EP0052041A1 - Method and device for pitch period determination - Google Patents

Abstract

The invention relates to the real-time detection of a voiced signal and to the measurement of the corresponding melody frequency in a speech signal. The method uses a set of data characteristic of the speech signal, supplied by processing circuits (10, 20, 30): measurements of the time intervals between zero crossings and measurements of the energy in the alternations of this signal. . The test procedure implemented by a microprocessor (40) retains alternations whose energies exceed thresholds characterizing the start of the melody period. These thresholds are predetermined for the first two successive sums retained, then depend on the energy values of the previous half-waves retained differently depending on whether the voiced character of the signal is acquired or not. Additional tests are implemented to minimize detection errors. Application, in particular, to the processing of speech signals for vocoders.

Description

The invention relates to the analysis of speech signals and more particularly to a method for detecting the melody frequency of voices heard in a speech signal and to a device intended for implementing this method.

In speech, voiced sounds are made up of vowels or liquid or sonic consonants and have well-defined spectral properties that are not found in unvoiced sounds made up of deaf consonants. These voiced sounds have a generally greater amplitude than unvoiced sounds and a very marked periodicity in the speech signal. The value of the frequency corresponding to this periodicity (linked to the vibration of the vocal cords) is the melody frequency included, depending on the people between 60 and 300 Hz.

This melody frequency is a fundamental parameter of speech which is evaluated in most vocoders, the quality of the detection of this frequency directly influencing the quality of the speech restored after decoding.

• Les premiers procèdent par analyse systématique du signal de parole, analyse du spectre ou autocorrélation, et mettent en oeuvre un volume de calculs généralement trop important pour conduire à des réalisations en temps réel au moyen de systèmes relativement simples.

Analysis of the state of the art makes it possible to distinguish two classes of methods and devices for detecting the melody frequency:

• The former proceed by systematic analysis of the speech signal, spectrum analysis or autocorrelation, and use a volume of calculations generally too large to lead to real time realizations by means of relatively simple systems.

The second, of temporal type, seek a periodicity directly on the temporal signal. They generally use a reduced set of data, for example time intervals between zero crossings (or between signal maximums), or counting the signal's zero crossings for a determined duration; decision criteria. take into account the properties observed on the speech signals. Calculations are more reduced with this type of detection, but the corresponding detection devices do not perform very well in the presence of noise and during the transitions between voiced signal and unvoiced signal. A method and a device for detecting the melody period using, as a data set, the measurements of the energy in the successive arches of the speech signal have also been described. This device benefits, compared to more common time type devices, from better noise immunity and from a more selective voicing criterion which limits false detections. However, detection requires the cutting of the signal into frames of fixed length, the calculations allowing the recognition of a voiced sound can only be carried out with a delay frame. Furthermore, there is a risk of detection of the frequency twice the melody frequency because the criterion making it possible to avoid such detection is only effective in the middle of a neighboring segment. Finally, the division of the signal into frames of fixed length which are not linked to the content of the speech signal harms the quality of the measurement, in particular during the transitions between voiced signal and unvoiced signal.

The subject of the invention is a method of real-time detection of the melody frequency in speech, of the time type, using measurements of the energy between zero crossings, as well as measurements of the time intervals between these passages at zero. zero. The method avoids false detections, in particular the detection of the double frequency, has good immunity against noise, and moreover does not significantly increase the complexity of the device intended for its implementation compared with known devices.

According to the invention, a method of detecting, in real time, the melody frequency in speech from a reduced set of data measured on this signal, is mainly characterized in that this set is composed of a measurements. (i variable) of the energy in the successive alternations of this signal and of measurements t _. associated with the durations of these alternations, and in that the test procedure implemented on these data includes an acquisition phase during which a first series of tests confers, when it is verified, the acquired character in the neighborhood and results in the calculation of a first melody period value, and a holding phase during which a second series when confirmed, confirms the acquired character of the voicing and results in the updating of the value of the melody period, this second series of tests being repeated as long as the acquired character of the voicing is preserved and a new phase d acquisition being implemented when the acquired character of the settlement is lost.

The invention also relates to a device intended for the implementation of this melody frequency detection method.

• - La figure 1 est le schéma du dispositif de détection suivant l'invention.
• - La figure 2 représente un exemple de segment de signal voisé, en début de parole ;
• - Les figures 3 et 4 représentent d'autres exemples de segments de signaux voisés, en début de parole, risquant de conduire à de fausses détections ;
• - Les figures 5, 6, 7 et 8 représentent des diagrammes séquentiels des différentes phases du procédé de détection de la fréquence de mélodie ;
• - La figure 9 représente un exemple de segment de signal voisé en cours de parole ;
• - La figure 10 représente quelques configurations particulières de l'énergie dans les alternances du signal voisé.

The invention will be better understood and other characteristics will appear from the following description with reference to the appended figures.

