Marian Boldea *
Speech Technology
Research at Computer Science Department, "Politehnica"
University of Timiºoara
Pronunciation dictionaries are
useful for both applications development and more elaborate speech
database designs, and their creation has to be supported by appropriate
procedures and dedicated tools. In our laboratory, an ASCII computer-compatible
phonetic transcription system was defined for Romanian starting
from the SAMPA phonetic alphabet [6,7];
using these conventions,
graphemes-to-phonemes transcription rules for a text-to-speech
system [8] were iteratively refined on a
lexicon of about 24,000
words, and the corresponding pronunciation dictionary was obtained
as an outcome of this process. To facilitate its manipulation
and future evolutions, an ad hoc system was built
[9] and its
utility was already proved during the design of some components
of a larger Romanian speech database [10].
The first of these is a database for concatenate
speech synthesis, designed to cover all practically possible Romanian
diphones [9] by taking into account phonetic
phenomena that occur
in connected speech and already used in our Romanian text-to-speech
system presented in the next section. As an evaluation companion
to this, speech material to be used as bench mark data in the
administration of a Romanian version of the Modified Rhyme Test
[11] was also recorded by the same speaker.
The second, and most important,
is a database of Romanian continuous speech, whose initially envisaged
use was in the training, development and evaluation of continuous
speech recognizers at the phoneme and word levels using speaker-independent
phoneme models with some context-dependence modelling capability.
During its design phase, the need for acoustic phonetics and phonology
studies appeared also very clearly to us, and we hope that it
will be used for this purpose, too. It is intended to be compatible
with the EUROM-1 speech database [12]
and will consist of high-quality
recordings from 100 speakers (50 men and 50 women), with equal
distributions for each sex in the age ranges: under 20; 21-30;
31-40; 41-50; over 50), every speaker recording mainly some read
material (four passages of a few logically connected sentences;
two or three independent sentences designed to increase the number
of occurrences of the least frequent phonemes; three to seven
independent sentences automatically selected to increase the number
of read diphones; four independent sentences, each designed to
determine - when correctly read - at least one occurrence of every
Romanian phoneme; 26 integer numbers between 0 and 9999; the Romanian
alphabet) as well as some semispontaneously spoken personal information
(name - spoken and spelled; series and number of the personal
identification document; telephone number; birth date; address).
More details on its design, collection and processing can be found
in [10]. Sixty-seven speakers
(38 men and 29 women) have been
recorded so far2, and work is in progress towards the automatic
labelling of the signal and its use for continuous speech recognition
experiments (see Section 4).
3. Text-to-speech conversion
Given the wide area of applications that can use it, the characteristics
of the human auditory system that allow synthetic speech to be
understood even if it does not sound naturally, and its being
necessary for realistic answer back capabilities during data collection
for spoken dialogue systems development, speech synthesis from
written text has been lately given special attention in our group.
Converting texts into synthetic speech encompasses,
on the one side, a natural language processing (NLP) module able
to transform the input text into an appropriate intermediate representation,
and on the other side, a signal processing (SP) part capable of
turning this representation into an output signal. Ideally, the
NLP module should "understand" the input texts the same
way a perfectly trained human reader would do, in order to produce
not merely phonemic symbols, but also information about prosodic
(intonation, duration, intensity) and phonetic (coarticulation)
phenomena, and the SP component has to apply its own phonetic
rules to generate an output signal as naturally sounding as possible.
Even with human readers though, the way a text is read can vary
(e.g., poetic or dramatic texts), so that contemporary text-to-speech
systems are quite limited as far as the incorporation of appropriate
linguistic knowledge is concerned. Consequently, a typical structure
was chosen for our system, as illustrated in Fig. 1.
The first four blocks in the diagram
[8,9] are the
constituents of the NLP component, while the last two are actually
collapsed into a single overlap-and-add concatenate synthetiser
which uses the database presented in Section 1 and allows for
phonetic phenomena to be accounted for through diphones segmentation
[9,13].
The text normalisation stage replaces numbers and abbreviations
with full orthographic forms, the spell checking tries to spot
mistyped words in order to prevent failures of the graphemes-to-phonemes
translation, and all these first three stages use rules and dictionaries
to implement their tasks [8].
The prosodic component [9] makes
three passes (prosodic parsing and phoneme duration assignment,
duration adjustments according to the prosodic structure, and
fundamental frequency contour generation) over the representation
it receives, and transmits to the synthesis module the appropriate
control parameters.
Experimental intelligibility evaluations of the system
using a Modified Rhyme Test [11] adapted to
Romanian [9] and natural
bench mark signals recorded by the same speaker used for the diphones
database, although limited to 6 subjects, demonstrated a good
performance (90.89% average intelligibility for synthetic speech
vs. 98.21% for natural speech). No attempt has been made to evaluate
the system from the viewpoint of the naturalness of the synthetised
speech.
2 January 20, 1997.
175
