ON THE POSSIBILITY OF USING “ARTIFICIAL LARYNX” DEVICES FOR SPEECH RESEARCH

N. F. KUCHER, V. M. RAZIN

(Presented by the Department of Computer Engineering)

“Artificial larynx” (AL) devices were developed as prostheses for people who, for one reason or another, have lost their vocal cords—for example, in connection with a surgical operation. They enable such people to restore the possibility of communication by means of speech communication.

At present there are several varieties of AL devices, which are divided into two classes: vibrational and sound-emitting.

The basic element of vibrational devices is a mechanical vibrator, oscillating at a specified fundamental-frequency pitch, which is applied to a certain area of the neck and in this way excites the speech-forming tract.

In sound-emitting devices, fundamental-frequency oscillations are generated in the form of pulsed acoustic pressure, which is created by a special sound-emitting element of the electrodynamic or electromagnetic telephone type, connected at the output of a pulse generator, and are introduced through a special sound conduit into the speaker’s oral cavity.

Artificial speech produced with the aid of AL devices approaches natural speech in quality and loudness.

In the Soviet Union, only one voice-producing device of the vibrational type is mass-produced—the AG-61, built on the buzzer principle. Despite the relative simplicity of the device, the speech obtained with its aid is of good quality.

Among foreign devices of the vibrational type, the device of the Wescon firm (USA), described in [1], is known. The vibrator in this device is driven by an electronic rectangular-oscillation generator implemented with transistors.

At present the Soviet Union does not mass-produce sound-emitting devices. However, AL devices of the “Golos-15,” “Golos-18,” and “Golos-21” types have been developed and prepared for mass production; they are characterized by the following data:

1. Fundamental frequency — 80–150 Hz.
2. Duration of the fundamental-frequency pulse — 0.3–0.6 msec.
3. Average loudness level — 70 dB at a distance of 1 m.

In addition, the “Golos-21” device makes it possible to generate noise sounds with an average loudness level of 50–60 dB.

Let us consider the process of speech formation with the aid of “artificial larynx” devices. According to the acoustic theory of speech formation, most fully developed by G. Fant [2], the process of forming speech sounds proceeds as follows (Fig. 1 serves to illustrate this process):

Fig. 1. Block diagram illustrating the process of formation of speech sounds

Diagram labels: constant air-flow source — lungs; converter of constant air flow into variable flow — vocal folds; \(S_0(\omega)\); linear passive filter — speech-forming tract; \(S(\omega)\); speech; lips.

The lungs send a constant air flow to the vocal folds, which, depending on the type of sound, transform this flow either into nearly periodic oscillations or into white noise. The oscillations emitted by the vocal folds have the spectrum \(S_0(\omega)\). These oscillations then enter the speech-forming tract, which is a linear passive four-terminal network having the frequency transfer function \(K(\omega)\). At the output of the speech-forming tract, which ends with the lips, a speech signal arises whose spectrum \(S(\omega)\), according to the theory of linear systems, is equal to

\[ S(\omega)=S_0(\omega)\cdot K(\omega). \tag{1} \]

The process of forming speech sounds with the aid of AL devices proceeds analogously to that described above (Fig. 2).

The oscillations of the “artificial larynx,” with spectrum \(S'_0(\omega)\), act directly on the speech-forming tract with transfer function \(K(\omega)\), at whose output an “artificial” speech signal is formed with spectrum \(S'(\omega)\), equal to

\[ S'(\omega)=S'_0(\omega)\cdot K(\omega). \tag{2} \]

The artificial speech signal has good intelligibility and loudness, almost not inferior to natural speech, and, moreover, possesses two remarkable properties: constancy of the fundamental-frequency pitch and constancy of the spectrum \(S'_0(\omega)\) of the excitation signal of the speech-forming tract. These properties are extremely important in problems of recognition, synthesis, and compression of speech messages.

Fig. 2. Formation of “artificial speech”

Diagram labels: artificial larynx; \(S'_0(\omega)\); speech-forming tract; \(S'(\omega)\); speech.

Thus, pronouncing speech with the aid of an AL device ensures normalization of the speech signal of different speakers with respect to the fundamental-frequency pitch and the excitation spectrum, and makes it possible, in analysis, to dispense with the complex device for extracting the fundamental frequency, since the fundamental-frequency signal can be taken directly from the generator included in the AL.

Among the wide range of problems associated with the analysis and synthesis of speech messages, the problems of determining the transfer functions of the human speech-forming tract and determining the excitation signal (fundamental-frequency pulses) stand out. The use of AL devices makes it possible to solve these problems.

Consider expression (2). The transfer function of the speech-producing tract \(K(\omega)\) is determined from it as follows:

\[ K(\omega)=\frac{S'(\omega)}{S'_0(\omega)}, \tag{3} \]

that is, in order to determine the unknown transfer function \(K(\omega)\), it is sufficient to know \(S'(\omega)\), the spectrum of the “artificial” speech signal, and \(S'_0(\omega)\), the spectrum of the oscillations of the AL device, whose determination by known methods presents no difficulty, since, unlike in natural speech, it is possible to measure the spectrum \(S'_0(\omega)\) of the excitation signal at the output of the sound duct of the AL.

Another method for measuring the transfer function of the speech-producing tract is based on a known property of linear systems, namely that the impulse transient function \(g(t)\) of a linear system (the weighting function) and the frequency transfer function \(K(\omega)\) are related to one another by the Fourier transform

\[ K(\omega)=\int_{0}^{\infty} g(t)e^{-j\omega t}\,dt. \tag{4} \]

It is assumed that, within one period \(T\) of the fundamental tone, the transient process (speech signal) in the speech-producing tract, when acted upon by an excitation pulse (of the fundamental tone), ceases. Therefore, if with the aid of the AL a sufficiently short excitation pulse is formed (for a fundamental-tone period \(T=10\) msec it is sufficient to choose an excitation-signal duration equal to 100–150 µsec), the resulting “artificial” speech signal may, with a sufficient degree of accuracy, be regarded as the weighting function of the speech-producing tract, from which, with the aid of expression (4), the desired function \(K(\omega)\) is easily obtained.

After determining \(K(\omega)\), from expression (1) one can determine the spectrum \(S_0(\omega)\) of the function of the natural human vocal cords by the formula

\[ S_0(\omega)=\frac{S(\omega)}{K(\omega)}. \tag{5} \]

With the aid of the inverse Fourier transform, the time function of excitation \(U(t)\) is determined from the spectrum \(S_0(\omega)\).

REFERENCES

1. H. Barney. Unitary Transistored Artificial Larynx. IRE Wescon Convention Record, Part 8. August, 1959.
2. G. Fant. Acoustic Theory of Speech Production. “Nauka,” 1964.