EP0543719A1 - Method and arrangement for voiced-unvoiced decision applied in a very low rate vocoder - Google Patents

Abstract

The method consists in considering the development of the speech signal over a given number K of successive frames, by assigning to each current frame n a score for each of the possible states of development of the speech signal, as a function of the correlation coefficients of samples distributed in each frame (1, 2, 3), in determining (5) within each current frame the state of development of the speech signal having the maximum score, in order to recover from this state the most probable state of the frame n-K, by successively passing through the maximum-score states of each preceding frame. Application: low bit rate vocoders. <IMAGE>

Description

The present invention relates to a voicing decision method and device for a very low bit rate vocoder.

In vocoders with very low bit rate of 1200 bits / second and less, the speech signal is segmented into frames of constant durations from 10 to 30 milliseconds so as to determine within them the periodicity still designated by "Pitch" "in Anglo-Saxon language, of the speech signal. These frames are arbitrarily positioned in the speech signal and only one voicing value is provided for each frame. As the frames are grouped in blocks of 2, 6 or 8 frames for example and so that the bit rate allocated to voicing is reduced, all possible voicing combinations are generally not authorized. It is generally considered in fact improbable that there may exist one or two isolated voiced frames in the middle of a packet of unvoiced frames, or vice versa. This leads in the usual embodiments to implementing a voicing decision process which operates in two passes, a first pass according to which a first local voicing decision is taken at the level of each frame, the latter resulting from the examination of the values of different parameters of the speech signal and a second pass to correct the decisions made during the first pass and delete in the voiced parts the small packets of unvoiced frames and vice versa. Naturally, the operation of the corresponding devices is based on a large part of heuristics (based on experiments, learning, etc.) which is most of the time satisfactory under mild conditions but which degrades very quickly as soon as the speech signal is disturbed for example by noise.

The object of the invention is to overcome the aforementioned drawbacks.

To this end, the subject of the invention is a voicing decision method for a very low bit rate vocoder according to which the speech signal is sampled and segmented into frames of constant duration, characterized in that it consists in considering the evolution of the speech signal on a determined number K of successive frames by assigning to each current frame n a score for each of the possible states of evolution of the speech signal as a function of the correlation rates of samples distributed in each frame, to be determined in each current frame state of evolution of the speech signal which has the maximum score to go up from this state to the most probable state of the frame nK passing successively through the states of maximum score of each previous frame.

• La figure 1 les différentes étapes du procédé selon l'invention mises sous la forme d'un organigramme.
• La figure 2 un diagramme d'états pour montrer le mécanisme des transitions mises en oeuvre dans le procédé selon l'invention permettant d'effectuer des décisions de voisement avec un maximum de vraisemblance.
• La figure 3 un exemple de fonctionnement du mécanisme d'état représenté à la figure 2 sur cinq trames successives.
• La figure 4 un graphe montrant comment une décision de voisement peut être prise en compte par un cheminement des états sur cinq trames successives.
• La figure 5 un mode de réalisation d'un dispositif pour la mise en oeuvre du procédé selon l'invention.

Other characteristics and advantages of the invention will appear below with the aid of the description which follows given with reference to the annexed drawings which represent:

• Figure 1 the different steps of the method according to the invention put in the form of a flowchart.
• FIG. 2 a state diagram to show the mechanism of the transitions implemented in the method according to the invention making it possible to make voicing decisions with maximum likelihood.
• Figure 3 an example of operation of the state mechanism shown in Figure 2 on five successive frames.
• FIG. 4 is a graph showing how a voicing decision can be taken into account by a routing of the states over five successive frames.
• Figure 5 an embodiment of a device for implementing the method according to the invention.

The method according to the invention which is represented diagrammatically by blocks 1 to 5 in FIG. 1 performs a voicing decision on speech signal ranges of duration multiple of the "pitch". This duration is calculated between two predetermined extreme values, which are a minimum value to take into account sufficient signal and a maximum value to both limit the computational load and take into account the speed of natural variation of the characteristics of speech. The treatment begins with the calculation of three parameters which are a long-term correlation rate note R _M , a correlation rate at order 1 note R₁ and a rate of transition to 0 note T _ppz . These calculations are shown in blocks 1, 2 and 3 in Figure 1.