• - Figure 1 is the diagram of the detection device according to the invention.
• - Figure 2 shows an example of a voiced signal segment, at the start of speech;
• - Figures 3 and 4 show other examples of segments of voiced signals, at the start of speech, likely to lead to false detections;
• - Figures 5, 6, 7 and 8 show sequential diagrams of the different phases of the melody frequency detection method;
• - Figure 9 shows an example of signal segment voiced during speech;
• - Figure 10 represents some particular configurations of the energy in the alternations of the voiced signal.

The melody frequency detection method uses, to search for the presence of a voiced signal, and to measure the corresponding melody period, a reduced set of data constituted as follows: the speech signal is first filtered by a low pass filter with cutoff frequency f = 800 Hz; this filtered signal is then sampled. Then, from the filtered and sampled signal, the data useful for detection is obtained by detecting the zero crossings of this signal and "integration" between consecutive zero crossings; the corresponding sums give an estimate of the energy in each positive or negative alternation of the signal. The time intervals t (i variable) between zero crossings are stored in a first table and the corresponding sums a. are stored in a second table. These two tables are established in real time. Finally, from this reduced data set, the discrimination between voiced and unvoiced signal segments is obtained by following different criteria depending on the phase: during a so-called "acquisition" phase, the device follows a first test procedure according to a first set of criteria, while during a second phase called "holding", the device follows a second test procedure according to a second set of criteria. When, during this holding phase, the test indicates that the voiced character of the signal is lost, a new acquisition phase begins.

During these procedures, additional protection tests are introduced to avoid false detections.

The device for detecting the melody frequency intended for the implementation of this process very succinctly described above is represented in FIG. 1. This device comprises an analog processing circuit 10 with two inputs, E ₁ and E ₂ respectively. suitable for connection to a microphone and line output amplifier. This analog processing device comprises: an amplifier 11 whose input is connected to the input E ₁ , a second variable gain amplifier 12 whose input is connected to the output of the amplifier 11 on the one hand, and directly at the entrance E ₂ on the other hand. This amplifier 12 has its output connected to the input of a low pass filter 13 whose cutoff frequency is, as indicated above, f = 800 Hz. The output of this low pass filter 13 is connected to the input signal from an analog to digital converter 20. This converter also includes a clock input H setting the frequency of the samples taken from the analog signal. This clock input is coupled to the output of a clock 1, delivering a signal at the frequency _H. , via a frequency divider 2 whose output delivers a clock signal H.

As an example, the converter can deliver the digital values of the samples in the form of words of 8 binary elements, one binary element being reserved for the sign of the sample.

The device also includes a set of digital circuits 30 and a microprocessor 40. The digital processing circuits are connected on the one hand to the output of the analog digital converter and to the clock output H, on the other hand to the microprocessor. These circuits are more precisely: an accumulator 31 intended to add up the values of the successive samples which are supplied to its multiple signal input in the form of words of 8 binary elements by the converter; the sums are supplied in the form of words of 12 binary elements of which only the 8 most significant are transferred to the microprocessor 40 to be put in memory. A zero detector 32 whose signal input receives the binary element characteristic of the sign of the samples supplied by the converter This zero-crossing detection circuit is a simple logic circuit which compares the sign of the sample present at the output of the converter with the sign of the previous sample stored in this circuit. This detector has an output which supplies an interrupt pulse 1 to the microprocessor 40. The zero detector also includes a clock input H. The digital processing circuits also include a counter 33 having an input connected to the output H of the divider 2 and a reset input, RESET; this counter allows the microprocessor to be given measurements of the time elapsed between two resets. Finally, these circuits 30 also include a frame counter 34, the input of which is also connected to the output H of the divider 2 and the output of which supplies interrupt pulses 1 to the microprocessor, for displaying and storing the results obtained during a test procedure; this circuit also has a reset input, not shown.

The microprocessor 40 comprises: an MPU processing unit, 41; a random access memory RAM, 44, the content of which can be modified and read at will, and which makes it possible to store the values of the sums a and of time intervals t. as well as the intermediate values useful for detection; a read only memory, PROM, 45 in which the test program making it possible to determine the melody frequency is recorded; a display device 46 displaying when appropriate the detected values. These elements 41 to 46 are connected to each other and to a PIA interface circuit, 42 via a bidirectional link bus 47, the interface circuit also being connected by bidirectional data buses 35, 36, 37 to the accumulator. reader 31 and counters 33 and 34. The address bus and address decoders have not been shown in this diagram for simplicity.

Data acquisition from the filtered and sampled signal is obtained by the digital processing circuits in connection with the microprocessor in the following manner.

As indicated above, an interrupt pulse I _e supplied by the zero crossing detector 32 to the interface circuit 42 controls the transfer of the content a. of the accumulator 31 in a first memory table 44 (via the link bus 35 between the accumulator and the interface circuit 42, the interface circuit 42 and the link bus 47 between the circuit interface and memory 44), and the transfer of content t. of the counter 33 in a second table of the memory 44 (via the link bus 36, the interface 42, and the link bus 47).

