EP0347338A2 - Method and apparatus for speech analysis, synthesis and coding - Google Patents

Abstract

A means of analysing and of synthesising speech is described, using the simulation of the acoustic behaviour of a tube divided into segments with variable cross-section. The variations in cross-section of the various segments of a tube allow the generation of phonemes when an air pressure and delivery source is positioned similar to the human vocal cords. By simulation, these phonemes can be generated in the form of electrical signals delivered to a loud-speaker. The invention bears upon the choice of the lengths of the tube segments and relates this choice to the fineness of the approximation which it is desired to make. For an approximation with three formants (the formants are the resonant frequencies of the tube), the tube is divided into eight segments of successive lengths L/10, L/15, 2L/15, 3L/15, 3L/15, 2L/15, L/15 and L/10; L is the total length of the tube. <IMAGE>

Description

The invention relates to speech analysis, synthesis and coding.

The processes of analysis, synthesis and coding of human speech come up against considerable difficulties which are: the great complexity of the frequency spectrum of the sounds emitted, the proximity of the spectra of neighboring phonemes, the multiplicity of the different phonemes used in the same language and a fortiori in different languages and dialects, and especially the multiplicity of the ways in which the sounds are effectively emitted according to the sounds which precede or follow (phenomena of coarticulation).

It is therefore very difficult either to recognize a succession of phonemes emitted at high rate, to reconstruct the words which have been pronounced, or to synthesize sequences of sounds and words which will be effectively recognized with their meaning by those who hear them.

A known method for synthesizing speech consists in using a device simulating the behavior of an acoustic tube with variable section which represents the vocal tract by which human speech is emitted.

The vocal tract, which starts from the so-called vocal cords (which act as a source of excitation at the upstream end of the tube) extends from the larynx to the lips, passing through the pharynx and the oral cavity. It is a duct whose section is not uniform over its length and varies within wide limits (for example 2 cm² in the larynx, from 3 to 7 cm² in the pharynx, from 0 to 15 cm² for the oral cavity, 0 cm² on the lips if the lips are closed, etc).

This vocal tract can be represented as an acoustic tube constituted by a succession of individual sections of constant length whose section at rest has a determined value.

The books of G. FANT, Acoustic Theory of Speech Production, 1960, Mouton and CO, Gravenhage, Pays Bas, and JL FLANAGAN, Speech Analysis Synthesis and Perception, 1972, SPRINGER-VERLAG - New-York, do being of this type of representation in which the vocal tract is cut into successive sections of the order of a centimeter long whose sections can be listed. The production of sounds can be modeled by variations in the cross-sectional areas of the individual sections.

It is therefore possible to produce sounds approaching phonemes of human speech by using a succession of sections of acoustic tubes with a source of air flow at the inlet, this source having characteristics similar to those of human vocal cords, and by varying the sections of the different sections.

Of course, with modern computer signal processing techniques, we will not use a material acoustic tube having sections that can materially vary section by section, but we will use a simulation of the air source and the vocal tract either by analog electrical circuits, or by a computer in which parameters can be varied representing in particular the tube sections, the total length of the tube, the spectrum of the air flow from the source.

The computer will provide an output to a loudspeaker (for speech synthesis) of an electrical signal whose spectrum and the variations of spectrum reproduce as faithfully as possible the spectrum and the variations of spectrum of the sound or successions of sounds that we want to broadcast. For speech analysis, it is a microphone which receives the acoustic message, which converts it into electrical signals, received and processed by the computer, for example after analog / digital conversions so that the computer can work in digital mode . The result of the analysis can be used directly in speech recognition or be coded and transmitted for reconstruction. The coding can be of scalar or vectorial type.

If the principle of the simulation of the vocal tract by a succession of sections of acoustic tubes of variable section is known, it has never been able to be implemented in a satisfactory manner to allow the analysis or the synthesis of continuous speech. Most often some tests are made for example with vowels or consonant-vowel sets; but we are very far from being able to synthesize or recognize rapid successions of sounds as they appear in human speech.

The reason is that automatic control from text is difficult and poorly understood; the acoustic tube must be configured by a large number of factors: there are many sections of tubes, each can undergo variations in section in very large proportions (just pronounce a [o] or a [a] to see the variation in cross-section of the air passage between the lips), and, if the area function is called the curve of the values of cross-sectional areas of the pipe sections along the succession of sections, there is no direct relationship between the area functions of the acoustic tube and the sounds emitted.

