Presentation | 2020-12-02 Approximation Method for Bayes Optimal Prediction in Phoneme Recognition Problem Taishi Yamaoka, Shota Saito, Toshiyasu Matsushima, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | In this paper, we propose a method of phoneme recognition. In the previous studies on phoneme recognition using the Hidden Markov Model, the Hidden Markov Model used for prediction is defined as one by a certain criteria. In addition, for the defined Hidden Markov Model, parameters were estimated from the training data, and the phonemes corresponding to the new speech data were predicted using paremters. In this peper, we assume 0-1 loss as the loss function, and formulate the optimum prediction based on Bayesian criterion. In other words, instead of selecting one Hidden Markov Model and estimating its parameters and making predictions using them, we propose a prediction that directly minimizes the probability of error in the prediction. Although this prediction is theoretically optimal, its calculation involves two problems: (i) The complexity of the sum calculation of the state transition series is on the exponential order with respect to the length of the voice. (ii) It is difficult to analytically calculate the integral by the posterior distribution of the parameters of the Hidden Markov Model. In order to solve these problems, in this paper, we apply the Viterbi algorithm for problem (i) and the Variational Bayesian method for problem (ii), and propose a Bayesian semi-optimal algorithm. This algorithm makes predictions by weighted averages of approximate posterior distributions of multiple Hidden Markov Models. By conducting numerical experiments using artificial data, it was confirmed that the proposed method has a smaller false recognition rate than the method of selecting and predicting one model as in the previous research. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Phoneme recognition / Hidden Markov model / Bayes criteria |
Paper # | IT2020-30 |
Date of Issue | 2020-11-24 (IT) |
Conference Information | |
Committee | IT |
---|---|
Conference Date | 2020/12/1(3days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Online |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | Lectures for Young Researchers, General |
Chair | Tadashi Wadayama(Nagoya Inst. of Tech.) |
Vice Chair | Tetsuya Kojima(Tokyo Kosen) |
Secretary | Tetsuya Kojima(Yamaguchi Univ.) |
Assistant | Takahiro Ohta(Senshu Univ.) |
Paper Information | |
Registration To | Technical Committee on Information Theory |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Approximation Method for Bayes Optimal Prediction in Phoneme Recognition Problem |
Sub Title (in English) | |
Keyword(1) | Phoneme recognition |
Keyword(2) | Hidden Markov model |
Keyword(3) | Bayes criteria |
1st Author's Name | Taishi Yamaoka |
1st Author's Affiliation | Waseda University(Waseda Univ.) |
2nd Author's Name | Shota Saito |
2nd Author's Affiliation | Waseda University(Waseda Univ.) |
3rd Author's Name | Toshiyasu Matsushima |
3rd Author's Affiliation | Waseda University(Waseda Univ.) |
Date | 2020-12-02 |
Paper # | IT2020-30 |
Volume (vol) | vol.120 |
Number (no) | IT-268 |
Page | pp.pp.32-37(IT), |
#Pages | 6 |
Date of Issue | 2020-11-24 (IT) |