After these transfers, the interface circuit 42 controls the resetting to zero of the accumulator 31 and of the counter 33. The test procedure takes place in real time, which makes it possible to limit the size of the RAM required, the two tables each having, for example, 256 memory boxes, and the new data being rewritten on the old already tested. For this, reading and writing indices of these tables are provided and an additional test, not detailed here, makes it possible to ensure in reading that the reading index does not exceed the writing index (in order to do not reuse values already tested) and in writing that the writing index does not exceed the reading index (which would cause the loss of untested values).

The test procedure implemented from this data takes into account the form of the speech signal and takes place from a test program recorded in the program memory 45. The test procedure characteristic of the detection method of the melody frequency will be explained in detail below in conjunction with the signal diagrams of FIGS. 2, 3, 4 and 9 and with the sequential diagrams of the test program represented in FIGS. 5 to 8.

FIG. 2 shows an example of a signal segment seen at the start of speech. This signal consists of positive and negative alternations whose maximum amplitude, duration, and energy are variable. The voiced signal is characterized by the fact that two successive alternations sives (of different signs) having energies higher than those of the alternations of the same sign preceding and following, can be detected in this signal. These particular alternations are repeated at an almost constant period, called the melody period.

• - pour la phase d'acquisition du signal voisé, à détecter trois groupes de deux alternances successives, dont les énergies (a_lp et a_1n, ^a ₂p et a_2n, a et a ) et la configuration dans le temps répondent à un ensemble de critères ; lorsque ces critères sont vérifiés, le caractère voisé du signal est acquis, trois débuts de période de mélodie ayant été trouvés, et une première valeur de la période de mélodie est calculée ;
• - pour l'entretien du caractère voisé en cours de test, il est vérifié que des alternances ayant des énergies dépassant des seuils définis en fonction des valeurs d'énergie des alternances précédentes retenues sont présentes dans le signal à des intervalles.de temps voisins de la période de mélodie initiale calculée ; la valeur de cette période est alors actualisée.

In general, the detection method according to the invention consists:

• - for the acquisition phase of the voiced signal, to detect three groups of two successive half-waves, whose energies (a _l p and a _1n , ^a ₂ p and a _2n , a and a) and the configuration over time respond to a set of criteria; when these criteria are verified, the voiced character of the signal is acquired, three beginnings of melody period having been found, and a first value of the melody period is calculated;
• - for the maintenance of the voiced character during the test, it is verified that alternations having energies exceeding thresholds defined according to the energy values of the previous alternations selected are present in the signal at intervals of time close to the initial melody period calculated; the value of this period is then updated.

When the maintenance test for the voiced character is not verified, a new acquisition procedure is initiated.

• - atest = 0 : début de la phase d'acquisition ; aucun test n'est vérifié ;
• - atest = 1 : la première alternance susceptible de caractériser le début de la première période voisée est retenue ;
• - atest = 2 : l'alternance successive de la première période voisée est retenue ;
• - atest = 3 : la première alternance susceptible de caractériser le début de la seconde période voisée est retenue ;
• - atest = 4 : l'alternance successive de la seconde période voisée est retenue ;
• - atest = 5 : la première alternance susceptible de constituer le début de la troisième période voisée est retenue ;
• - atest = 6 : l'alternance successive de la troisième période voisée est retenue ;
• - atest = 7 : la première alternance susceptible de constituer le début d'une n^ième période voisée est retenue ;
• - atest = 8 : la seconde alternance de la n^ième période voisée est retenue.

An "atest" pointer is provided for switching in the various elementary tests, the state of this register being characteristic of the progress of the detection:

• - atest = 0: start of the acquisition phase; no test is verified;
• - atest = 1: the first alternation likely to characterize the start of the first voiced period is retained;
• - atest = 2: the successive alternation of the first voiced period is retained;
• - atest = 3: the first alternation likely to characterize the start of the second voiced period is retained;
• - atest = 4: the successive alternation of the second voiced period is retained;
• - atest = 5: the first alternation likely to constitute the start of the third voiced period is retained;
• - atest = 6: the successive alternation of the third voiced period is retained;
• - atest = 7: the first alternation likely to constitute the start of a n ^th voiced period is retained;
• - atest = 8: the second alternation of the ^nth voiced period is retained.

Before being able to make a first measurement of the melody period, the first test makes it possible to find two successive alternations of contrary signs, whose energies exceed given thresholds, S _l p and S _1n , the beginning of the first of these two alternations which can constitute the beginning of the melody period when the following tests are also verified.