Furthermore, the spectra of sounds emitted in human speech are characterized by "formants" (which are successive maxima present in the spectrum: first forming for the frequency lowest resonance quence, second forming, third forming, ...). These formants represent resonances of the vocal tract, resonances which modulate the spectrum of the sound source (the vocal cords) to result in a spectrum modulated at the output of the vocal tract. The vowels for example are characterized by fixed values of the frequencies of formants (that is to say the values of the frequencies of maximum amplitude of the spectrum). Consonants are rather characterized by the relative variations of the frequencies of formants.

But the combination of a sequence of syllables is difficult to formalize in the form of variations in the frequencies of the formants because for an element of the sequence considered the frequencies of the formants depend on the preceding and following sounds (coarticulation phenomenon).

Speech synthesizers called "formants synthesizers" have been produced: they consist in using (or simulating) resonant circuits whose resonant frequency can be controlled individually. By combining several resonance frequencies corresponding to the formant frequencies of a determined vowel, this vowel can be synthesized. By varying the resonant frequencies of the circuits in the same way as the forming frequencies of a consonant vary, this consonant can be reproduced artificially.

In general, knowing the first three formants or their variations over time is a good approximation for analyzing or synthesizing sounds. But one could be satisfied with two formants for a simplified analysis or synthesis, or on the contrary go up to four formants, or even more, for a more elaborate analysis or synthesis.

In the synthesis with formants, one analyzes or one reconstitutes spectra of signals presenting maxima of amplitude for given frequencies, but obviously one does not know how to analyze or reconstitute exactly the whole spectrum and the variations of spectrum which exactly define the constitution of the sound considered. And the problem is of course considerably complicated if, as a result of phenomena of coarticulation between successive vowels and consonants, the spectra and variations in spectrum of the signal mix.

The present invention starts from the remark that one can combine in an entirely original and particularly interesting way the propositions of speech analysis and synthesis using simulation by an acoustic tube with variable section and the knowledge that one has. acquired in formant analysis and synthesis, to achieve extremely effective analysis and synthesis devices. Their effectiveness comes from the fact that they provide a very good representation of sounds while minimizing the number of parameters of representation of these sounds and from the fact that they operate in a mode which seems very similar to the operating mode of human speech.

According to the invention, an apparatus for analyzing, coding or synthesizing speech is proposed using a device for simulating the acoustic behavior of a tube formed by a succession of sections of different and variable sections placed end to end, characterized in that the tube comprises a set of N sections, divided into subsets of successive rows in the following manner: the set of N sections is divided into two subsets of rank 1, the first subset, on the upstream side of the tube, corresponding to a negative sensitivity to variations in section for the first forming and the second to a positive sensitivity, each sub-assembly of rank i being divided in the same way into two sub-assemblies of rank i + 1 if there is a change in sign of the sensitivity of the forming i + 1 in this subset, one of the subsets corresponding to a negative sensitivity for the (i + 1) th forming and the other to a positive sensitivity, each subset of rank (n-1) being e nfin divided into two sections, one of the sections corresponding to a negative sensitivity of the n ^th forming and the other to a positive sensitivity, the sensitivity of the i ^th forming to section variations of a section representing the relative variation of the frequency of the i ^th forming as a function of a section variation of this section; the device having, for analysis or synthesis control parameters, on the one hand the section variations of some of the tube sections thus defined, and on the other hand the total length of the tube; the device receiving signals from a microphone or supplying signals to a loudspeaker depending on whether it functions as a speech analyzer or synthesizer.

What is important is the way in which the acoustic tube is subdivided into successive sections, which is linked to the existence of formants and to the sensitivity of these formants to variations in the local section of the tube.

While in the past the subdivision into sections was either arbitrary or linked to different data, here we propose a very particular subdivision linked to formants and depending on the number of formants with which the analysis or synthesis approximation must be made.

More precisely, we will show that if we want an approximation with two formants, that is to say an approximation similar to that which we obtain in an analysis, coding or synthesis with two formants but obtained by simulation of the behavior d 'a tube with successive sections of variable section, the tube will be divided into four sections of successive relative lengths substantially equal to 1/6, 1/3, 1/3, 1/6 (referred to a unit length of tube).

If we want an approximation to three formants, we will use a simulation of a tube divided into eight sections of successive lengths relative 3/30, 2/30, 4/30, 6/30, 6/30, 4/30, 2 / 30, 3/30.

