Amalia Todiraºcu *
A Unification-Based Model for Speech Generation
The attributes which are used
to describe the type phoneme correspond to each phoneme
feature. The features were chosen to be the most general ones
in describing the phonemes. According to the theory of distinctive
features [8] (due to Trubetzkoy),
the articulatory and acoustic
properties of sounds can be described by using bilateral or multilateral
oppositions. According to Jacobson, sounds can be classified by
using binary oppositions, like: vowel - non-vowel, consonant -
non-consonant, compact - diffuse, nasal - oral, continuant - interrupted,
supralaryngeal - laryngeal, voiced - voiceless, lenis - fortis,
strident - mat, constricted - non-constricted.
Example:
Consonants which are dental and labial are ANTERIOR. CORONAL sounds
are produced with the blade of the tongue raised from its neutral
position (like dentals, alveolars, alveolopalatals).
The features are classified in
two groups, depending on the parts of the speech organs which
influence them - LARYNGEAL or SUPRALARYNGEAL. The LARYNGEAL features
are: SPREAD, CONSTRICTED, and VOICED. The SUPRALARYNGEAL features
are: the manner of pronunciation (MANNER, with the attributes
NASAL, CONTINUANT, and STRIDENT) and the place of articulation
(PLACE, with the attributes CORONAL, ANTERIOR, and DISTRIBUTED).
The attribute TYPE gives the type of the phoneme (VOWEL, CONSONANT,
or SEMIVOWEL). The attribute CORRESP gives an entry to the effective
sound, which is recorded separately.
This representation will be used
for the transformation of the text into an orthographic string.
Some context-dependencies rules are applied in order to obtain
the final orthographic string. The formalism can be applied in
a system for speech synthesis, but the result of this process
will be a nonnatural utterance. A more interesting approach can
be to store word sound or syllable sound, in order to obtain natural
speech.
3. Hyphenation
The third part presents an application
of the model for hyphenation. Syllables are important in the pronunciation
of the word, the alternation between stressed and unstressed syllables
giving the correct intonation and rhythm for the word or of the
phrase. Hyphenation is treated in a constraint-satisfactory approach,
rather than in a generative one. The value of type phon
can be represented as follows
where C1,...,Ck
are constraints to be applied to the phonological string. The
constraints are also described as feature structures.
The model is applied as an example
for the Romanian language. The word itself is described by a structure
of type foot and must be divided into a list of syllables.
The list will satisfy the set of constraints imposed by the hyphenation
rules and the rules for combining phonemes. For the Romanian language,
there are some examples of the rules to be applied when it is
necessary to hyphenate a word:
Examples
of the rules to be applied when hyphenating a word:
- VCV --> V-CV (ex. e-lev; re-vol-ta);
- VCCV --> VC-CV (fap-ta);
- VC{l, r}V --> V-C{l, r}V, where C is one of the consonants
b, c, d, g, p, t, v, f (ex. ta-bla);
- VrlV --> Vr-lV (mier-la).
For English, there are also some examples of hyphenation rules:
- VCS --> VC-S (ex. a-ble; no-tion), where C is a
voiced consonant and S is a syllabic one;
- VCCV --> VC-CV (ex. in-vite; kip-ling);
- VCCCCCV --> VCC-CCCV (ex. bubble-spray).
Rules for combining phonemes inside
a syllable for the Romanian language:
- a syllable may contain only
one open vowel (e, a);
- a vowel combines with one
or more semivowels in order to build a syllable (iau);
- a vowel combines with one
or more consonants in order to form a syllable (vor);
- the labials combine with 'r'
(brumã, prin, frig), and the strident consonants
need a previous 's' (scad, spate).
169
