Emphasized Accent Phrase Prediction from Text for Advertisement Text-To-Speech Synthesis

PACLIC 28 !170 Emphasized Accent Phrase Prediction from Text for Advertisement Text-To-Speech Synthesis Hideharu Nakajima, Hideyuki Mizuno, and Sumitaka Sakauchi NTT Media Intelligence Labs., NTT Corporation, 1-1 Hikarino-oka, Yokosuka, Kanagawa, 239-0847 JAPAN Abstract Realizing expressive text-to-speech synthesis needs both text processing and the rendering of natural expressive speech. This paper fo- cuses on the former as a front-end task in the production of synthetic speech, and in- vestigates a novel method for predicting em- phasized accent phrases from advertisement text information. For this purpose, we exam- ine features that can be accurately extracted by text processing based on current Text-to- speech synthesis technologies. Among fea- tures, the word surface string of the main con- tent and function words and the part-of-speech of main function words in an accent phrase are found to have higher potential on predict- ing whether the accent phrase should be em- phasized or not through the calculation of mu- tual information between emphasis label and features of Japanese advertisement sentences. Experiments confirm that emphasized accent phrase prediction using support vector ma- chine (SVM) offers encouraging accuracies for the system which requires emphasized ac- cent phrase locations as context information to improve speech synthesis qualities. 1 Introduction The introduction of corpus-based speech synthesis methods such as unit selection synthesis ((Hunt, et al., 1996) etc.) and Hidden Markov Model speech synthesis ((Zen, et al., 2009) etc.) makes expres- sive speech synthesis possible if an adequate speech database is prepared. However, the synthesized speech often fails to recreate emphasis or phrase boundary tone, even though both are key character- istics of expressive speech. The location markers of emphasis and phrase boundary tone have been con- firmed useful in improving expressive speech syn- thesis qualities; they form part of the context in- formation for speech synthesis (Meng, et al., 2012; Maeno, et al., 2014; Strom, et al., 2007; Yu, et al., 2010). For establishing Text-To-Speech (TTS) synthesis for expressive speech, it is necessary to predict lo- cations of emphasis and phrase boundary tone from the input text. The phrase boundary tone occurs at the phrase end, and existence/non-existence of the tone can be accurately classified, from the text to be synthesized, by using machine learning approaches (Nakajima, et al., 2013; Ross, et al., 1996). Thus, this paper focuses on the remaining target of empha- sis positions. In this work, we use the word “empha- sis (emphasized)” to denote portions that are percep- tually more salient to the listeners in a sentence. In human speech, emphasis can be regrouped at least into four functions based on analysis in con- ventional literatures as (Hovy et. al, 2013; Sridhar et. al, 2008) (bold portions show emphasized words and phrases). 1. expressing linguistic “focus”: (e.g., “ Taro did.” (as an answer to “who did ...?”)) 2. expressing “contrast”: (e.g., “not A but B”) 3. expressing “element of surprise”: (e.g., “I heard he was sick, but he had much energy.”) 4. disambiguating grammatical structure: clarifying parallel and dependency structure (e.g., to dis- tinguish “{old men} …

Content truncated to 500 words