 2nd, and 3rd formant frequencies. Then the upper and the lower bounds of each region are modified so that the estimated frequency comes closer to the major peak of the spectrum, utilizing the second‐order and the third‐order moments that represent the variance and skewness of the spectral pattern. The process repeats until the k th estimation equals the (k − 1) th estimation. This modification improves the estimation precision significantly. An experiment with model spectra generated by an all‐pole model gave estimation precision of 3% using formant frequencies typical of the five Japanese vowels. Speech materials spoken by five male and five female speakers were used for this experiment. The speech waveform was sampled at a rate o...