Unsupervised Segmentation of Phoneme Sequences based on Pitman-Yor Semi-Markov Model using Phoneme Length Context

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 243–252, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Unsupervised Segmentation of Phoneme Sequences based on Pitman-Yor Semi-Markov Model using Phoneme Length Context Ryu Takeda and Kazunori Komatani The Institute of Scientific and Industrial Research, Osaka University 8-1, Mihogaoka, Ibaraki, Osaka 567-0047, Japan {rtakeda, komatani}@sanken.osaka-u.ac.jp Abstract Unsupervised segmentation of phoneme sequences is an essential process to ob- tain unknown words during spoken dia- logues. In this segmentation, an input phoneme sequence without delimiters is converted into segmented sub-sequences corresponding to words. The Pitman-Yor semi-Markov model (PYSMM) is promis- ing for this problem, but its performance degrades when it is applied to phoneme- level word segmentation. This is be- cause of insufficient cues for the seg- mentation, e.g., homophones are improp- erly treated as single entries and their dif- ferent contexts are also confused. We propose a phoneme-length context model for PYSMM to give a helpful cue at the phoneme-level and to predict suc- ceeding segments more accurately. Our experiments showed that the peak per- formance with our context model out- performed those without such a context model by 0.045 at most in terms of F- measures of estimated segmentation. 1 Introduction 1.1 Motivation The final goal of our current project is to achieve the development of robots or systems that ac- quire knowledge during spoken interactions be- tween them and human beings in the open world. Unknown or new words appear frequently in our daily lives, and because their meanings may be different for the systems deployed in different ar- eas, automatic lexicon acquisition is a useful func- tion for maintenance-free spoken dialogue sys- tems. I ate tempura yesterday Goal: Unsupervised lexicon acquisition through spoken dialogue Focus of this study: Segmentation of phoneme sequences Word Phoneme I ate témpərə yesterday άɪ éɪt témpərə jéstɚdèɪ άɪéɪttémpərəjéstɚdèɪ Signal segmentation témpərə is unknown Phoneme recognition Phoneme sequence Input: Segmented tokens Output: Word conversion & context understanding Symbol Figure 1: Our target problem In this paper, we focus on the phoneme-level representation of utterance – not in the signal-level – to relax the problem, which results in an issue with phoneme sequence segmentation. The seg- mentation converts an input phoneme sequence into segmented sub-sequences corresponding to individual words. The focus of our study is illus- trated in Figure 1. When the robot listens to a hu- man utterance that includes an unknown word like “tempura,” the robot will estimate the unknown segment “tempura” by the iterative search based on trial-and-error of many hypotheses among dif- ferent layers. Here, the phoneme sequence of an unknown word is necessary as an intermedi- ate representation between signals and lexicon be- cause we cannot directly obtain the spelling of an unknown word just from sound information. Note that our assumption of a phoneme sequence given is partly supported by the high accuracy of state-of-the-art speech and phoneme recognition (Dahl et al., 2012; Hinton et al., 2012; Seide et al., 2011a,b). Approaches based on …

Content truncated to 500 words