dans laquelle M représenté une valeur de "pitch" en nombre d'échantillons, S_n et S_n-m sont des amplitudes d'échantillons de signal, N désigne le nombre d'échantillons analyses, K est une constante et Δ représente une fraction de M. Ce calcul permet de fournir la vraie valeur du "pitch" lorsque le taux de corrélation qui est obtenu présente une valeur maximale. Pour un son parfaitement voisé, R_M est égal à 1, tandis qu'il est pratiquement nul pour un son aléatoire (non voisé). La constante K₁ évite les divisions par zéro et fournit une valeur très faible de R_M dans les silences de parole où la puissance est très faible.The long-term correlation rate R _M is calculated according to the expression

in which M represents a value of "pitch" in number of samples, S _n and S _nm are amplitudes of signal samples, N denotes the number of samples analyzed, K is a constant and Δ represents a fraction of M This calculation makes it possible to provide the true value of the "pitch" when the correlation rate which is obtained has a maximum value. For a perfectly voiced sound, R _M is equal to 1, while it is practically zero for a random sound (unvoiced). The constant K₁ avoids divisions by zero and provides a very low value of R _M in speech silences where the power is very low.

dans laquelle S_n et S_n-1 représentent comme précédemment les amplitudes de N échantillons et K₂ est une constante.The correlation rate at order 1, R₁ is calculated according to the expression

in which S _n and S _n-1 represent as before the amplitudes of N samples and K₂ is a constant.

For a voiced sound, R₁ is very close to 1 and it approaches -1 for an unvoiced sound. The constant K₂ makes it possible to give to R₁ a value practically zero during the rests.

The rate of passage through zero, T _ppz provides the ratio between the number of sign changes of each sample S _n , and N / 2 this rate varies between a value almost zero for a voiced sound and a value close to 1 for a sound not seen.

The processing continues, as shown by block 4 in FIG. 1 by a normalization of the parameters R _N , R₁ and T _ppz determined previously. This normalization has the effect of fixing the values of the parameters R _N , R₁ and T _ppz between the two values 0 and 1, the value 0 corresponding to the perfectly unvoiced sound and the value 1 corresponding to the perfectly voiced sound. The normalized values obtained are then weighted by weighting coefficients α₁, α₂ and α₃ to form a quantity Z defined by the relation

Taking care to choose the values of the weighting coefficients such that the relation α₁ + α₂ + α₃ = 1 is verified, Z then takes the value 1 for perfectly voiced sounds and the value 0 for perfectly unvoiced sounds. The parameters α₁, α₂ and α₃ can for example be determined as follows α₁ = 0.45, α₂ = 0.35 and α₃ = 0.2 but their optimal values can only be really found by an experimental adjustment of the values of the coefficients α₁ as a function in particular of the filterings on the signal samples S _n .

sert alors d'indicateur brut de voisement compte tenu du bruit.The adaptation to a variable level of ambient noise takes place by noting that adding noise to the voiced signals has the consequence of reducing the correlation rates R _M and R₁ and increasing the rate of passage to zero T _ppz , that is to say that is, to globally decrease the maximum value Z _MAX possible for Z at a given time. For this, an update of an estimate of Z _MAX is carried out and the relative quantity $${\text{"z" = Z / Z}}_{\text{MAX}}$$

then serves as a rough indicator of voicing given the noise.

dont la décroissance est définie par une constante de temps qui est fonction de ε pour suivre l'évolution du bruit, cette constante de temps étant d'autant plus élevée que ε est petit.The quantity Z _MAX can be estimated at each new value of Z by updating Z _MAX with each new value of Z, if Z is greater than the value Z _MAX or if the new value of Z is less than Z _MAX , by calculating the quantity Z _MAX according to the relation $${\text{Z}}_{\text{MAX}}{\text{= (1-ε) Z}}_{\text{MAX}}$$

whose decrease is defined by a time constant which is a function of ε to follow the evolution of the noise, this time constant being all the higher as ε is small.

The voicing decision which is represented by block 5 in FIG. 1 takes place on a criterion of decision by maximum likelihood, by considering the possible evolution of the speech signal over several successive frames, by assigning in each new frame a score to each of the possible states of evolution depending on the value of the number z obtained from the relation (4).

To account for a voicing state over a minimum duration, for example of three consecutive frames, this leads to consider a combinational logic with 6 possible voicing states, as defined in the table in FIG. 2, in which "V "means" voiced "and" N "means" unvoiced ", and the graph of the transitions in FIG. 3 representing the transitions between possible states to reach the current frame n from the possible states of the previous frame n-1. It should be noted in FIG. 3 that each state of a frame n is preceded by one or two states in the previous frame n-1. Thus in frame n the NNV state is preceded by the NNN state, the NVV state is preceded by the NNV state, the VVV state is preceded either by the NVV state or by the VVV state, the NNN state is preceded either by the NNN state itself or by the VNN state, the VNN state is preceded by the VVN state and the VVN state is preceded by the VVV state.

