DE2312356A1 - METHOD AND DEVICE FOR DETERMINING THE BASIC FREQUENCY OF VOICE SIGNALS - Google Patents

Abstract

1383621 Excitation frequency detector INTERNATIONAL BUSINESS MACHINES CORP 6 March 1973 [28 March 1972] 10914/73 Heading H4R Fundamental frequency of a speech signal is determined by sampling the speech signal and generating digital values representative of each sample, determining the absolute value of the difference between each successive pair of sample values, continuously summing the absolute differences to obtain for each sample "i" a function "S(i)" and detecting variations in the slope of that function to indicate the time at which the mouth cavities are excited, i.e. the time at which differences between successive sample values start to increase after a period in which they have been decreasing. As described speech is sampled at 100 Ásec intervals and digitally coded in pcm coder 1. Differences between successive samples are provided by subtractor 7 fed directly from coder 2 on one input and via a one sample period delay 4 on the other input. The absolute value of the difference is provided by unit 10. Adder 11 and register 13 enable a running total to be provided for the absolute differences, while shift registers 17 and 19, both of which provide twenty sample periods delay, allow the comparison, in subtractors 20 and 21, of the running totals at three sampling points spaced twenty sampling periods apart, to give an indication of the slopes before and after the middle sampling point. The slopes are compared in subtractor 23 to provide on line 26 the magnitude of the difference in slopes, and on line 27 the sign of the difference in slopes. When line 27 is positive, corresponding to an increase in slope of the sum of the difference signals which occurs at the commencement of the excitation period, AND gate 28 is enabled passing the signal on line 26 to line 29. The slope difference signal is filtered in digital filter 30, to remove small high frequency variations, and applied to comparator 44 and register 42, via AND gate 43 which is enabled only when the comparator 44, which compares current slope and the slope stored in register 42, provides an output indicating that the current slope is larger than the stored slope. When the opposite condition occurs the value then stored in register 42 is held and the comparator 44 output signal is fed through inverter 47, and gate 48 to bi-stable circuits 52, 53, to produce a short pulse on the output of AND gate 57 indicating the start of the excitation period. When the sign of the slope difference signal on line 49 first goes positive, prior to the commencement of an excitation period, the register 42 and bi-stable 52, 53, are reset ready to detect the following maximum slope difference indicating start of the excitation period. The units shown dotted are provided to give a variable threshold to the maximum slope difference value detector by enabling and gate 48 only when the maximum detected by the comparator 44 bears a predetermined relation, e.g. more than half, the previous maximum detected.

Description

Method and device for determining the fundamental frequency of speech signals

The invention relates to a method for determining the basic speech frequency by determining the times of excitation of the oral cavities in the case of voiced sounds, as well as a device to carry out this procedure.

The problems associated with the determination of fundamental speech frequencies occur, have proven to be extremely difficult, which may partly be due to the fact that the definition of the Concept of basic speech frequency causes difficulties. The main problem lies in the large number of influential parameters than there are: temporal changes in the excitation curve and quasi-periodic excitation of the oral cavities, temporal changes the amplitudes of the basic frequency or determination of the times of application of the oral cavities. The last of these parameters is of particular importance because it is used in a number of devices for measuring the fundamental frequency will. This use is not unsurprisingly when it is found that the precise determination of the exact times of excitation has not yet yielded a satisfactory solution.

309842/0790

leads could be. In fact, this difficulty has hitherto been circumvented by making the hypothesis that that the maximum amplitudes of the speech signal shown as a function of time mark the beginning of voiced sound segments and thus also determine the point in time at which the oral cavities are stimulated.

Based on this hypothesis, the previously known Devices for determining fundamental speech frequencies from an analog representation of the speech signal. A maximum detector localizes the amplitude maxima of the analog signal or an arrangement for determining rapid spectrum changes determined these changes. The fundamental frequency is then determined by measuring the time intervals between the amplitude maxima or determined the observed spectral changes. However, the relation between one is established in this way Time interval and the real value of the fundamental frequency because of the instability and variations of the excitation signal Oral cavities, as produced by the subject's speaking apparatus, are never correct. Added to this instability The inaccuracy resulting from the hypothesis mentioned, namely the poor correlation between the Excitation times of the oral cavities and the maxima of the Speech vibration.

Accordingly, the most important object of the invention is to determine the actual excitation times of the oral cavity vibrations. Furthermore, a method can be specified, the determination of the excitation times with great accuracy by means of digital, and not analogue calculations allowed. That the arrangement necessary for this should be as simple as possible so that it is both economical exploitable works as well as sufficient operational safety owns, of course.

