Natural language research announcements

[Margaret Fleck <mfleck AT cs.uiuc.edu>]

-------- Original Message --------
Subject: Linguistics Seminar (next week/G17)
Date: Fri, 26 Jan 2007 10:23:59 -0600
From: Eunah Kim 
<ekim39 AT uiuc.edu>
Reply-To: 
ekim39 AT uiuc.edu
To: 
LING-DEPT-L AT LISTSERV.UIUC.EDU

Dear all,

Here is the information on the next seminar. Please note that the talk will be presented in G17, not 
in Lucy Ellis.
-------------------------------------------

Presenter: Tae-Jin Yoon (Ph.D. Candidate of the Linguistics Dept.)

Time: Thursday, February 1, 2007, 4PM,

Place: G17, FLB

Title:
Modeling Phonetics and Phonology for Speech Technology: Dual Advances in Linguistic Science and 
Engineering

Abstract:
In this talk, I present two distinct, but related research goals: One is improving automatic speech 
recognition using phonetic knowledge, and the other is finding phonetic encoding of linguistic 
structure with the aid of automatic speech recognition.

In the first part of the talk, I demonstrate that phonetic knowledge can be used in improving the 
performance of automatic speech recognition. Specifically, I show that spectral and temporal 
measurements of speech sounds that are related to voice quality can be used in improving word 
recognition accuracy in automatic speech recognition system. The spectral measurement of H1-H2 and 
the temporal measurement of the mean autocorrelation ratio are used to objectively label voice 
quality categories on each sonorant phone on a subset of Switchboard corpus (Godfrey et al. 1992). 
Results from a Support Vector Machine (SVM) classification experiment show that these features are 
predictive of Perceptual Linear Predictive Cepstra (PLPC, Hermansky 1990) that are used as an input 
to speech recognition. I further demonstrate that by incorporating voice quality knowledge into a 
speech recognition system, we can improve word recognition accuracy. Acoustic phonetic models of 
voice quality, th!
 us!
, lead to improved speech technology.

In the second part of the talk, I illustrate that tools from speech technology, in turn, can be used 
for linguistic analysis, enabling the analysis of large-scale databases, and incorporating 
laboratory and non-laboratory recordings. In particular, the phonetic encoding of prosodic structure 
is investigated through a study of the acoustic correlates of prosodic boundary and their 
interaction with accent at three levels of prosodic phrasing, with the three phrasing levels being 
word, intermediate phrase, and intonational phrase (see Ladd 1996). Acoustic cues to prosodic 
boundaries are observed in the lengthening of segments in the preboundary syllable rhyme, with 
greater effects of lengthening at successively higher levels of prosodic domains (Wightman et al. 
1992). A second dimension of prosodic structure is the encoding of prominence, which also gives rise 
to lengthening effects in the prominent syllable (stressed or accented) (Turk and Sawusch 1997). 
Given two distinct sou!
 rc!
es of lengthening, the question arises whether lengthening on its own can serve as a cue to either 
prosodic context. For the analysis, we obtained a large amount of phone-aligned data from Boston 
University Radio News Corpus (Ostendorf et al. 1995) using forced alignment based on HMM (Hidden 
Markov Model). I present results showing effects of prosodic boundary and accent at the three 
prosodic levels in measures of duration local to the domain-final rhyme.

In conclusion, phonetics and phonology provide an important source of knowledge for the design of 
speech technologies, and speech technologies provide useful and crucial aids for the investigation 
of complex linguistic phenomena based on diverse and large-scale speech databases.

Eunah