Naturally, this type of mechanism can very well be adapted to different durations from which the number of states to be considered will depend in each case.

Each of the possible arrival states in the frame is assigned a score. For each of them, the score retained is the largest possible among those obtained by adding or subtracting, according to the signs indicated on the arrows representing the transitions between states in FIG. 3, the value of z calculated by the relation (4 ) previous to which it is optionally subtracted a predetermined threshold value Z₀ between 0 and 1. To take into account the fact that certain starting states become prohibited because their score value becomes infinite, the corresponding ending states are also prohibited.

For example, the score SA₂ of the arrival state 2 (VVV) in FIG. 3 is determined by one of the relationships:
- SA₂ = MAX ((SD₁ + Z-Z₀); SD₂ + (Z-Z₀)
if the starting states SD₁ and SD₂ are both authorized.
- SA₂ = SD₁ + (Z-Z₀) if the start state 2 is prohibited.
- SA₂ = SD₂ + (Z-Z₀) if starting state 1 is prohibited
where SA₂ = - Infinite if the starting states 1 and 2 are both prohibited. In the latter case, state 2 is prohibited as the starting state for the next frame.

Naturally with each new frame the method stores the arrival scores which are then used as starting scores for the next frame and the table of the previous states is also stored.

The final decision takes place on all the frames with a certain delay, that is to say that the voicing decision at the time of arrival of the frame n will be that of the frame nK in the manner represented by the flowchart. of FIG. 4. The method consists in searching among the six arrival states for the one with the highest score, then it goes back K less once in a row from the state it reached the frame ni (i = 0, 1, ..., K-1) in the previous state which is memorized for this frame. In the example of FIG. 4 which represents a situation with a delay of five frames, at frame n state 1 is assumed to have the best score. Because this state has, according to the diagram of the transitions of FIG. 3, as state preceding the state 0 at frame n-1, the method goes back to this one and continues in a similar way until frame n- 5 along the solid bold lines in FIG. 4. The state to which the method thus goes back is the state VVN which corresponds to the n-5 frame not seen in state 5. However, the method is only valid if the constraint of three successive frames of the same voicing state are respected, which is not always the case if the process stops at the previous stage. In fact, since the frame n-5 corresponds to state 5, the frame n-4 can in this case only be in state 4, the only state which it is possible to achieve from FIG. 3 in starting from state 5, the ascent in the states during the next frame must necessarily lead to state 5 in frame n-4. The rest of the processing then consists, for all the possible states in frame n, of eliminating the states (here states 0, 3, 4 and 5) from which it is not possible to go back to state 5, in the frame n-4 by positioning their scores at infinity. In the example of figure 4 state 1 is of course not eliminated, because it is from it that the ascent was started in the states and state 2 is preserved because the process can go back from it to the state 5 in the frame n-4 (bold dashed lines).

If in another example, the initial ascent would have led to state 3 (frames n-5), knowing that state 3 can only be followed by states 0 or 3 (frame n-4), all the states in frame n which does not go back to states 0 or 3 in frame n-4 would have been eliminated. This method of eliminating states has, among other advantages, that of outputting a stream of voicing values systematically satisfying the constraint (absence of voicing island / non-voicing isolated). If on the reception side, this constraint is not verified, this means that there was a transmission error which allows curative measures to be taken, for example, by positioning the voicing indicator in the incriminated area at "voiced" .