The procedure that solves the task is drawn

309842/0790

FR 9 71 002

is characterized in that an analog signal representing the speech oscillation is sampled in sections and the sampled Value is digitized that for each sample the amount of the difference d (i) between that for this sample and the for the previous sample determined value is determined that the magnitudes of the differences d (i) between successive Samples according to the formula

i - 1

I d (il) 1

be added to a sum S <i), which is the amount of up to indicates the differences that have occurred during this sampling in order to obtain a signal whose mean value curve corresponds to the successive Sums S (i) of the amounts of sudden changes in the course shows the points in time of the excitation of the oral cavities, and that these points in time of excitation from the center of the curve will.

An advantageous device for carrying out this process that meets the above requirements Fulfilled economy and simplicity, is characterized by a voice signal picking up and in a sequence digital value converting encoder, a subtraction circuit for calculating the difference between successive scanning results, means for determining the amount the difference in successive sampling results, an adding circuit which the calculated difference values respectively to the sum of the previously calculated difference values, the successive sums via a feedback between The output and input of the adder circulate, and a curvature detector that the successive sums of the amounts picks up and emits a signal whenever a positive curvature occurs with sufficiently high sum values.

Further advantageous embodiments of both the method and

309842/0790

FR 971 022

the device according to the invention can also meet the claims can be removed. Details of the method and the mode of operation of the device can be found in the description below taken together with the drawings.

On the drawings shows:

Fig. 1 is a graphical representation of the sound vibrations the syllable "an" in French, for example from the word "chant", in pulse code modulation (PCM);

2 shows a similar representation of the sound "i",

for example from the word "cigale";

3 shows a section of an analog representation of the

sound produced by a human voice;

4 shows the value of the successive as a function of time

Sums of the absolute values of the differences in successive patterns for each pattern;

Fig. 5 is a diagram that as a function of time

shows the concavity of the curve shown in Figure 4;

6 shows a detector for carrying out the method according to the invention.

As already mentioned above, FIG. 1 shows the time-dependent amplitude program of the vocal sound that occurs during the French pronunciation the syllable "an" occurs from the word "chant". Such a representation is achieved by sampling the analog oscillation

309 84 2/07 90

FR 971 022

and coding of the scanning result in PCM (Pulse Code Modulation) - each of the crosses on the curve corresponds a sampling point, the time interval between two sampling points being 100 microseconds. A solid one. Line connects the cross marks, which correspond to the encoded sampling results of the voice signal, to indicate the change to show this signal in a form corresponding to the analog representation. Comparing a signal like it shown in Fig. 1, with an analog sound signal as shown in Fig. 3, shows both kinship and also differences in the different modes of representation. Fig. Shows the general form of a signal that consists of excitation and Damping periods. The real purpose of such a digital representation is to make visible the fact that the signal shown consists of high and low frequencies. The prior art shows that The vibrations of the vocal cords stimulate the cavities both in the back of the mouth and near the lips. The three cavities have different sizes; while the former are relatively extensive the cavities near the lips are of smaller size. So are the Resonance frequencies relatively low for the areas in the back of the mouth and correspondingly high for areas near the lips. At the time the vocal cords are stimulated, the various cavities are stimulated; their vibrations can be considered in phase can be viewed with the excitation oscillation. The amplitudes of the different frequencies are added and the maxima Ml, M2, M3, M4 of the sound signal are generated in this way. Immediately after the excitation, however, the vibrations are damped with different damping coefficients, the low frequencies are attenuated weaker than high frequencies. The attenuation is shown in Fig. 1 by means of a solid line made visible, which also illustrates the much faster attenuation of the high frequencies. The entire curve after point P shows that the high frequencies are due to the

971 O ₂₂ 309842/0790

There has already been strong damping, practically no change in amplitude more subject.

However, these remarks, which are familiar to those skilled in the art, become a new point of concern should be added below with regard to two techniques already used to measure the fundamental vocal frequency will be explained. For explanation, reference should also be made to the representation shown in FIG. 2, which is the PCM representation of the sound "i", as it is in the French language in the word, "cigale" shows up. Please also refer to the above-mentioned analog representation of a speech oscillation in Fig. 3.

The first method, which can be assigned to the prior art, consists in equating the times of occurrence of maxima with the times of excitation of the cavities and assigning these maxima measure up. However, the equation mentioned is only an approximation. Indeed, at the time of the Excitation of the oral cavities the sound signal in the vicinity of the zero line, as can be seen at point A in Fig. 2, and it only reaches its highest maximum at point F within the two illustrated excitation methods. You can see clearly that between the points A and F of the excitation periods Pl and P2 are 1.4 and 1.7 milliseconds apart, i.e. that below assuming one sampling per 100 milliseconds, a total of 14 or 17 samples of the signal take place. Under According to this requirement, an error is introduced at the time of the measurement of the fundamental frequency, which results from each excitation period inherent inaccuracies. They rely on the fact that you can start each excitation period at the time at which the oral cavities reach their maximum stimulation instead of having them begin where the oral cavities begin, getting excited. With the help of such devices, the actual stimulation times of the oral cavities can accordingly never be determined.

