Makoto Hiroshige

Estimation of multi-templates for speech recognition by using spectrum reforming and self-organized clustering

This report proposes a new recognition method of continuous Japanese speech. In order to overcome some difficulties in a large vocabulary recognition system, which requires large calculation cost and memory area, the proposed system only recognizes the single phonemes which are automatically selected from continuous speech. The automatic selection can be realized using a new spectrum reforming technique, i.e., a method which smooths the estimated parameters using a nonlinear filter and a parameter quantization method based on a priori rules. This system can also easily search the optimal template for speech phonemes using the self-organized clustering method. This technique is suitable for optimum estimation of multi-templates for continuous speech recognition.

Effectiveness of A Direct Speech Transform Method Using Inductive Learning from Laryngectomee Speech to Normal Speech

This paper proposes and evaluates a new direct speech transform method with waveforms from laryngectomee speech to normal speech. Almost all conventional speech recognition systems and other speech processing systems are not able to treat laryngectomee speech with satisfactory results. One of the major causes is difficulty preparing corpora. It is very hard to record a large amount of clear and intelligible utterance data because the acoustical quality depends strongly on the individual status of such people.

Evaluation of Direct Speech Translation Method Using Inductive Learning for Conversations in the Travel Domain

This paper evaluates a direct speech translation Method with waveforms using the Inductive Learning method for short conversation. The method is able to work without conventional speech recognition and speech synthesis because syntactic expressions are not needed for translation in the proposed method. We focus only on acoustic characteristics of speech waveforms of source and target languages without obtaining character strings from utterances. This speech translation method can be utilized for any language because the system has no processing dependent on an individual character of a specific language. Therefore, we can utilize the speech of a handicapped person who is not able to be treated by conventional speech recognition systems, because we do not need to segment the speech into phonemes, syllables, or words to realize speech translation. Our method is realized by learning translation rules that have acoustic correspondence between two languages inductively. In this paper, we deal with a translation between Japanese and English.

Phoneme recognition using modified TDNN and a self-organizing clustering network

This paper presents a new approach to phoneme recognition system. A modified Time-delay Neural Network (TDNN) based on similarity vectors of clustering node information is developed for this purpose. The speech data have been analysed first by time varying ARMA-D model to have better response of its time varying characteristics. For the generation of the similarity vectors of the clustering nodes, Self-Organising Clustering process is used. To study the performance of this system, the speaker-independent recognition of the voiced explosive(stop) consonants /b,d,g/ in varying phonetic contexts is taken as the initial recognition task. This system gives a recognition rate for the stop consonants of about 84.3% for speaker independent speech data. For all these experiments, Japanese speech data is used supplied by ATR, Japan. The time taken for the training and recognition by the system can be considered reasonable.

An Adaptive Signal Processing System for Voiced Speech with Automatic Spectrum Selector

In this paper, adaptive spectrum analysis algorithms for voiced speech are discussed. A new adaptive signal processing system which uses a modified MIS algorithm and has a spectrum selector using a neural network is proposed. Based on the properties of voiced speech, we first test an identification algorithm which estimates AR parameters from the first moment of the observed signal and investigate its sensitivity to noise. In order to compare the modified MIS for voiced speech analysis with the above first moment analysis, the frequency domain properties of weighting factor A which is used in the modified MIS are presented. This paper shows that we should select only the accurate spectra from among the results given by the modified MIS. The selection of spectra is automatically performed using a neural network. Using above methods, we construct a new analysis system for voiced speech. Experiments on real speech show that the proposed system is effective for speech analysis.