Automatic Distinction between the Words “Start” and “Stop”

B. N. Epifantsev, N. F. Kucher, A. N. Osokin, V. M. Razin

(Submitted by the Department of Computer Engineering)

The problem of objective recognition of auditory patterns has recently attracted great interest. The question is the possibility of perceiving—or, more precisely, recognizing—speech by special devices, speech that until recently was perceived only by humans [1]. The tendency toward automation of various kinds of processes is quite naturally beginning to penetrate this area of human activity as well.

It has already been noted [2] that solving the problem of constructing devices capable of recognizing connected speech (individual words, phrases) is, in the general case, made extremely complicated by differences in the degree of clarity and distinctness of pronunciation. As Academician L. V. Shcherba points out [3], an infinite number of transitional degrees in pronunciation is possible, ranging from absolute clarity and precision to careless patter, when all unstressed syllables are half “swallowed.”

As is known [2], at present sufficiently clear invariant features ensuring good recognition of these sounds have not been found for most phonemes. Therefore the nearest task in the problem of voice control is considered [6, 7] to be the development and use of devices recognizing a limited set of phonemes and their combinations, which in essence reduces to recognition of a limited set of words. According to the terminology accepted in the literature [2], recognition of a limited set is understood as the identification of a phoneme or word included in the set, while the device does not respond to any other sounds. In distinguishing a limited set, correct identification is carried out only when words or phonemes of the set are pronounced. To any other sounds the device either does not respond or outputs some word of the set.

A device that distinguishes a limited set of words proves to be simplest and most reliable when so-called “differential” features of speech signals are used for recognition, such as the presence or absence of a fundamental tone, noise, closures, etc. [4, 5, 6].

We carried out a search for features that are simple to analyze by circuit methods and that do not depend on the quality of pronunciation or on a change of speakers. From published sources [4] the following is known:

a) when the sound “S” is pronounced—a noisy, front-lingual sound—there is no fundamental tone regardless of who pronounces this sound, a man or a woman;

b) the sounds “P,” “T,” “K” are voiceless consonants, plosives, pronounced without participation of the vocal cords;

c) the sounds “O,” “U” are vowels, labial; the main source in pronouncing these sounds is the vibrating vocal cords, i.e., when “O,” “U” are pronounced the fundamental tone is present.

Analysis of oscillograms of the words “start” and “stop,” pronounced by different announcers, showed the following:

1. In both words under study there is a voiced segment.
2. From the beginning of the speech signal to the appearance of the fundamental tone, which corresponds to the beginning of the voiced segment in the words “start” and “stop,” there is a certain time interval.
3. The time difference from the beginning of the speech signal to the appearance of the fundamental tone in the word “start” is always smaller than the analogous difference in the word “stop.”

On the basis of the foregoing, the following differential features of the words “start” and “stop,” invariant with respect to the announcer, were selected for distinction:

1. Presence or absence of a voiced segment in the word.
2. Presence or absence of a noise segment at the beginning of the word.
3. Magnitude of the time interval from the beginning of the word to the beginning of the voiced segment.

The authors developed an algorithm for distinguishing the Russian words “start” and “stop” and constructed a device that automatically distinguishes these words; its functional diagram is shown in Fig. 1.

Labels in Fig. 1: speech signal; fundamental-tone signal; Amplifier I; Amplifier II; Peak detector I; Peak detector II; Schmitt trigger I; Schmitt trigger II; Monostable multivibrator III; Monostable multivibrator II; Monostable multivibrator I; Coincidence circuit II; Coincidence circuit I; Trigger I; Monostable multivibrator IV; Trigger II; Indication; “stop” pulse; Potential; “start” pulse.

Fig. 1. Functional diagram of a device distinguishing the words “start” and “stop”

In order to simplify the apparatus, it is proposed to pick up the fundamental-tone signal by means of a laryngophone placed on a certain part of the announcer’s neck. In order to increase noise immunity against extraneous sounds, which is especially important when the device operates in a real environment, and to ensure uniformity of the elements used, it is proposed to pick up the speech signal from another laryngophone, also placed on another part of the neck.

The operation of the device proceeds as follows: The speech signal is amplified by amplifier 1. Peak detector 1 extracts the envelope of the speech signal, from which the Schmitt trigger forms a rectangular pulse equal in duration to the word pronounced. The leading edge of this pulse starts monostable multivibrator 1, which generates a pulse of duration \(T_1\), somewhat longer than the duration of the plosive “P” at the beginning of the word “start.” The trailing edge of the pulse from monostable multivibrator 1 starts monostable multivibrator 2, which generates a pulse of duration \(T_2\). The trailing edge of this pulse starts monostable multivibrator 3, which generates a pulse of duration \(T_3\), with the sum \(T_1 + T_2 + T_3\) having to be less than the duration of the segment “ST” in the word “stop.” The trailing edge of this pulse starts monostable multivibrator 4, which generates a pulse of duration \(T_4\).