309842/0790

FR 971 022

In addition, FIG. 2 utilizes the PCM representation makes clear not only one, but even several maxima of the speech oscillation. In contrast, Fig. 3 shows only a single one Maximum that clearly stands out from a period of maximum stimulation of the oral cavities. Those shown in FIG Maxima all correspond to a significant stimulation of the oral cavities. As a result, one of the prior art associated maximum detector, for example, select one of the maxima D, E or F, this being done during the excitation period P1 selected can be different from that determined during the subsequent period P2; here should be the periods begin at the time of vocal cord stimulation. This procedure therefore becomes an additional error for measuring the fundamental frequency which might explain the instability and variations that the excitation signal has attributed to the oral cavities.

The second method, also belonging to the state of the art, consists in the sudden changes in the speech vibrations to be determined in relation to time. For this purpose, a threshold value detector is generally used, which is determined by the maximum amplitudes of the speech signal is excited at the time of maximum excitation of the oral cavities. Here, too, it becomes clear that, since each excitation period Pl and P2 has several abrupt amplitude changes, the threshold detector both by the first of these sudden changes (D) - if this maximum is above the threshold value - can be switched on, as well as one of the following maximum amplitudes (E or F), depending on the setting of the threshold value in the detector. In the end, this results in the same typical error as already explained above. On top of that, it doesn't matter which sudden amplitude change is recorded for each excitation period, the point in time corresponding to this change never coincides with the moment of stimulation of the oral cavities. One might assume that by defining a very low threshold it should be possible, for example

309842/07 90

FR 9 71 022

in the excitation period P2 the sudden change in the speech signal, which corresponds to the maximum B to recognize. In reality, however, this is impossible, as it risked To determine maxima that have nothing to do with triggering the excitation of the oral cavities " Definition of the detector threshold value below the amplitude of point B in period P2 also includes curve points such as the point X trigger the detector. Consequently, the threshold is set to a sufficiently high / Level in order to rule out any uncertainty regarding the determined maximum.

It follows that none of the prior art known arrangements allow the measurement of the time of excitation of the oral cavities at all. All realized so far Measurement methods are flawed and in principle only approximate methods. A procedure according to the invention allows with the precision given by the electronic possibilities the determination of the actual time of excitation the oral cavities. To do this, the voice signal, which is originally in analog form, is PCM-coded, as shown in FIGS. 1 and 2 is represented by the cross marks. It. has already been said that the State-of-the-art devices characteristic Statements (maximum amplitude or exceeding a given threshold value) from the speech signal in analog form to have. The method according to the invention and the devices that allow it to be carried out determine characteristics of the speech signal after the latter has been subjected to a special treatment. For example, let it be a series of sampling results in PCM modulation of the speech signal given, whose individual values are preceded by a sign (+) or (-) are provided, depending on whether the speech signal is positive or negative at the time of the corresponding sampling. In the table below, which relates to FIG. 1, is

^{FR 971} ° ²² 309842/07 9Ö

continued this example.

Scan no.
value

1
+ al

2 + a2

3 + a3

4 + a4

5 + a5

7th
+ a7

8 + a8

9 + a9

10 -alO

11 + all

By determining, for each sample, the difference between the value assigned to it and that of the previous one, and determines the absolute value of the successive differences, one obtains the absolute values of the amplitude differences of the Speech signal between two consecutive samples.

The values given in the table above are accordingly obtained for the curve shown in FIG. 1.

Looking at FIG. 1, it will be found that the values d1 ... d10 are in the vicinity of the start of the excitation period as well as in the vicinity of the excitation maximum of the oral cavities increase and maintain a higher value while they decrease beyond the time of maximum excitation and have a low value. One states accordingly at the beginning of a voiced period:

dl

d2 <d3 <d5 <d7

d8

Etc..

In contrast, the values d (i) at the end of the same form voiced Period a decreasing series or stay the same. Some even go to zero:

d (n-7)> d (n-4)> d <n-3)> d (n - l)> d (n + 3)> d (n + 5) ■ 0

In between there are values d (i) that increase or decrease do not follow the series of values, e.g. d4 and d6, d (n-6), d (n-5), d (n-2), etc. In any case, one will find that the general tendency for values to rise immediately

309842/0790

FR 971 022

after the time of stimulation of the oral cavities and a drop in the further course of a voiced period. this agrees with the statement made above regarding the stimulation of the vocal cords and the breakdown of those caused by the oral cavities generated frequencies match.