The flowchart of the corresponding test program is shown in FIG. 5, this test being designated by test I below. After an initialization phase of all the variables, the reading index of the tables of the memory 44, i, is incremented. Then a sum a _i and the corresponding time interval t. are read from memory. A test on the sign of the sum a _i then makes it possible to test the value of the sum a _i with respect to the thresholds defined above, S _1p and S _1n . When this test is negative, the "atest" pointer is set to zero. A new reading of the variables is then undertaken. When one of these tests is positive, the corresponding value of the sum a _i is loaded into a register and constitutes the value a _l p or a _1n , according to the sign of the sum, value likely to constitute the first sum of the beginning of the melody period. The value of the corresponding time interval t. is loaded into a register and constitutes a value t _p or t _n according to the positive or negative sign of the corresponding sum. This sign is also kept in memory in a "first sign" register in order to subsequently search for the start of the following periods only on sums of the same sign. In addition, the value of the read index, i, is also kept in memory in an "initial" register for possible later use. When this first sum is detected, the pointer "atest", initially at zero, is incremented by 1. A test on the value of this pointer with respect to 2 is then implemented before searching for the following sum making it possible to characterize completely the beginning of the melody period. This second sum must exceed the corresponding sign threshold. If it does not exceed the threshold, atest is reduced to zero and the test resumes with the following sum. When this second sum of opposite sign is also found, the point "atest" is again incremented and the test of the value of this pointer with respect to 2 is then checked. The first two values a _l p and a _1n greater than the thresholds S _l p and S _1n , are then found.

The test procedure then continues to search for the start of the second melody period, at the same time as the time intervals between zero crossings are added to allow a value of the melody period to be determined later.

FIG. 6 represents the test procedure making it possible to determine the start of this second period and the first values of time intervals between the sums with the same sign of the first two groups. As before, the reading index is first incremented, then a sum and the corresponding time interval, a _i and t _i , are read from the memory. The sign of the sum a _i is tested and two parallel branches are possible according to the sign of the sum. At the beginning of each branch, a check of the alternation of the sums sign is carried out. When this alternation condition is not verified, a referral makes it possible to change branches after correction of the overflow. These branch changes are shown in dotted lines in the figure. When the alternation condition is well verified, the time interval, called "current", t _12p or t _12n between the sum of the first group, a _1n or a _1p of the same sign as the sum a. being tested and the start of the alternation corresponding to this sum being tested is calculated as follows: t ₁₂ p new value is equal to t _12p old value plus t _p plus t. Then the value of the time interval between zero crossings, t _i , corresponding to this sum under test is stored in a register (t or t _n according to its sign) which makes it possible to calculate the time interval current.

• - Lorsque cet intervalle de temps courant est supérieur à T_M , les premières alternances retenues, correspondant aux sommes a_lp et a_ln' ne pouvaient pas correspondre à un début de période de mélodie et le programme est réaiguillé vers le premier test, après réinitialisation des valeurs de temps courant et de la variable "atest", et incrémentation de la valeur du registre "initial" mise en mémoire.
• - Par contre, lorsque la valeur de temps courant ne dépasse pas la période maximale T_M , la valeur de la somme correspondante a_i est comparée à un seuil fonction de la valeur de la première somme de même signe retenue.
• - En effet, les sommes du deuxième groupe permettant de caractériser le début de la deuxième période ont des valeurs situées dans un voisinage des valeurs des premières sommes retenues. Dans l'exemple mis en oeuvre, le test est effectué par rapport à des valeurs de seuil :
  
  c'est-à-dire que ces valeurs de seuil sont égales aux plus grandes, en valeur absolue, des deux valeurs ¾a_1p et S_lp pour la première, et ¾a_1n et S_1n pour la seconde :
• - Lorsque le résultat de ce test est négatif, un test sur la valeur du pointeur "atest" est réalisé, de façon à incrémenter l'indice de lecture i et à calculer directement la valeur du temps courant sans effectuer de test sur la valeur suivante de la somme ; en effet cette somme suivante ne peut pas constituer le début de la seconde période compte tenu de son signe (atest est alors égal à 2).
• - Par contre, lorsque le résultat du test sur la valeur de la somme est positif, la valeur de la somme correspondante peut constituer la première somme a₂p ou a_2n du deuxième groupe, correspondant au début de la seconde période, et la variable "atest" est incrémentée. Seule la première des deux sommes a été trouvée et un test du pointeur "atest" par rapport à "4" permet d'engager une nouvelle procédure de test puisque, à ce moment atest = 3. Les mêmes tests sur la valeur suivante permettent soit de vérifier les mêmes critères, au signe près, sur la somme suivante, soit de retourner au début du test I après réinitialisation lorsque le critère de durée par rapport à la période maximale n'est pas vérifié ou au début du test II lorsque le critère de durée est vérifié mais pas le critère sur la valeur de la somme. Alors atest est ramené à la valeur 2 car la somme précédente retenue ne peut constituer le début de la deuxième période puisque la somme suivante ne peut être retenue.
• - Lorsque les deux valeurs successives ont été trouvées le pointeur "atest", qui est à nouveau incrémenté, a alors la valeur quatre ; ce qui indique que le second test est terminé. Une dernière comparaison de la différence entre la valeur du temps courant t_12p et la valeur du temps courant t_12n (chacune de ces deux valeurs pouvant donner une valeur de la période de mélodie) permet de vérifier que cette différence est inférieure à un écart de temps déterminé, t_pn; ; ce test permet de s'assurer que le signal est suffisamment régulier pour qu'on puisse caractériser une période de mélodie et permet d'éliminer les erreurs évidentes. t_pn peut être choisi égal à 256 µS (soit 20 échantillons à 7,8 kHz). Cet écart entre t₁₂p et t_12n est aussi l'écart entre les premières alternances des deux groupes retenues.

