SELECTION OF A FEATURE SYSTEM FOR DISTINGUISHING A LIMITED SET OF WORDS

N. F. KUCHER, A. N. OSOKIN

(Presented by the scientific seminar of the Department of Computer Engineering)

Information-input devices using optical-reading systems, high-speed output devices, high-speed output devices with data representation on CRTs and microfilm, and means that provide simultaneous input and output operations using CRTs in combination with a light pen, plotters, and speech communication with a digital computer are considered highly promising for further solving the problems of operator–machine communication [1].

The problem closest to practical implementation at present is that of distinguishing a limited set of words separated by pauses during pronunciation. It is also possible to distinguish continuous phrases by treating them as a long word.

More than 30 devices are known that distinguish a limited set of words. Unfortunately, none of them has found application, since none satisfies the specified requirements. The reason for the failures is due to the problem of distinguishing features, which must be invariant to the speaker, independent of the tempo of pronunciation, reliably and simply identified by hardware methods, and characterize the phonetic structure of the pronounced word.

In our work [2] it was proposed, for distinguishing a limited set of words, to use the following features:

1) voiceless stops — $C_{\mathrm{Mг}}$;
2) voiced stops — $C_{\mathrm{Mз}}$;
3) voiceless consonants — $C$;
4) voiced consonants and vowels — $Г$.

These features characterize, to a certain degree, the process of speech formation, which makes it possible to use simple circuits for their identification. To form these features one may use the following signals characterizing the process of speech formation:

1. The threshold-amplified (clipped) envelope of the fundamental-tone signals (vibrations of the vocal cords during pronunciation of words) — $x(t)$. This envelope will be equal to zero when a word is pronounced in the portion of voiceless consonants, plosive and fricative phonemes, namely: s, sh, f, kh, t, k, p, ch, ts. When all voiced phonemes are pronounced, the vocal cords necessarily vibrate. Therefore the envelope of the fundamental-tone signal will be different from zero.

2. The clipped amplitude envelope $A(t)$ of the speech signal $S(t)$, represented in the form $S(t)=A(t)\cos\varphi(t)$ — $y(t)$. Clip-

normalized amplitude envelope of the speech signal will be equal to zero before the pronunciation of the phonemes p, t, k, ch, ts (voiceless stop) and the phonemes b, d, g (voiced stop). It should be noted that a voiced stop at the beginning of a word is characterized by the absence of a speech signal and the presence of laryngeal vibrations [3].

1. The envelope of the whole word, \(z(t)\). The use of this signal makes it possible to increase the reliability of feature extraction.

The indicated envelopes are shown in Fig. 3 of [4].

In the adopted notation for the initial extracted signals, the features may be written as follows:

\[ \text{voiceless stops: } C_{\mathrm{mg}} = x(t)\cdot \overline{y(t)}\cdot z(t); \]

\[ \text{voiced stops: } C_{\mathrm{mz}} = x(t)\cdot y(t)\cdot z(t); \]

\[ \text{voiced stops at the beginning of a word: } C_{\mathrm{mz}} = x(t)\cdot \overline{y(t)}\cdot \overline{z(t)}; \]

\[ \text{voiceless consonants: } C = \overline{x(t)}\cdot y(t)\cdot z(t); \]

\[ \text{vowels and voiced consonants: } \Gamma = x(t)\cdot y(t)\cdot z(t). \]

We shall obtain the greatest amount of information about the word pronounced if, in the process of pronouncing the word, the extracted features are used to compose an approximate phonetic structure of the word. For this purpose, when a word is pronounced, it is necessary not only to extract features, but also to record the sequence of their occurrence. We give the phonetic structures of several words under the proposed system of features:

\[ \text{zero} \equiv \Gamma; \]

\[ \text{one} \equiv \Gamma C_{\mathrm{mz}}\Gamma; \]

\[ \text{two} \equiv C_{\mathrm{mz}}\Gamma; \]

\[ \text{nine} \equiv C_{\mathrm{mz}}\Gamma C_{\mathrm{mg}}C; \]

\[ \text{priority} \equiv C\Gamma C_{\mathrm{mg}}C\Gamma C_{\mathrm{mg}}C; \]

\[ \text{comma} \equiv \Gamma C_{\mathrm{mg}}C\Gamma C_{\mathrm{mg}}C\Gamma; \]

\[ \text{photo-light-guide-end} \equiv C_{\mathrm{mz}}\Gamma C_{\mathrm{mg}}C\Gamma C_{\mathrm{mg}}C\Gamma CC_{\mathrm{mz}}\Gamma C_{\mathrm{mg}}\Gamma C_{\mathrm{mg}}C. \]

From the examples given it is clear that some words are characterized by only a single feature. This is explained by the absence of preliminary segmentation of the speech stream into separate phonemes. Thus, since in the word “zero” all sounds are voiced, the voicing feature \(\Gamma\) will be the only one. In order to expand the set of distinguishable words, it is highly desirable to have a device for preliminary segmentation of speech into phonemes (before feature extraction).

Using these features, it is possible to distinguish 25–30 words. If the set of words is selected specially, then up to 50 words can be distinguished. If phrases are pronounced as a set of coarticulated words, then they may be considered as long words. In that case the set of distinguishable phrases is much larger.

References

1. New Directions in the Technology of Information Input and Output (A Survey Based on Foreign Publications). Collection: Means of Computer Technology and Office Equipment. Series TS-2, Issue 2, Moscow, 1972.

2. N. F. Kucher, A. N. Osokin, V. M. Razin. On One Approach to Recognizing a Limited Set of Words. Proceedings of the 1st Final Conference of the Research Institute of Automation and Electromechanics, Publishing House of the Tomsk TsNTI, 1972.

3. M. A. Sapozhkov. The Speech Signal in Cybernetics and Communications. Svyazizdat, 1963.

4. N. F. Kucher, A. N. Osokin. Feature Extraction for Distinguishing a Limited Set of Words (this collection).