The second step of the method according to the invention consists in adding each difference d (i) to the sum of the absolute values of the previous differences. This gives you the following values S (i):

51 = dl

52 = dl + d2

53 = dl + d2 + d3

54 = dl + d2 + d3 + d4

I.
Sx = dl + d2 + d3 + d4 + d5 +. ' + dx

If one puts these different values S (i) as a function of time graphically, a cluster of points is obtained, the mean value curve of which is shown in FIG. The relationships of this Signal shape related to the function of the oral cavities are the following:

3 0-9 842/0790

FR 971 022

DIFFERENCE VALUE ABSOLUTE VALUE Sampling 1 + al Sampling 2
Sampling 1 + a2 - al dl Scanning 3
Sampling 2 + a3 - a2 d2 Sampling 4
Scanning 3 + a4 - a3 d3 Sampling 5
Sampling 4 + a5 - a4 d4 Sampling 6
Sampling 5 -a6 - a5 d5 Sampling 7
Sampling 6 + a7 + a6 d6 Sampling 8
Sampling 7 + a8 - a7 d7 Scanning 9
Sampling 8 + a9 - a8 d8 Sampling 10
Scanning 9 -alO - a9 d9 Sampling 11
Sampling 10 + all - alO dlO

309842/0790

FR 971 G22

Those sections of the curve which have a sharp rise - for example section A - correspond with the times following the time of maximum excitation of the oral cavities; are plotted in the ordinate direction the successive sums S (i) as a function of the difference values d (i), their amount immediately after the point in time the stimulation of the oral cavity increases.

The parts of the curve which - like section B - are less Have slope correspond to the respective ends of the excitation periods; during these times the values d (i) are low or even zero.

The points C indicate a sudden change in slope the curve as well as a change in curvature; the turning points C occur at the transition from a section B. in a section A on, these points correspond to the times of stimulation of the vocal cords at the beginning of each voiced Period. In fact, the points C on the curve are between the points that mark the end of a period Characterize B of the excitation and those points that denote the beginning of an excitation period (section A).

Both the change in the slope of the curve and the occurrence of the turning point is used to determine the excitation times to characterize the oral cavity. These times are from ^ easy to determine going from the curve points, provided that there is a linear axis between the time axis and these points Relation exists.

The information that there is a sudden change in slope and curvature is obtained by adding in at each point on the curve the sums of the differences from FIG. 4 are formed, a measurement of the difference between two Slopes, which are to be referred to here as inferior slope PI and superior slope PS. Do you perform this operation information is obtained for each point on the curve

P _R 971 022 309842/0790

not only in terms of the sudden change in slope, but also in terms of the curvature of the curve in each of them Points. If it is found that the curvature of the function shown in Fig. Is generally negative, except for the When the vocal cords open, one notices a rapid change in curvature in points C. This change in curvature occurs approximately at the same time as the abrupt change in slope and is in the device used according to the invention.

It has already been mentioned that the curve shown in FIG. 4 represents a mean value curve, the mean value from the Values S (i) is formed, which at the beginning of a voiced period a series with increasing values and for the remainder the period is a series of falling values. It was also found that certain values of the sequence that are to be referred to here as irregular values, do not show this behavior of rise or fall. Consequently will in order to prevent the occurrence of errors that could be introduced by these irregular values, the operations the determination of proven changes in curvature and slope through evaluation - for each point of the curve in Fig. 4 - the values of the slopes PI and the slopes PS to an equal number N of the previous values S (i) or follow, where the value S (i) corresponds to the calculation point of the curve. It should be noted that the number N is such should be selected so that it allows the timing of the opening of the vocal cords to be determined for each speaker. Consequently, one must take into account the speaker whose point in time is closest to the opening of the vocal cords lying together. This condition is fulfilled by the highest or sharpest female voices, whereby the vocal cords open every three milliseconds on average. Hence, the numerical values of N must be below approximately 3 milliseconds choose such that for each point of the curve in FIG. 4 for calculating the curvature a section A is larger

fr 971 022 309842/07 90

Slope and such a B small slope is available for calculation. In the case of a different design, i.e. if N is greater is chosen as 3 milliseconds, the number of points on the curve used to calculate the slope PS and the slope PI also includes all points of a voiced period Pl or P2 from FIG. Include all of these in the calculation If points are included, the determination of the curvature at points C of the curve shown in FIG. 4 proves to be impossible. A favorable value for N is the number 20. This value corresponds to a speech signal lasting 2 milliseconds provided that the values S (i) shown in FIG. 4 are 100 microseconds apart.