We will show in the following how these values are obtained.

The theoretical values of these lengths can be calculated precisely, but of course the practical values can only be approximations of the theoretical values without fundamentally changing the overall result of speech analysis or synthesis.

For the determination of the sensitivity of the formants to variations in cross-sections, the following approximation can be made, consisting in plotting the function of sensitivity of the formant to variations in cross-section of a section as a function of the position of this sections between the end upstream and downstream end of the tube.

For the first component, this function can be assimilated to a half-period of sinusoid, the sensitivity being negative and maximum at the upstream inlet of the tube, zero in the middle, and positive and maximum at the outlet. By positive sensitivity is meant an increase in the frequency of the formant for a growth in the section considered. A negative sensitivity is a decrease in frequency for a growth in section.

For the second component, the sensitivity function can be compared to three half-periods of sinusoid between the input and the output. For the i ^th forming, the function can be assimilated to a sinusoid whose half-period is L / (2i-1) where L is the total length of the tube, the sensitivity being maximum and negative at the upstream input (there therefore has 2i-1 half-periods between the entry and exit of the tube for the sensitivity function of the i ^th forming).

The zero crossing zones of the sensitivity of the different formants constitute the boundaries of the sections of successive tubes. There are N = 2 + n (n-1) total sections if we make an approximation to n formants.

The action on the sections of the pipe sections of the simulation device can be exercised in several different ways:
- action on the overall section of the section
- action on the section of a local portion of section located towards the middle of the section (to act on all the formants at the same time)
- action on the section of a local portion of section located at the border between two sections (if we want to voluntarily remove the action on one of the formants: the one whose sensitivity is canceled out at this border).

By this judicious organization of well chosen sections of tubes, we directly linked the analysis and synthesis of human speech to the concept of formants, which minimizes the number of control parameters of the simulation device when we want produce sounds whose formants and their variations have rightly been listed.

This organization is therefore fundamentally different from the proposals already made in terms of simulation by tubes of variable section since until now we have been content to subdivide the tubes into sections in an artificial way: subdivision into regular sections of the order of 1 cm long or, by analogy with the vocal tract, subdivision between an area of larynx, pharynx and arbitrary subdivision in the mouth.

• - la figure 1 représente la forme générale d'un conduit vocal humain ;
• - la figure 2 représente la schématisation de ce conduit sous forme d'un tube divisé en tronçons de sections différentes, variables individuellement ;
• - la figure 3 représente le schéma-bloc d'un dispositif de synthèse de parole ;
  
• - la figure 4 représente le tracé des courbes de sensibilité des quatre premiers formants d'un tube uniforme ;
• - la figure 5 représente la division d'un tube selon l'invention en quatre tronçons pour une approximation limitée aux deux premiers formants ;
• - la figure 6 représente la division d'un tube selon l'invention en huit tronçons pour une approximation limitée aux trois premiers formants ; et
• - la figure 7 représente la division d'un tube selon l'invention en quatorze tronçons pour une approximation limitée aux quatre premiers formants.

Other characteristics and advantages of the invention will appear on reading the detailed description which follows and which is given with reference to the appended drawings in which:

• - Figure 1 shows the general shape of a human vocal tract;
• - Figure 2 shows the diagram of this conduit in the form of a tube divided into sections of different sections, individually variable;
• - Figure 3 shows the block diagram of a speech synthesis device;
  
• - Figure 4 shows the plot of the sensitivity curves of the first four formants of a uniform tube;
• - Figure 5 shows the division of a tube according to the invention into four sections for an approximation limited to the first two formants;
• - Figure 6 shows the division of a tube according to the invention into eight sections for an approximation limited to the first three formants; and
• - Figure 7 shows the division of a tube according to the invention in fourteen sections for an approximation limited to the first four formants.

FIG. 1 represents in section the simplified anatomy of a human vocal tract with different regions and organs such as the vocal cords CV constituting the source of air flow (of very specific periodic wave form), the LU uvula, palate PL, tongue LN, teeth DN, upper lips LS and lower LI.

FIG. 2 represents a diagrammatic representation of the vocal tract in the form of an acoustic tube 10 composed of adjacent cylindrical sections T1, T2 ... T16, having sections which are different from one another at rest, these sections may vary independently of each other; the combination of the variations of section of the different sections makes it possible to produce sounds. The vowels are mainly expressed by relationships between the different sections. Consonants tend to translate into transitions between a first combination of sections and a second combination.

