Automatic Prosodic Labeling with Conditional Random Fields and Rich Acoustic Features

Automatic Prosodic Labeling with Conditional Random Fields and Rich Acoustic Features Gina-Anne Levow University of Chicago Department of Computer Science 1100 E. 58th St. Chicago, IL 60637 USA levow@cs.uchicago.edu Abstract Many acoustic approaches to prosodic la- beling in English have employed only lo- cal classifiers, although text-based classifi- cation has employed some sequential mod- els. In this paper we employ linear chain and factorial conditional random fields (CRFs) in conjunction with rich, contextually-based prosodic features, to exploit sequential de- pendencies and to facilitate integration with lexical features. Integration of lexical and prosodic features improves pitch accent pre- diction over either feature set alone, and for lower accuracy feature sets, factorial CRF models can improve over linear chain based prediction of pitch accent. 1 Introduction Prosody plays a crucial role in language understand- ing. In addition to the well-known effects in tone languages such as Chinese, prosody in English also plays a significant role, where pitch accents can indicate given/new information status, and bound- ary tones can distinguish statements from yes-no questions. However, recognition of such prosodic features poses significant challenges due to differ- ences in surface realization from the underlying form. In particular, context plays a significant role in prosodic realization. Contextual effects due ar- ticulatory constraints such maximum speed of pitch change (Xu and Sun, 2002) from neighboring sylla- bles and accents can yield co-articulatory effects at the intonational level, analogous to those at the seg- mental level. Recent phonetic research (Xu, 1999; Sun, 2002; Shen, 1990) has demonstrated the im- portance of coarticulation for tone and pitch accent recognition. In addition context affects interpreta- tion of prosodic events; an accent is viewed as high or low relative to the speaker’s pitch range and also relative to adjacent speech. Some recent acoustically focused approaches (Sun, 2002; Levow, 2005) to tone and pitch accent recognition have begun to model and exploit these contextual effects on production. Following the Par- allel Encoding and Target Approximation (PENTA) (Xu, 2004), this work assumes that the prosodic tar- get is exponentially approached during the course of syllable production, and thus the target is best ap- proximated in the later portion of the syllable. Other contextual evidence such as relative pitch height or band energy between syllables has also been em- ployed (Levow, 2005; Rosenberg and Hirschberg, 2006). Interestingly, although earlier techniques (Ross and Ostendorf, 1994; Dusterhoff et al., 1999) employed Hidden Markov Models, they did not ex- plicitly model these coarticulatory effects, and re- cent approaches have primarily employed local clas- sifiers, such as decision trees (Sun, 2002; Rosenberg and Hirschberg, 2006) or Support Vector Machines (Levow, 2005). Another body of work on pitch accent recog- nition has focused on exploitation of lexical and syntactic information to predict ToBI labels, for example for speech synthesis. These approaches explored a range of machine learning techniques from local classifiers such as decision trees (Sun, 2002) and RIPPER (Pan and McKeown, 1998) to se- quence models such as Conditional Random Fields 217 (CRFs)(Gregory and Altun, 2004) more …

Content truncated to 500 words