The fundamental-tone signal is amplified by amplifier 2. Peak detector 2 extracts the envelope of this signal, and Schmitt trigger 2 forms from it pulses that arise when a voiced segment is present in the word and that are equal in duration to the voiced segments of the words “start” and “stop.”

When the word “start” is pronounced, the beginning of the plosive “P” corresponds to the leading edge of the pulses of Schmitt trigger 1, whose pulses are fed to monostable multivibrator 1. Monostable multivibrator 1 generates a pulse that starts monostable multivibrator 2. The pulse of monostable multivibrator 2 is fed to coincidence circuit 1, to the second input of which the pulse of Schmitt trigger 2 is applied. If the word “start” has been pronounced, then the pulse of monostable multivibrator 2 coincides in time with the voiced segment, and at the output of coincidence circuit 1 a pulse appears corresponding to the pronounced word “start,” which sets trigger 2 to the one state; the indicator lamp “start” lights up. The pulse from the output of coincidence circuit 1 sets trigger 1 to the “0” state, and coincidence circuit 2 is closed. When the word “stop” is pronounced, the beginning of the noise “S” corresponds to the leading edge of the pulse of Schmitt trigger 1, which starts monostable multivibrator 1. The pulse of this monostable multivibrator sets trigger 1 to the “1” state, and coincidence circuit 2 is prepared for further operation. This same pulse starts monostable multivibrator 2, which in turn starts monostable multivibrator 3, which starts monostable multivibrator 4. Since \(T_1 + T_2 + T_3\) is less than the duration of the segment “ST” in the word “stop,” there will be no pulse at the output of coincidence circuit 1, because the pulse at the output of Schmitt trigger 2 appears only after the end of the time interval \(T_1 + T_2 + T_3\).

Coincidence of the pulses of monostable multivibrator 4 and Schmitt trigger 2 in coincidence circuit 2 occurs during the sounding of the voiced “O” in the word “stop”; at the output of coincidence circuit 2 a pulse appears, which corresponds to the presence of the pronounced word “stop.” From the output of coincidence circuit 2 the signal is applied to the zero input of trigger 2; at its one output the potential will be close to zero, while the potential taken from the zero output will light the “stop” indicator lamp.

The pulse durations of the monostable multivibrators were selected experimentally and are respectively: \(T_1 = 15\) msec, \(T_2 = 30\) msec, \(T_3 = 60\) msec, \(T_4 = 15\) msec.

To evaluate the operability of the device when it was used by different operators, the following tests were carried out. A team of speakers was chosen arbitrarily (5 men and 5 women). They pronounced the word “start” 30 times, then the word “stop” 30 times, and then “start-stop” 30 times. The result of the distinction was recorded by the lighting up—

...indicator lamps and by the pulse outputs of the device with the aid of a dual-beam oscilloscope.

The results of the tests indicate 100% distinguishability of the indicated words by the device.

The use of a two-channel circuit for the device, as well as the use of monostable multivibrators 1 and 3 for the purpose of delay, makes it possible to obtain increased noise immunity in the operation of the device, since the device can operate only when a signal is present from both laryngophones, which is possible only when voiced sounds are pronounced, as occur in the words “start” and “stop”; and the delay produced by monostable multivibrators 1 and 3 makes it possible to eliminate false triggering of the device by short, sharp sounds, for example coughing. The principle underlying the operation of the device, as already noted, is based on the properties of the structure of the words “start” and “stop,” which is invariant with respect to the speaker. Therefore, the result of the distinction does not depend on whether a man or a woman, one person or several people, use the device, since this device does not take the individuality of the voice into account.

REFERENCES

1. A. A. Kharkevich. Pattern Recognition. Radiotekhnika, No. 5, 1959.

2. G. I. Tsemel. Automatic distinction of an organic set of words. Izvestiya AN SSSR, OTN, “Energetika i avtomatika”, No. 5, 1960.

3. L. V. Shcherba. Selected Works on the Russian Language. Uchpedgiz, 1957.

4. E. T. Semenova, I. M. Tetelbaum. Recognition features for a limited set of words. In: Problems of Bionics, edited by M. G. Gaaze-Rapoport, Nauka, 1967.

5. L. A. Chistovich et al. Speech, Articulation, and Perception. Nauka, 1965.

6. E. N. Myasnikova. Objective Recognition of Speech Sounds. Energiya, 1967.

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