For speech synthesis, the tube is placed behind an air flow source reproducing the characteristics of the vocal cords, that is to say in particular a periodic flow wave of period approximately 10 milliseconds having a sawtooth shape. very rounded, the growth front being slower than the decreasing front.

Given the difficulty of mechanically producing such an acoustic tube, we will rather use modern technologies. last computer simulation, in which the acoustic behavior of the tube can be determined, that is to say that the air flow and pressure at each point and in particular at the outlet of the tube, can be calculated; the characteristics of the electrical signal that must be applied to a loudspeaker to reproduce this flow rate and this pressure are also calculated, and an electrical signal having these characteristics is produced by a generator controlled by the computer.

FIG. 3 very schematically represents this hardware embodiment of a speech synthesizer by simulation: a data input member defines the succession of phonemes to be produced. This member may for example be an alphanumeric keyboard CL on which keys or combinations of keys represent phonemes.

These data are applied to the CALC computer in the form of electrical signals in a conventional manner, via a link bus.

The computer controls an electrical signal synthesizer (GEN) which itself controls an HP speaker.

The operation of the computer is controlled as follows: from the sequence of keyboard commands, a set of parameters is generated; these parameters correspond to the values of the sections of the sections of the acoustic tube representing the vocal tract and to the variations of these sections over time.

Data processing is simply simulation by calculation of the behavior of the tube having these sections and these section variations. This behavior is now well known and is described, for example, in the aforementioned work by J.L. Flanagan.

The treatment first results in air flow and / or air pressure at the outlet of the tube, then in the generation of the characteristics of the electrical signal which must be applied to a loudspeaker to reproduce the pressure at the outlet. We can assume for simplicity that the air pressure caused by the loudspeaker is proportional to the instantaneous electric current which feeds it. In this case, the processing consists in determining at each instant what is the waveform of the air pressure representing the desired sound, the synthesizer of electrical signals providing a waveform of current corresponding exactly to the shape of wave of the calculated air pressure. Of course, if the loudspeaker has a nonlinear air pressure / electric current response curve, the calculation must take this into account.

Given that the invention does not relate to the principle of synthesis or analysis of speech by simulation of the acoustic behavior of a tube, principle which is known, but on the choice of the parameters of the simulation, we will now detail this choice.

The choice relates to the lengths of the tube sections used in data processing.

That is to say that the parameters stored in memory in the computer will not be the section variations of sections of a tube cut arbitrarily into sections of any length (as is the case in FIG. 2 where we have taken for convenience of the sections which all have the same length) but these parameters will represent the section variations of sections of well determined lengths resulting from the cutting according to the invention which will now be explained in detail.

We start from a tube of total length L (for example from 15 to 20 cm, which corresponds to the length of the vocal tract). The acoustic response of this tube presents formants, that is to say more or less pronounced resonances at certain frequencies. The spectrum of an acoustic signal emitted at the entry of the tube will be modulated by these formants and will present local maxima at the frequencies of the formants.

The theoretical acoustic study of a tube of length L shows that the frequency of formants varies depending on the section of the tube. But it does not vary in the same way everywhere: if one varies the section of the tube only locally in the middle of the length of the tube, one realizes that the frequency of the formants does not vary at all; if, on the contrary, the section is varied only at the mouth of the tube or at its outlet, it is found that a variation in section varies the frequency of the formants: if it is at the mouth of the tube that the section varies, the frequency of formants increases as the section decreases; if, on the contrary, it is at the outlet of the tube that the section varies, the frequency of formants increases as the section increases.

Finally, if the section of the tube is varied at any location, the frequencies of the different formants will vary with different amplitudes and directions.

In fact, for a tube initially of uniform cross-section, it is possible to give a theoretical representation of the sensitivity of the formants, that is to say of the direction of variation of the frequencies of the forming as a function of a local variation in cross-section of the tube, because the sensitivity of the formants varies sinusoidally along the tube between the mouth and the outlet, the period of the sinusoid being different for each of the formants.

This is what is shown in FIG. 4: the diagram 4a represents the sensitivity curve SF1 of the first forming F1 of the tube as a function of the position x (x varying between 0 and L) at which a variation in section is produced.

Diagram 4b represents the sensitivity curve SF2 of the second forming F2, diagram 4c represents the sensitivity curve SF3 of the third forming F3, and diagram 4d represents the sensitivity curve SF4 of the fourth forming F4.