The following describes the individual stages in determining the points in time at which sudden changes occur the curve gradients are available. The determination is made for each point i of the curve in FIG. 4, in which .on the one hand the Difference between the value S (i) at this point and the value S (i-20) at point i-20 is formed to determine PI (i), and on the other hand, to determine PS (i), the difference between the values S (i + 20) and S (i) is determined. After that, will PS (i) subtracted from PI (i). The following calculation is carried out for each point i:

PI (i) - PS (i) = [S (i) - S (i-20)] - [S (i + 20) - S (i)] = D (i)

For some points on the curve the difference PI (i) -PS (i) is positive, negative for others. It can also be seen in Fig. 4:

PIl - PSl <0
PI2 - PS2> 0

The experiments carried out show that the consequence of the differences PI (i) -PS (i) decreases in front of the zone of positive curvature surrounding point C, some positive or negative and close Runs through zero values and then increases very sharply.

309842/0 7 ^ 9 0

FR 971 022

The section of the curves in which the increase in the difference sequences PI (i) -PS (i) takes place very quickly also has positive ones Curvature. In FIG. 5, the course of the mean values is shown, obtained from the values corresponding to the values D (I) = PI (i) -PS (i) Points, in addition, this representation shows the relative course of the curve shown in FIG. 4 with indication of the positive or entered with a negative sign. In this illustration one can clearly see the sudden variations in the value D (i). The different Points forming the curve in FIG. 5 were determined from the points on the curve in FIG has been explained, have time-dependent ascending order; so they became corresponding to values S (i), each partial are below or above the mean values forming the points in FIG. 4.

If one no longer presents the mean value curve, but rather the curve of the exact values D (i) for a part of FIG. 5, as it is is shown in the enlargement of a section of this figure, it is noted that the series of positive values of the difference PI (i) -PS (i) is not characterized by the fact that every value is systematically greater than the one that precedes it in time. Therefore you can, if you bl ... b6..etc. than the positive values of PI (i) -PS (i) at times ti ... t6 ... etc. referred to, starting read the following relationships from Fig. 5:

bl <b2 <b3 <b5 <b6 <b4 ... etc.

to identify those among these values that are the greatest Have magnitude, i.e., which correspond to the strongest positive curvature of the mean value curve in Fig. 5, one becomes a smoothing perform the curve, preferably in digital form, e.g. by digital filtering of the successive values D (i) = PI (i) -PS (i) and then determine the value that has the large amount. These steps of the process of the invention are more detailed Form in the context of the description of an implementation of the present invention, as shown in Fig. 6, explain

309842/0790

FR 971 O22

FIg. 6a shows the circuits for calculating the values SCi) ' the curve shown in Fig. 4 and the values DCi) = PICi) -PSCi) the one in Flg. 5 curve shown; Fig. 6 represents the circles Smoothing of the curve shown in FIG. 5 and the detector circles to determine sudden changes in incline and maximum Curvatures.

The signal originating from the speaker, its stimulation times The oral cavities should / will be determined via line 1 given to circuit 2, one of which is the incoming analog signal in a PCM modulated transforming encoder. The output signals of the encoder 2 are successively via line 3 entered into register 4 via gate circuit 5 and at the same time via line 6 into subtraction circuit 7. Register 4 is so large that there is just one sample in PCM modulation of the speech signal can accommodate. The subtraction circuit 7 receives from the output of the register 4 via line 8 another input signal. Each scan entered in register 4 is then removed from it again when the following The scanning result is entered into the circuit 7. Moving on the scanning result into or out of register 4 takes place under control of the gate circuit, which is activated by a command signal CSl. Entering a new scan result causes the previous one to move on at the same time In the subtraction circuit 7 »are accordingly the sampling result in the subtraction circuit 7 for each time point t (I) aCD and the subsequent scanning result a (i + I). The line 9 formed by the subtraction circuit 7 is used Difference given on the transmitter 10, its output line 15 now the absolute value - or the amount id (i) i of the input signal.

The loop of adder circuit 11, line 12, register 13 and line 14, from the output of the register 13 to a

309842/0790

FR 971 022

Input of the addition circuit 11 returns, allows the calculation of the successive sums S (i) obtained by taking the sum of the difference d (i) at time t (i) on line 15 and the total of the differences from previous periods ^

= | a (i + l) - a (i) | + I | a (i) - a (il)

1 ·

The addition circuit 11 receives the absolute values d (l) = | a (i + l) -a (i) I via one of its inputs, which it receives via line 15 from the transmitter 10 is supplied. She also receives the sum

She receives this sum via her second input through the Line 14. The size of the register is just sufficient to accommodate this value S (i), which corresponds to the sums indicated above. The output of the register 13 therefore carries the corresponding sum S (i) at each point in time, which corresponds to this point in time is equivalent to. The sequence of values occurring as a function of time at the output of register 13 is shown in FIG.

