Latent Prosody Model-Assisted Mandarin Accent Identification

Latent Prosody Model-Assisted Mandarin Accent Identification Yuan-Fu Liao1, Shuan-Chen Yeh2, Ming-Feng Tsai3, Wei-Hsiung Ting4, and Sen-Chia Chang5 1,2,3,4Department of Electronic Engineering, National Taipei University of Technology 5Advanced Technology Center, Information and Communications Research Laboratories, Industrial Technology Research Institute 1,2,3,4yfliao@ntut.edu.tw, 5chang@itri.org.tw Abstract A two-stage latent prosody model-language model (LPM-LM)-based approach is proposed to identify two Mandarin accent types spoken by native speakers in Mainland China and Taiwan. The frontend LPM tokenizes and jointly models the affections of speaker, tone and prosody state of an utterance. The backend LM takes the decoded prosody state sequences and builds n-grams to model the prosodic differences of the two accent types. Experimental results on a mixed TRSC and MAT database showed that fusion of the proposed LPM-LM with a SDC/GMM+PPR-LM+UPR-LM baseline system could further reduced the average accent identification error rate from 20.7% to 16.2%. Therefore, the proposed LPM-LM method is a promising approach. Keywords: Accent recognition, latent prosody model, Mandarin, Taiwan 1. Introduction Over the past decades, many approaches have been proposed to deal with language identification (LID) tasks. They tried to capture the specific characteristics of different languages. These characteristics roughly fall into three categories: the phonetic repertoire, the phonotactics, and the prosody. The mainstream system (as shown in NIST language recognition evaluation (LRE) 2007) [1] is usually based on the fusion of multiple acoustic and phonotactic systems. Although LID is extensively studied, less works have been done on accent identification (AID), especially for native speakers, such as American and Indian English, Mainland China and Taiwan Mandarin, Hindi and Urdu Hindustani and Caribbean and non-Caribbean Spanish. Comparing with LID task, AID of native speakers is more challenging because, (1) some linguistic knowledge, such as syllable structure, may be of little use since native speakers seldom make such mistakes; (2) difference among those speakers is relatively smaller than 125 that among foreign (non-native) speakers. In other words, the capacities of the popular acoustic and phonotactic approaches may be limited in this case. Many approaches have been proposed to model the prosodic differences between languages, dialects or accents [2], recently. Most of them are based on direct modeling of surface prosodic features, i.e., the raw prosodic features. For example, frame-level pitch flux features and GMMs were proposed in [3]; segmental-level pitch features were extracted using Legendre polynomials and modeled by ergodic Markov model in [4]; and supra-segment-level prosodic features were captured by n-gram in [5]. Figure 1. The block diagram of the proposed LPM-LM-based Mandarin accent identification system. Figure 2. The block diagram of the proposed LPM framework (speaker factor is omitted to simply this figure). However, surface prosodic features are often affected by many other non-prosodic latent factors, such as channel, speaker, phonetic context, and so on. Therefore, it is necessary to apply some feature normalization methods [6] to alleviate the unwanted affections. To absorb those unwanted affections, in this study a two-stage latent prosody model-language model (LPM-LM)-based approach as shown in Fig. 1 and 2 is proposed. The aim is to discriminate two …

Content truncated to 500 words