On these curves, we did not worry about the relative value of the sensitivities SF1, SF2, SF3, SF4 relative to each other to others. Only the form of variation, the signs, the positions of the maxima and minima and the passages through zero interest us according to the invention. We therefore gave a maximum unit value to each of the sensitivities.

The theoretical form of the sensitivity curves of the formants as a function of the position x to which a section variation is applied is very simple: it is a sinusoid whose half-period is L / (2i-1) where i is the rank of the form: i = 1 for the first form F1, that is to say for the lowest resonant frequency; i = 2 for the next resonant frequency immediately; And so on. This sinusoid has a minimum (maximum sensitivity in absolute but negative value) at the mouth of the tube (x = 0) and a maximum (maximum and positive sensitivity) at the end of the tube (x = L).

We can verify that the tube is antisymmetrical, that is to say that an action on the section at any point on the abscissa x acts on the different formants in exactly the same way, but with an opposite sign, that an action on the section at a point of abscissa Lx.

For x = L / 2 the action is therefore zero: the sensitivity goes through zero at this point for all formants regardless of their rank.

This remark will be important for the rest because it will limit the number of command parameters of the speech analysis or synthesis device: the same variation in frequency of formants is obtained, for all the formants at the same time by acting on the section at the abscissa point x instead of the abscissa point Lx provided that the section at this point is varied in the opposite direction to that which would have been used at point L-x.

The explanations above were in the context of a tube initially of uniform section to the sections of which small variations are applied. Experiments carried out by The inventors have shown that in the case of a tube divided into sections of variable sections and in the case where significant variations are applied to these sections, the directions of variation are preserved even if the sensitivity functions are no longer sinusoidal.

The invention proposes to divide the tube into sections whose limits correspond exactly to the zero crossings of the sensitivity of the formants with which one wishes to make an approximation of speech analysis or synthesis: each passage through zero defines the limit of a section.

The zero crossings of the formants' sensitivity are located on the x-axis:
- A0 for the first forming F1
- B1, A0, B′1 for the second forming F2
- C1, C2, A0, C′2, C′1 for the third forming F3
- D1, D2, D3, A0, D′3, D′3, D′2, D′1 for the fourth forming F4 and so on.

The values of these abscissas are as follows: A0 = L / 2 (middle of the tube) B1 = L / 6 B′1 = L - L / 6 C1 = L / 10 C′1 = L - L / 10 C2 = 3L / 10 C′2 = L - 3L / 10 D1 = L / 14 D′1 = L - L / 14 D2 = 3L / 14 D′2 = L - 3L / 14 D3 = 5L / 14 D′3 = L - 5L / 14

We will give three examples of cutting according to the invention then a general rule:

First example : we want an approximation to two formants F1 and F2.

The tube is cut into four sections which are:
- a first section from 0 to B1 (length L / 6)
- a second section from B1 to A0 (length L / 3)
- a third section from A0 to B′1 (length L / 3)
- a fourth section from B′1 to L (length L / 6)

The corresponding tube is shown in Figure 5.

Second example : we want an approximation to three formants F1, F2, F3.

The tube is divided into eight sections which are:
- a first section from 0 to C1 (length L / 10)
- a second section from C1 to B1 (length L / 15)
- a third section from B1 to C2 (length 2L / 15)
- a fourth section from C2 to A0 (length 3L / 15)
- And four other symmetrical sections of the first four relative to the middle of the tube.

The tube is shown in Figure 6.

Third example : we want an approximation to four formants F1, F2, F3, F4.

The tube is divided into 14 sections which are shown in Figure 7 and which are:
- a first section from 0 to D1 (length L / 14)
- a second section D1 to C1 (length L / 35)
- a third section C1 to B1 (length L / 15)
- a fourth section from B1 to D2 (length L / 21)
- a fifth section D2 to C2 (length 3L / 35)
- a sixth section C2 to D3 (length 2L / 35)
- a seventh section D3 to A0 (length L / 7)
- and seven other symmetrical sections of the first with respect to the middle of the tube.

To generalize the method to an approximation with n formants (although it is unlikely that we want to exceed n = 4), we determine the abscissa Xi, j of the j ^th passage through zero of the sensitivity of the i ^th form, for all formants (i = 1 to n) and over the entire length of the tube (j = 1 to 2i - 1).

We have Xi, j = L (2j - 1) / (2i - 1) x 2.