The values S (i) from the output of the register 13 are transmitted via line 16 to the shift register 17, the output of which in turn is connected to the shift register 19 via line 18. The subtraction circuit 20 receives an input signal from line 14, which forms the output of register 13. The output of the register 17 is connected to each of the subtraction circuits 20 and 21 via line 18. The latter receives over a line 22 also the output signals of the shift register 19. Finally, the output lines of the subtraction circuits 20 are and 21 are connected to the subtraction circuit 23 via lines 24 and 25. The overall circuit therefore allows the determination sudden slope variations of that shown in FIG

309842/0790

FR 971 022

Curve as well as the determination of the curvature of this curve. the Shift registers 17 and 19 have a capacity / which is sufficient to store 20 values S (i). They work in such a way that as soon as a new value is shifted from register 13 into register 17, it is transported further within the latter and, if, for example, the value S (40) has been introduced, the value S (20) is output on line 18. This The value is then fed to the shift register 19, which then outputs the value S (I) on line 22. As a result, there are too at each instant t (i) on lines 14, 18 and 22 the values S (i + 20), S (i) and S (i-20). By means of the subaction circuits 20, 21 and 2 3 the following differences are calculated:

PS (i) = S (i + 20) - S (i)
PKi) = S (i) - S (i-20) D (i) = PI (I) - PS (I)

The result of the successive differences D (i) is output on line 26, while the sign of each difference D (i) is output on line 2 7 is available. The information on the sign of the differences D (i) allows sudden changes in slope the curve in Fig. 4, as well as the curvature of this curve, as it was already in connection with the description for Fig. 5 was explained.

Accordingly, the two decisive factors for determining the stimulation time of the oral cavity are located on lines 26 and 2 7 Information. If the level on line 27 corresponds to a positive sign of the difference D (i), this means that. a sudden change in slope and positive curvature of the curve shown in FIG. 4 and in the maxima of that shown in FIG Curve present. In this case, the gate circuit 28 lets the values D (i) determined by the subtraction circuit 23 through. It for example, let bl, b2, b3 ..., which are the first maxium of the in The curve shown in FIG. 5 corresponds to values D (i). Just like before-

971 022 309842/0790 ......

if shown in FIG. 5, the sign of D (i) is positive and consequently the series of values hl, b2 _r b3 ... is transmitted via line 29 to the digital filter 3O, the latter containing the directional coefficients. The filter 30 is composed in a known manner of four shift registers 31, 32, 33 and 34 as well as a summing circuit 35, which the latter is connected _{to the shift register inputs via lines 36, 37, 3 V} 8 and 39 and to the output of the via line 40 Shift register 34 is connected. The output line 41 of the summing circuit 35 transmits the values D ¹ Ci) produced by smoothing from the values D (i). Accordingly, from the sequence of values bl, b2, b3 ... which are fed to the filter '30, output values b'1, b'2, b'3 ... are obtained on line 41. The curve obtained from the entirety of the points b'1, b'2, b'3 ... by connection is shown in dash-dotted lines in FIG. 5. This curve or straight line was created by smoothing the solid line.

The fourth sequence b'l, b'2, b'3 .. * is then examined in order to determine the largest of the values, which - as has already been explained in the description of the method - is used as a means for determining the point in time of excitation Serves oral cavities. The maximum value is determined with the aid of an overall circuit, which will be described in more detail below. The register 42 successively receives the values b'1, b ^r 2, b'3 ... via the gate circuit 43, which are output via line 41. The subtraction and comparison circuit 44 receives the values mentioned at one of its inputs , while at the other input it receives the output values output by the register 42 via line 45. The output line of the circuit 44 carries information which, as regards the sign of the comparison, is fed via line 46 to the gate circuit 43 and to the inverter 47. The output 50 of the inverter 47 forms an input of the gate circuit 48. This gate circuit also receives an input signal via line 49 which originates from the output 27 of the subtraction circuit 23. This signal is the difference