The value of this current time interval, either t _12p or t _12n , is then compared with the maximum value T _M of the melody period; this value T _M being a prerecorded data:

• - When this current time interval is greater than T _M , the first half-waves selected, corresponding to the sums a _l p and a _{ln '} could not correspond to the start of the melody period and the program is redirected to the first test, after reinitialization of the current time values and of the "atest" variable, and incrementation of the value of the "initial" register put in memory.
• - On the other hand, when the current time value does not exceed the maximum period T _M , the value of the sum corresponding to _i is compared with a threshold depending on the value of the first sum of the same sign retained.
• - Indeed, the sums of the second group used to characterize the start of the second period have values located in the vicinity of the values of the first sums used. In the example implemented, the test is carried out with respect to threshold values:
  
  that is, these threshold values are equal to the larger, in absolute value, of the two values ¾a _1p and S _l p for the first, and ¾a _1n and S _1n for the second:
• - When the result of this test is negative, a test on the value of the pointer "atest" is carried out, so as to increment the reading index i and to directly calculate the value of the current time without performing a test on the following value of the sum; indeed this next sum cannot constitute the beginning of the second period given its sign (atest is then equal to 2).
• - On the other hand, when the result of the test on the value of the sum is positive, the value of the corresponding sum can constitute the first sum a ₂ p or a _2n of the second group, corresponding to the start of the second period, and the variable "atest" is incremented. Only the first of the two sums has been found and a test of the pointer "atest" compared to "4" allows to initiate a new test procedure since, at this time atest = 3. The same tests on the following value allow either to verify the same criteria, except for the sign, on the following sum, either to return to the start of test I after reinitialization when the duration criterion in relation to the maximum period is not verified or to the start of test II when the criterion duration is checked but not the criterion on the value of the sum. Then atest is brought back to the value 2 because the previous sum retained cannot constitute the beginning of the second period since the following sum cannot be retained.
• - When the two successive values have been found, the pointer "atest", which is again incremented, then has the value four; which indicates that the second test is complete. A final comparison of the difference between the value of the current time t _12p and the value of the current time t _12n (each of these two values can give a value of the melody period) makes it possible to verify that this difference is less than a difference of determined time, t _pn ; ; this test ensures that the signal is regular enough to characterize a melody period and eliminates obvious errors. t _pn can be chosen equal to 256 µS (i.e. 20 samples at 7.8 kHz). This difference between t ₁₂ p and t _12n is also the difference between the first half-waves of the two groups selected.

Test II is then finished and test III, making it possible to search for the start of the third voiced period, can then begin.

FIG. 7 and FIG. 8 represent the test III which makes it possible, from the first and the second group of sums retained, to search for the third group of sums which can characterize this beginning of the third period; the acquisition of all the values of sums retained and the values of corresponding time intervals indicates that the voiced character of the signal is acquired and makes it possible then to calculate a value of the melody period which takes account of the time intervals between beginning of period.

Before describing the course of test III, the various tests carried out there are presented below.

As for the first two tests, the values of sums a are compared with threshold values; these threshold values S ₃ p and S _3n depend on the preceding sums of the same sign retained as follows:

In addition, as in the first two tests, the current time intervals (between the sum retained with the same sign characterizing the start of the second period and the sum under test), t ₂₃ p and t _23n , are compared with duration values defined as follows:

T _m characterizing a minimum melody period and e a maximum tolerated time difference, are prerecorded data. The first two tests, (1) and (2) on the value of the current time verify that the current time is long enough to constitute a melody period. The third is on the contrary intended to ensure that this current time value is not too large.

• 1 ^a ₂ - a₁| + |a₂ - a₃| ≤ q_max, q_max étant une donnée préenregistrée, les sommes a₁, a₂ et a₃ étant affectées des indices p ou n selon la branche du test en cours.

An additional condition of monotony in the progression of the sums is also required in order to avoid detecting the half-period. FIG. 3 represents a segment of voiced signal which, if this additional condition was not imposed, would lead to a detection of double frequency by retaining the sums indicated at _1p and at _1n , a2p and ^at _2n , and ^at ₃ p and a _3n , whereas a ₂ p and a _2n correspond to half-cycles in the middle of the melody period. This condition of monotony is:

• 1 ^to ₂ - to ₁ | + | a ₂ - a ₃ | ≤ q _max , q _max being a prerecorded datum, the sums a ₁ , a ₂ and a ₃ being assigned indices p or n depending on the branch of the test in progress.