We classify all Xi, j in ascending order along the tube at their respective positions; each tube section is delimited by two adjacent abscissas of this classified sequence, the first section starting at the abscissa 0 and ending at the abscissa Xn, 1 = L / 2n-1 and the last section starting at the abscissa Xn , 2n-1 = L - L / (2n-1) and ending at the abscissa L.

The total number of sections is N = n (n-1) +2.

We have thus precisely defined a series of very important parameters for operating the speech analysis or synthesis device, these parameters being the number of sections and the length of each.

These parameters are supplied to the computer and the data processing consists of an action on the section of the sections defined by these parameters. The action may relate to a number of sections equal to half the total number, for the reason of symmetry indicated above.

By detailed studies we will determine what are the section variations to be performed on each section to produce this or that phoneme (and we are guided by the knowledge already established on the frequencies of formants and variations of frequencies of formants corresponding to these phonemes ).

A data memory can be associated with the computer, memory directly containing for each phoneme the sequences of section variations of the sections thus defined.

In a speech synthesis device, the triggering of these variation sequences results, after processing in the computer, in the generation of electrical signals transmitted to the loudspeaker, and in the production of the desired phoneme.

In a speech analysis device, this is done by looping back: a microphone receives the sounds, converts them into electrical signals. These signals are processed by the computer. A comparison is made between data from the processing and data generated by sequences of section variations corresponding to known sounds.

The invention can be used as an educational speech synthesis toy allowing a better understanding of the development of sounds by the human vocal system. In this case, the source could be a mouthpiece comprising a reed in which the user will blow. We can also use a white noise source. We will use 4 or 8 sections whose volumes are controlled by pistons controlled by the fingers of the hand. The device can be manufactured by plastic molding.

Claims (7)

1. Apparatus for analysis, coding or synthesis of speech using a device for simulating the acoustic behavior of a tube constituted by a succession of sections (T1, T2 ...) of different and variable sections placed end to end, characterized in that the tube comprises a set of N sections, divided into subsets of successive rows as follows: the set of N sections is divided into two subsets of rank 1, the first subset, of the upstream side of the tube, corresponding to a negative sensitivity to variations in section for the first forming and the second to a positive sensitivity, each sub-assembly of rank i being divided in the same way into two sub-assemblies of rank i + 1 s '' there is a change in sign of the sensitivity of the forming i + 1 in this subset, one of the subsets corresponding to a negative sensitivity for the (i + 1) ^th forming and the other to a positive sensitivity , each subset of rank (n-1) being finally divided into d them sections, one of the sections corresponding to a negative sensitivity of the n ^th forming and the other to a positive sensitivity, the sensitivity of the i ^th forming to the section variations of a section representing the relative variation of the frequency of the i ^th forming as a function of a variation in section of this section; the device having, for analysis or synthesis control parameters, on the one hand, the section variations of some of the tube sections thus defined, and, on the other hand, the total length L of the tube; the device receiving signals from a microphone or supplying signals to a loudspeaker depending on whether it functions as a speech analyzer or synthesizer. 2. Apparatus for analysis or synthesis of speech according to claim 1, characterized in that the number of sections of the tube is N = 2 + n (n-1) if one wishes to make a synthesis or ana speech lysis with an approximation corresponding to the first n formants of the tube. 3. Apparatus for analysis or synthesis of speech according to claim 2, characterized in that n = 2, and in that the tube is divided into N = 4 sections whose successive lengths are L / 6, 2L / 6, 2L / 6 and L / 6 respectively. 4. apparatus for analysis or synthesis of speech according to claim 2, characterized in that n = 3, and in that the tube is divided into N = 8 sections whose successive lengths are L / 10, L / 15, 2L / 15, 3L / 15, 3L / 15, 2L / 15, L / 15 and L / 10 respectively. 5. apparatus for analysis or synthesis of speech according to claim 2, characterized in that n = 4, and in that the tube is divided into N = 14 sections whose successive lengths are L / 14, L / 35, L / 15, L / 21, 3L / 35, 2L / 35, L / 7, L / 7, 2L / 35, 3L / 35, L / 21, L / 15, L / 35 and L / 14. 6. speech synthesis apparatus according to claim 1, characterized in that it is produced physically in the form of a succession of cavities corresponding to said sections, means for adjusting and controlling the volume of each cavity being provided. 7. speech synthesis apparatus according to claim 6, characterized in that the upstream cavity is associated with a mouth into which a user can blow and in that the adjustment means comprise pistons which can be actuated manually.

Images