To: Distribution
From: David Kahaner ONRFE [kahaner@xroads.cc.u-tokyo.ac.jp]
Re: Studies on a Vocal Tract Model for Speech Synthesis and Analysis
3 April 1990

Researches of the Electrotechnical Laboratory, Number 905, December 1989.
"Studies on a Vocal Tract Model for Speech Synthesis and Analysis"
(126 pages including 124 item bibliography)
by               Hiroshi Ohmura 
                 Speech Processing Section
                 Machine Understanding Division
                 Electrotechnical Laboratory
                 1-1-4 Umezono, Tsukuba-shi
                 Ibaraki 305, JAPAN
                    Phone: 0298-58-5933

In Japanese, with English summary, given below.
Summary: Both the articulatory and acoustical sides of speech research 
are very important for the understanding of speech sound phenomena 
generated by complicated movements of vocal organs. However, the scope of 
our information about these articulatory-acoustical relationships is 
quite limited. Therefore, the development of a vocal tract model 
integrating speech synthesis and analysis is essential to clarify and 
systematize all the relationships. The first purpose of these studies is 
to show that the articulatory data such as vocal tract shape are 
effective in continuous speech synthesis by use of a simplified vocal 
tract model controlled dynamically by a compact set of rules. To get a 
further improvement of synthetic speech, it is necessary to put the 
articulatory information in real speech to practical use. Therefore, the 
second purpose is to propose a new vocal tract model which gives a plane 
of articulatory-acoustical descriptions of speech in more general 
conditions. Through the synthetic usage of this model, we wish to make a 
contribution to describing a complete system existing in the relations 
between these two realms. This monograph consists of an introduction and 
six chapters.  

Chapter 2 deals with speech synthesis by rules using a dynamic vocal 
tract simulator, which is a kind of analog computer system. The control 
data for the simulator are generated from input sentences by a 
programming system. This system consists of two parts of articulatory and 
prosodic rules. The articulatory data such as area functions and 
durations are stored in a table with all of the intermediate symbols for 
each phoneme. Through the speech synthesis experiments of Japanese and 
English fairy tales, it was certified that the articulatory 
representation of speech functions sufficiently as an effective 
description of the feature.  

Chapter 3 deals with actual states in speech wave analysis for extraction 
of articulatory information. This chapter consists of three parts. The 
first part describes an adaptive inverse filtering method of vocal tract 
area function estimation. The assumptions of the vocal tract model are as 
follows; lossless tube section, zero impedance termination at the lips, 
pure resistance termination at the glottis, and sound source located at 
the end of the tube sections. With these assumptions, the adaptive area 
function estimation method has been developed. The merit of this method 
is the almost realtime estimation of area functions without using any 
transcendental information about formants and vocal tract length.  
However, the applicaiton of the model is limited to vowels and vowel-like 
sounds because of its all-pole transfer function. The second part 
describes an iterative method for consonantal area function estimation 
based on the extended vocal tract model of which the sound source can be 
located at any section inside the vocal tract. Now, its transfer function 
becomes a pole-zero function. As a result of the area function estimation 
for voiceless stops, a problem of modeling for sound source 
characteristics of consonants has been left to appropriate estimation of 
those locations which conform to the articulatory points. In the third 
part we discuss acoustic correlates regarding the manner of articulation 
of dentals in Japanese. These acoustic quantities are also important, in 
the same way as vocal tract shapes for specification of speech sound 
features.  

Chapter 4 deals with a new vocal tract model constructed under 
generalized acoustic conditions and recursive calculation methods of 
transfer functions for typical articulations such as voiceless stops, 
nasals, and liquids. A digital circuit of this model is presented by 
reflection coefficients and propagation constants both of which have some 
frequency characteristics. This model can by used in speech synthesis and 
analysis fields.  

Chapter 5 deals with a speech production model including the vocal tract 
model described in the previous chapter and introduces a basic 
application algorithm of a generalized vocal tract ARMA model for 
consonantal area function extimation. The speech production model 
consists of the sound source frequency characteristic, vocal tract 
transfer characteristic and the transfer characteristic from the lip-end 
to an external point.

Chapter 6 deals with the computation of the nomograms on systematic 
articulatory variations and their power spectra as a primary application 
of the new vocal tract model. Using those nomograms for articulatory 
interpretations of spectra for vowels and several consonants in real 
speech, we may conclude that this model is effective in the elucidation 
of articulatory-spectral relationships in real speech.

Chapter 7 in the conclusion of this report. The articulatory-spectral 
nomograms are expected to be very useful in adding to our stock of 
knowledge regarding a view of internal acoustic phenomena and a norm 
which separates consonantal vocal tract characteristics from sound source 
characteristics. For the future, it is necessary to introduce functional
elements of articulation in the presented vocal tract model. Usage of 
nomographic information for many different situations in real speech 
makes it possible to find a new approach for consonantal area function 
estimation, and also to develop a new method of feature extraction.

The report is in Japanese, but the author speaks reasonably good English.  
He was not familiar with the Office of Naval Research, and was initially 
reluctant to discuss his work with the "military". This reaction has 
occurred before, but is usually a minor problem that solves itself when 
the type of research that ONR supports is explained.  Nevertheless, 
American researchers (especially those from national labs) who wish to 
communicate with Japanese scientists should be sensitive. 

-------- END OF MEMO--------------------------------------------------
ustical de