Furthermore, to protect against acquisition errors likely to occur in a voiced signal configuration such as that shown in FIG. 4, where the midpoints of the period are retained instead of the start of the period (which can lead to a loss of synchronization in the middle of a segment or at the subsequent detection of half a period, double melody frequency), another additional condition is imposed: this condition is that values of sums a _i rejected are not greater than the previous sums same sign retained. For the voiced segment represented in FIG. 4, a _l p, ^a ₂ p and a _1n , ^a _2n would normally be retained, but the condition described above implemented in test III will not be verified because a'3p, rejected by the duration criteria, is greater than a ₂ p retained. In this case, these are the values a 'which correspond to the start of the period, and should have been retained, and all of the research has been resumed since the start of test I.

The flow of test program III is shown in Figures 7 and 8. These figures also show the flow of test IV used when the voiced character of the signal is acquired to verify that the voiced character is maintained. Indeed, the sequences corresponding to the third test, test III, and to the fourth test, test IV, differ only by internal connections which depend on the value of the pointer "atest", and by the values of thresholds to which the sums are compared at. being tested. These threshold values and the corresponding test are defined as follows:

). De plus ces seuils, qui pourraient devenir trop faibles ou même changer de signe en fin de segment voisé sont bornés par les seuils prédéterminés S_1p et S_1n. Enfin, et surtout, lorsqu'une seule de ces conditions est vérifiée, le caractère voisé du signal continue à être considéré comme acquis pourvu que les conditions sur les intervalles de temps soient vérifiées. En effet, si cette disposition n'était pas adoptée, une diminution de l'énergie dans une seule des alternances du signal voisé pourrait conduire à décider que le caractère voisé est perdu, ou à détecter une période de mélodie double alors que la présence de la somme de signe contraire suffit à maintenir une décision correcte. Les tests sur les intervalles de temps sont exactement les mêmes que ceux mis en oeuvre dans le test III.These conditions are close to those of test III but the tolerance on the thresholds is wider (¾ and .no more

). In addition, these thresholds, which could become too low or even change sign at the end of the voiced segment, are bounded by the predetermined thresholds S _1p and S _1n . Finally, and above all, when only one of these conditions is verified, the voiced character of the signal continues to be taken for granted provided that the conditions over the time intervals are verified. Indeed, if this provision were not adopted, a reduction in the energy in only one of the alternations of the voiced signal could lead to decide that the voiced character is lost, or to detect a period of double melody while the presence of the sum of the opposite sign is sufficient to maintain a correct decision. The tests on the time intervals are exactly the same as those implemented in test III.

Certain branches of the sequence are common to tests III and IV. In addition, those which, after testing the "atest" pointer correspond to atest = 4 or 5 are test branches III and those which correspond to atest = 6 or 7 are test branches IV. To simplify the figures, only the branches relating to positive sums have been shown. Non-detailed symmetrical negative branches correspond to the positive branches detailed in these figures. They differ only in the index of variables and thresholds (n instead of p and the direction of comparison for the test with respect to the threshold).

The diagram shown has a first entry, start of test III, l, when the voiced character is not acquired; another entry, 2, start test IV, allows when the voiced character is acquired, to reset the test variables and to update the previous values retained at ₂ , ^a ₃ and ^t ₂₃ in a ₁ , a ₂ and t ₁₂ ( for positive and negative values) when the search advances by a period. This offset appears in FIG. 9 which represents a voiced signal segment tested during a maintenance phase (the old values are put in parentheses above the new values). Then in a branch common to test III and test IV, the reading index is incremented; the sum a and the time interval t _i are read from the memory. A test on the sign of the sum allows to choose the branch of the suitable test procedure. In the following it is assumed that the first sum retained in test I is positive, that is to say that the first sum tested in test III is also positive. The current time interval t ₂₃ p is calculated and this time interval is tested.

If this interval is too short to be able to correspond to a melody period (t _23p <t _12p - e or t _23p <t _min ) and the sum under test is however greater than a _2p , the first sums retained were bad (Figure 4) and all the search is reset from test I, because the voiced character was not acquired (atest = 4). On the other hand, if this sum is not greater than a ₂ p, which is the normal case, the current time is updated and the reading index is incremented for the reading of a time value t _i , set memory in the register for calculating the current time, and the current time is calculated. Then the test is resumed at the level of the first increment of the reading index (point 3), which makes it possible to test the following alternation of the same sign.

If the time interval t ₂₃ p is not too short but on the contrary that it exceeds the value t _12p + e, all the variables are reset and the search is resumed from test I because the beginning of a third period was not found.