fr 971 O22 309 8 42/07 9 0

PI (i) - PS (i). During operation, the values are in the register 42 and the comparator 44 entered. First, b'l entered into register 42 as soon as b'2 is established / becomes it is introduced into the comparator 44. At the same time, it is entered into register 42, from which Xtfert b'l transmits to the comparator 44 via line 45. If b'2 is greater than b'l, the output line 46 leads the circuit 44 high level. As a result, the gate circuit 43 lets the value b'3 through, whereupon this is recorded in the register 42. As a result, the value b'2 on line 45 is again fed into the comparator 44, which at the same time receives the value b'3 via line 41. There will be a new sign determination carried out by comparison (b'3 - b'2) and this The process is repeated until the difference between successive values becomes negative. If one considers Fig. 5, it is found that the condition "low level" is given on line 46 for the value b'6 / since b'6 - b'5 are smaller than is zero. At this point in time the gate circuit 43 is no longer open, so that the register 42 does not have any further values can accommodate. This determines the point in time when a sudden change in gradient occurs. Reigns on the line 46 is low, line 50 is high due to inverter 47. Hence there is also an entrance the gate circuit 48 at a correspondingly high level. The second input of the gate circuit 48 to which the line 49 is connected is also at a high level, provided that a curve area is scanned (Fig. 4), the positive curvature (Fig. 5) and consequently on line 27, which forms the output of the subtraction circuit 23, there is also a high level. Up to this point, the entire drawing indicated by dashed lines in FIG. 6b is in the description has not yet been mentioned, so that the third input of the gate circuit 48 has not yet been taken into account can. Since it has just been seen that at the moment of determining the first negative difference - in the example chosen b'6 - b'5 - lines 50 and 49 both on high

FR ₉₇₁₀₂₂ 309842/0790

Level, the output line 51 of the gate circuit will be 48 are also at a high level. The one on this line

A change from low to high level that has occurred is caused by the two bistable multivibrators connected in cascade

52 and 53 transferred. The output of circuit 52 is also fed to circuit 53 via line 54. The line 54 is also connected to one input of the gate circuit 55, the second input to the output of the multivibrator 53 via line 56 is connected. This circuit avoids a continuous high level on line 57, which is the output of the Gate circuit 55 forms as long as there is also a high level on lines 27 and 49 (condition: positive curvature); in fact, the multivibrator 53 causes the line 57 to drop very rapidly at a low level. So you can at the output of the circuit according to the invention in each case at the point in time take a short rectangular pulse from the stimulation of the oral cavity. A line 58 connects the output of an inverter 59 with the register 42 and the bistable multivibrators 52 and 53. This line is at a low level as long as the difference values PI (i) - PS (i) are positive. As soon as this difference becomes negative, the level of conduction increases 49, whereupon line 58 again leads to a low level, whereby the multivibrators 52 and 5 3 as well as the register 42 deferred and therefore ready for the next voiced period.

The circuit just described in connection with FIG. 6 for carrying out the method according to the invention can also be carried out the circuit parts shown in dashed lines are perfected. These circuit parts are intended to create a variable detector threshold are provided. That such a variable threshold is useful becomes apparent on closer inspection of Fig. 5. The totality of the values D (i) shown in this figure is not only positive when suddenly in FIG Incline changes occur. It can also be seen that Fig. 4 has sections of positive curvature of short duration,

FR ₉₇₁ 022 309842/0790

which occur due to weak slope changes. This type of variation is only random and may be that shown in FIG Do not allow the system to emit any false signals indicating the occurrence of excitation times.

The circuit for the threshold variation basically consists of a threshold register 60, which receives input signals from a Gate circuit 61 receives, and a subtraction comparison circuit 62, one input with the output of the register 60 and the other input with the output of the register 42 is connected via line 45a. The output of the comparator 62 delivers its signals via line 6 3 to the gate circuit 48 and via line 63a to the gate circuit 61.The latter has a second input, which via line 46b, the inverter 64, Line 46a and line 46 are connected to the output of the comparison circuit 44. A third input of the gate circuit is connected to the output of register 42 via line 45b.

In operation, the register 60 is a threshold of the value S (Fig. 5) fed. This value is selected so that the sections of positive curvature of the curve in FIG. 4 are not taken into account and via line 51 (Fig. 6b) as soon as they are detected. As soon as the series of values b '1, b · 2, b'3 ... from the Register 42 is transferred to the comparison circuit 44, these values are successively fed to the comparator 62 z via line 45a. So in the event of a spurious occurrence positive curvature when comparing the value b'l in the circuit 62 via line 6 3a output a low level because b'l - S is less than zero. As a result, the level of the first input of the gate circuit 61 is then low. Hence will the content of register 60 is not changed. At the same time, the level on line 6 3 is also low, so that, as a result, as soon as the level drops on line 56, no information whatsoever. tion can run via the gate circuit as the second input this gate circuit low level, the third input (line

