Korean Sentence Symbol Preprocess System for the Improvement of Speech Synthesis Quality

Abstract

In this paper, we propose a Korean sentence symbol preprocessor for a SSML (speech synthesis markup language) supported speech synthesis system in order to improve the quality of the synthesized result. After the analysis of Korean Wikipedia documents, we propose 8 categories for the meaning of sentence symbols and 11 regular expression for the classification of each category. After the development of a Korean sentence symbol preprocess system we archived 56% of precision and 71.45% of recall ratio for 63,000 sentences.▸Keywords : sentence symbol processing, speech synthesis, quality improvement, Korean sentence symbols, preprocessing•제1저자 : 이호준 •교신저자 : 이호준

A comparative analysis of English nuclear stress principles in conversation

“…The following conversational data set comes from a recording of a Midwestern American family's dinnertime conversations in their home 6 (from Lee, 2013). The recording was made discreetly (with the researcher out of view, and a camera and microphone placed inconspicuously near their dinner table).…”

“…Some like Zubizarreta (1998) and Tajsner (2008) in the minimalist framework propose fairly complex and even unnatural derivations for stress placement, which lack empirical evidence for their proposals. A major shortcoming of derivational generative accounts, especially the more complex ones, is that they simply take the assigned focus as a given, and fail to consider or include context, speaker intentions, and communicative intent (Lee, 2013;Szwedek, 2017). 4 Due to the lack of empirical testing, they also miss important patterns and are thus unable to explain the primacy of sentence-final nouns in stress assignment (Szwedek, 2017).…”

“…Costa (2001) provides a simple analysis for canonical sentences, involving a constraint that aligns stress with the main focus; this is essentially similar to the Nuclear Stress Rule and other generative proposals. Lee (2013) provides a more detailed OT analysis that attempts to explain the various linguistic domains involved: syntax, lexical classes, prosodic phonology, and focus as a specific feature within information structure. For normal stress, at least nine OT constraints are proposed to align prosodic stress, lexical stress, compound words, primary focus, content and function words, prosodic phrase constituents (namely, intonational phrases, or IPs), informational phrase domains, and syntactic boundaries.…”

“…Some constraint interactions are illustrated below. Two sample constraint evaluation tableaux are provided below, adapted from Lee (2013); the data and methodology of this paper are explained in the data analysis section below. The first shows the relevant constraints for a sentence with normal but non-final focus.…”

“…As this excerpt attests, these few CG principles or OT constraints can account for most or all stress patterns in the conversation. The OT analysis adopted here treats Contrast as the more general category for special stress (Lee, 2013), while the CG approach here prefers Special as the more general designation. Otherwise, the relevant OT constraints and CG principles generally seem similar, and often one may seem like a notational variant of the other, and seemingly without offering any advantages over the other.…”