If the time interval t ₂₃ p is not too short and at the same time it does not exceed the value t _12p + e, this interval may correspond to the melody period. Consequently, the test on the value of the sum with respect to the threshold S _p (S _3p in this test III) is carried out. If this test is not verified, the value of the current time is updated, the reading index is incremented and the time interval t _i which corresponds to it is stored in memory. The following alternation test of the same sign is undertaken by returning to point 3 of the test.

When the sum a _i is greater than the threshold, the first sum a ₃ of the third period (a ₃ p in the example used, "first sign" being positive) is found provided that the criterion of monotony between the values a ₁ , a ₂ and a ₃ indicated above be also checked. Whereas a _3p = a .. Otherwise the test is repeated at the start of test I.

The value of atest is then incremented (atest = 5) (Figure 8), then this value of atest is compared with 6 and 8. As test III is not finished, this test is negative. By repeating test III in point 3, it remains to be verified by the other branch (negative branch in the example used) that the energy in the following alternation also exceeds the threshold associated with it so that this sum can be retained like the second of the third period. For this the same tests on the time interval are carried out. When this interval (t _23n in the example used is too short and the sum a. _Under test is greater than a _2n , the whole search is reset from test I, because the voiced character was not acquired (atest = 5) On the other hand if this sum is not greater than a _2n , the current time is updated, the value of atest is reduced to 4 and test III resumes at point 3 on the following sum to start the search again from the start of the third period.

If the time interval (t _23n ) exceeds the maximum value, the search is reset from test I. Similarly, if the value under test does not exceed the corresponding threshold S _3n (as during a failure on the first two duration tests) the current time is calculated, the time interval t. is put in memory and atest is brought back to 4 in order to cancel the previous sum retained and to start again the search for the beginning of the third period. After the test of the monotony criterion (return to the start of test I if this criterion is not verified) atest being equal to 5, a "first sign" test is carried out. This test makes it possible to ensure that the value about to be retained (a _3n in the example retained) is indeed of opposite sign compared to the first sum retained.

Then as before, the pointer is incremented and atest atest is then equal to 6. The second cycle of the ^3rd period is found. The same criterion as in test II on the difference of the periods starting with alternations of contrary signs is then checked to eliminate the obvious errors: t _23n - t ₂₃ p | <t _pn - (4). If this condition is satisfied, the value of the melody period is calculated:

A new test, which is then the fourth test, is carried out (by switching to entry point 2, start of test IV) to find out whether the voiced character of the signal is maintained.

If condition (4) over the time intervals is not checked, the value of atest is reduced by 2 and the test is repeated at point 3.

For the fourth test, the basic procedure is similar to that of the third test but additional branches are provided so that particular signal configurations which do not satisfy all the conditions indicated above (which would lead for test III to a definitive rejection of the alternation considered) are interpreted as voiced signals when the voiced character was previously acquired. These particular configurations are shown in FIG. 10. They are such that one of the alternations of the start of the n ^th period, the first or the second, which may be positive or negative, has an energy below the threshold S ₄ p or S _4n fixed, the other exceeding the corresponding threshold. For each of these configurations, the values of the different variables used for the procedure are shown in Figure 10 next to the corresponding configuration.

ou

alors que le signe, respectivement négatif et positif, ne correspond pas à celui attendu, "signe premier" étant respectivement positif et négatif, la procédure de test est telle que les branches de correction "cas 1", "cas 2", permettent de sortir du test IV -en retenant la somme précédente rejetée a_i-1 et en calculant la période de façon normale.When, atest being equal to 6, the sum retained a; is such that

or

while the sign, respectively negative and positive, does not correspond to that expected, "first sign" being respectively positive and negative, the test procedure is such that the correction branches "case 1", "case 2", make it possible to exit the IV test -by retaining the previous sum rejected at _i-1 and calculating the period in a normal way.

When, atest being equal to 7, the sign of the sum being tested is that expected, a _i <0 and positive prime sign or a _i > 0 and negative prime sign, but this sum is less in absolute value at the threshold or when atest being equal to 7, the current time interval has become too large, t _23p > t _12p + e or t _23n > t _12n + e, only the first sum of the n ^th period (respectively a ₃ p and a _3n for cases 3 and 4) is retained and the melody period is then equal to the corresponding time interval, t ₂₃ p or t _23n . These corrections are very important because these particular configurations occur frequently and would lead to double period detection if they were not taken into account.

The voiced - unvoiced decision is made directly from the test result, by the value of the period. When the decision is requested at a rate different from that of the test, at the frame rate (given the frame counter 34) by the output interrupt pulses 1 ₅ applied to the microprocessor 40, the value of the period, result of the procedure of test, can be corrected by calculating an average value. In fact, the measurement of the value of the melody period can be given in real time or with a delay frame, an output register being provided for storing the current value of the melody period at suitably chosen times. . When during the test procedure, test III or test IV fails, or when no zero crossing is detected during a frame, this output record is reset.