_FR 97X022. 309842/0790.,

tung 49) has a high level. As a result of this measure, no section of positive curvature that is lower than the threshold S is determined and transmitted via line 51. In the case of an excitation time, the values of the section of the positive curvature will be above the threshold S, so that, for example for the value b'l, the line 6 3a is at a high level, since b'l-S is greater than zero. Throughout the rise of the smoothed curve in FIG. 5, the level on line 46a will be at the top and on 46b at the bottom. As a result, no value whatsoever can be transferred from register 42 via line 45b and gate circuit 61 into register 60. However, as soon as the first value (b'6, for example on FIG. 5) is lower than the previous value, the comparator 44 will drop the level of line 46 because b'5-b'6 is less than KuIl. Thereupon nothing more can be transferred to the register 42. Likewise, line 46a will be at a low level , so that line 46b is at a high level and the value b'6 can be brought from register 42 into threshold value register 60 via gate circuit 61. This new value b'6 will be the decisive threshold for determining the subsequent excitation time. If there is a risk that the changes in the fundamental frequency are so great that such a threshold turns out to be too high, it is sufficient, for example, to place a divider between registers 42 and 60 that halves the threshold value. The threshold value contained in register 60 will change at each point in time of excitation of the dog cavities, since the maximum of the curve shown in FIG. 5 changes with each excitation.

Therefore, at the time of excitation of the oral cavities, conduction 50 will have a high level as soon as the first negative result of comparator 44 appears (e.g., b'6-b'5 in Figure 5); management 49 will lead to a high level, since the result of the difference PI (i) - PS (i) in the subtraction circuit 23 is positive, likewise, lead 63 will lead high level as the result

fr 9 71 022 309842/07 90

of the comparison between the values from register 42 and the threshold is positive. As a result, as shown above was energized - line 52.

Also on the output line 5 7 of the overall circuit like them is shown in Fig. 6, is accordingly the excitation times information characterizing the oral cavities occur. With every excitation a square-wave pulse occurs which leads to Measurement of the fundamental frequency of speech vibrations can be used, e.g. by measuring the time interval between two successive pulses determined. Suitable circuits for measuring such time intervals are known in the art Technology known.

309842/0790

FR 9 71 022

Claims (8)

PATENT CLAIMS 1. j Method for determining the basic speech frequency by determining the times of excitation of the oral cavities for voiced sounds, characterized in that an analog signal representing the speech oscillation is sampled in sections and the sampled value is digitized, that for each scan the magnitude of the difference d (i) between that for this scan and that for the previous one Sampling determined value is determined that the amounts of the differences d (i) between successive Samples according to the formula i - 1 Σ d (i-l) be added to a sum S (i), which is the amount of indicates the differences that have occurred up to this sampling in order to obtain a signal whose mean value curve of the successive sums S (i) of the amounts sudden changes in the course at the times of the excitation showing the oral cavities, and that these excitation times are determined from the course of the curve. 2. The method according to claim 1, characterized in that the digitized scanning signals are PCM-modulated. 3. The method according to any one of claims 1 or 2, characterized in that the determination of the excitation times from the course of the curve includes the following process steps: a) the change in slope of the mean value curve of these sums is determined from the sums S (i) of the amounts, by calculating for each sum: FR 972 022 309842/07 90 a first difference between the value of the sum S (i) and that of the sum S (i- η), a second difference between the value of the sum S (i + n) and that of the sum S (i), a third difference between the results of the first and second difference formation, where the The result of this third difference formation is the information relating to a sudden change in gradient in the form of a sign change from negative to positive; b) within the group of positive results of the third subtraction, the result is determined which has the greatest value; c) depending on the result of the third difference formation, which positive sign and within a predeterminable one Group has maximum size, a signal indicating the excitation of the oral cavities is generated. 4. The method according to claim 3, characterized in that the number "n ^;: = 20 is selected. 5. The method according to claim 3, characterized in that a stimulation of the oral cavities characterizing Signal is only generated when the amount of the third subtraction is a function of the speech signal the adjustable threshold. 6. Device for performing the method according to at least one of claims 1 to 5, characterized by an encoder that picks up the speech signal and transforms it into a sequence of digital values, a subtraction circuit for calculating the difference between successive sampling results, a device for determining the amount of the difference between successive scanning results, an adder circuit that calculates the difference fr 971 022 3098 427 07 90 values each to the sum of the previously calculated difference values, whereby the successive sums over a feedback circuit between the output and input of the adder circuit, and a curvature detector that sums the successive amounts which picks up amounts and emits a signal whenever a positive curvature occurs. 7. Apparatus according to claim 6, characterized by a maximum detector to which the sums are entered correspond to positive curvature, and the sum entered from the group determines the one with the greatest value, whereupon a signal that is only issued once for a group is generated. 8. The device according to claim 7, characterized by a threshold value detector, which the successive sums of the amounts receives and emits a signal for each sum lying above a predeterminable threshold, which the transfer of this amount above the threshold in the maximum detector. _FR 971 022 309842/0790 Blank page