However, the voiced - unvoiced decision logic can be a little more elaborate: For example, an additional duration criterion is introduced so that a voiced segment is always greater than 25 ms for example. Likewise, a segment whose detection method would indicate the unvoiced character but whose duration would be less than 25 ms is masked by the insertion of melody values interpolated from those evaluated on adjacent voiced segments.

The melody frequency detection method described above can be performed with a microprocessor of modest performance. It was implemented during its study on a ROCKWELL microcomputer, AIM 65, built around an MPU 6502 microprocessor.

The test procedure described above by way of example and the detection device associated with it can be modified without departing from the scope of the invention.

For example, the device represented in FIG. 1 comprises an interface circuit 42. It is also possible to use two PIA interface circuits, these possibly making it possible to carry out additional interruptions and to introduce several modes of execution, continuous real-time execution mode for a system in operation, or launched execution for a certain number of frames when the processing is carried out on recorded data.

Furthermore, the flowcharts of the test procedures described above can be modified, for example by modifying the order of the elementary tests when possible, without departing from the scope of the invention. In addition, the threshold values indicated above by way of example can also be chosen for example according to the type of voice (male voice and female voice).

Claims (10)

1. Method for detecting, in real time, the melody frequency in speech from a reduced set of data measured on this signal, characterized in that this set is composed of measurements a _i (i variable) of l energy in the successive alternations of this signal and associated measurements t _i , of the durations of these alternations, and in that the test procedure implemented on these data comprises an acquisition phase during which a first series of tests confers, when it is verified, the acquired character in the voicing and results in the calculation of a first value of melody period, and a maintenance phase during which a second series of tests confirms when it is verified, the acquired character of the voicing and results in the updating of the value of the melody period, this second series of tests being repeated as long as the acquired character of the voicing is preserved and a new acquisition phase being implemented when the character acquired from the void is lost. 2. Detection method according to claim 1, characterized in that the first series of tests consists in selecting, in the series of energy measurements in the successive alternations of the signal, a _i , three groups of two successive measurements a _1p a _1n , a _2p - a _2n , a _3p - a _3n exceeding predetermined thresholds S _1p and S _1n for the first group and thresholds S ₂ p and S2n, S ₃ p and S _3n defined according to the energies in the previous alternations selected for the following groups, the time intervals between alternations of the same sign selected, calculated from the durations t _i of the alternations, fulfilling defined criteria, these three groups of alternations characterizing the beginnings of three successive melody periods. 3. Detection method according to claim 2, characterized in that the thresholds S _2p and S _2n are defined as being the largest value of ¾a _1p and of S _1p for the first and of ¾a _1n and S _1n for the second. 4. Detection method according to claim 3, characterized in that the thresholds S ₃ p and S _3n are defined by the relationships:

and

5. Detection method according to any one of claims 2 to 4, characterized in that the second series of tests consists in selecting, in the following a _i energy measurements in successive alternations, two successive measurements including at least one exceeds the threshold S ₄ p or S _4n according to the sign of the corresponding alternation, these thresholds S ₄ p and S _4n defined as a function of the energies in the previous alternations selected limiting wider neighborhoods with respect to the previous values retained than those defined by the thresholds S ₃ p and S _3n used in the first series of tests, the time intervals between alternations of the same sign selected, calculated from the durations t _i of the alternations, fulfilling defined criteria, these alternations retained characterizing the beginning of an n ^th melody period. 6. Detection method according to claim 5, characterized in that the thresholds S ₄ p and S _4n are defined as being the largest value of ¾ [a _2p + (a _2p - a _1p )] and of S _1p for the first and from

(a _2n + (a _2n - a _1n )] and S _1n for the second. 7. Detection method according to any one of the preceding claims, characterized in that, in addition to the threshold criteria on the energy measurements, a criterion of monotony in the variation of these energy measurements in the alternations selected is also checked in the test series to avoid detection of the frequency double the actual melody frequency. 8. Detection method according to any one of the preceding claims, characterized in that protection tests are provided in the first and in the second series of tests to rule out alternations which cannot characterize the start of a new period of melody because of their position in time compared to the previous alternations selected. 9. Detection method according to any one of the preceding claims, characterized in that at the end of the first series of tests, a test on the energy measurements rejected with respect to the energy in the previous alternation of same retained sign is made, to avoid initialization during the acquisition phase, during the melody period and not at the start of the period. 10. Device for implementing the detection method according to any one of the preceding claims, characterized in that it comprises an analog processing circuit with amplifier (12) and low pass filter (13), an analog converter -digital (20), digital processing circuits (30) delivering, from the filtered and digitized signal, measurements (t.) of time intervals between zero crossings and measurements of energy (a.) in the half-waves of this signal; and a microprocessor (40) comprising a data memory (44) for storing these measurements, a processing unit (41) for testing these measurements according to the test procedures stored in a programmable memory (45 ) of this microprocessor, an interface circuit (42) being provided between the microprocessor and the digital processing circuits.

Images