et au taux de passage à zéro-T_ppz. Les résultats des calculs effectués par les circuits multiplieurs 13, 14, 15 sont appliqués sur les entrées d'un circuit sommateur 16 fournissant sur sa sortie k valeur Z. La valeur Z de la relation (3) est appliquée à l'entrée du bloc de normalisation 7. Le bloc de normalisation 7 comporte un registre 17 de mémorisation du nombre Z maximum. La sortie du registre 17 est couplée à une première entrée d'un circuit comparateur 18 dont la deuxième entrée reçoit la valeur Z fournie par le dispositif de calcul 6. L'entrée du registre 17 est reliée à la sortie d'un circuit multiplexeur 19 commandé par la sortie du circuit comparateur 18 pour appliquer à l'entrée du registre 17 soit la valeur Z telle quelle, soit le résultat du calcul fourni par un circuit multiplieur 20 possédant une première entrée d'opérande reliée à la sortie du registre 17 et une deuxième entrée d'opérande sur laquelle est appliquée le coefficient 1-ε de la relation (5) décrite précédemment. Un circuit diviseur 21 est couplé par une première entrée d'opérande à la sortie du registre 17 et par une deuxième entrée d'opérande à la sortie du dispositif 6 pour calculer la quantité "z" égale à Z/Z_MAX. La quantité "z" est appliquée à l'entrée du dispositif de décision finale 8 qui comporte un dispositif de calcul 22 des scores, de remontées des états et d'éliminations des états interdits couplé à une mémoire 23 des scores de départ SDi, à une mémoire 24 des scores d'arrivée SAi et à une mémoire 25 des états précédents. Le dispositif de calcul des scores 22 reçoit d'autre part, sur une deuxième entrée la valeur de seuil Z₀, sur une troisième entrée le numéro de trames n et fournit sur une sortie les décisions de voisement pour chacune des trames n-K.A device for implementing the method according to the invention is shown in FIG. 5. This device comprises represented respectively inside closed dotted lines a device 6 for calculating the parameter Z, a device 7 for normalization, a device 8 for final decision. The device for calculating the parameter Z comprises a device 9 for calculating long-term autocorrelation of the value R _M , a device 10 for correlation to the order 1 of the coefficient R₁ and a device 11 for calculating the zero crossing rate T _ppz . The calculation devices 9, 10 and 11 each receive simultaneously on a first input the signal samples S _n and carry out the calculations corresponding to the preceding relations (1), (2) and (3) inside a window d analysis provided by a device 12 which determines an analysis duration for each frame n as a function of the value of the "pitch" M. The multiplier circuits 13, 14, 15 are coupled to the output of the calculation devices 9, 10 and 11 to apply the weighting coefficients α₁, α₂ and α₃ to the correlation rates R _M and $$\frac{\text{R₁ + 1}}{\text{2}}$$

and at the zero crossing rate-T _ppz . The results of the calculations carried out by the multiplier circuits 13, 14, 15 are applied to the inputs of a summing circuit 16 providing on its output k value Z. The value Z of the relation (3) is applied to the input of the normalization block 7. The normalization block 7 comprises a register 17 for storing the maximum number Z. The output of the register 17 is coupled to a first input of a comparator circuit 18, the second input of which receives the value Z supplied by the computing device 6. The input of the register 17 is connected to the output of a multiplexer circuit 19 controlled by the output of the comparator circuit 18 to apply to the input of the register 17 either the value Z as it is, or the result of the calculation provided by a multiplier circuit 20 having a first operand input connected to the output of the register 17 and a second operand entry to which the coefficient 1-ε of the relation (5) described above is applied. A divider circuit 21 is coupled by a first operand input at the output of the register 17 and by a second operand input at the output of the device 6 to calculate the quantity "z" equal to Z / Z _MAX . The quantity "z" is applied to the input of the final decision device 8 which includes a device 22 for calculating the scores, feedback of the states and elimination of the prohibited states coupled to a memory 23 of the starting scores SDi, at a memory 24 of the arrival scores SAi and a memory 25 of the previous states. The score calculation device 22 receives on the other hand, on a second input the threshold value Z une, on a third input the number of frames n and provides on an output the voicing decisions for each of the frames nK.

Naturally, the implementation of the method according to the invention which has just been described is not unique and it goes without saying that other embodiments using microprocessors for signal processing in particular are just as possible, their programming being within the reach of the skilled person.

Claims (10)

Voting decision method for a very low bit rate vocoder according to which the speech signal is sampled and segmented into frames of constant duration, characterized in that it consists in considering the evolution of the speech signal over a determined number K of successive frames by assigning to each current frame n a score for each of the possible states of evolution of the speech signal as a function of the correlation rates of samples distributed in each frame (1, 2, 3), to be determined (5) in each frame current the state of evolution of the speech signal which has the maximum score to go up from this state to the most probable state of the frame nK passing successively through the states of maximum score of each previous frame. Method according to Claim 1, characterized in that it consists in calculating (1) a long-term correlation rate R _M for an interval M equal to the value of the pitch of the speech signal. Method according to claims 1 and 2 characterized in that it consists in calculating (2) a correlation rate at order 1 to determine the voiced or unvoiced nature of the sound corresponding to the correlated samples. Method according to any one of Claims 1 to 3, characterized in that it consists in determining (3) the rate of zero crossings of the speech signal. Method according to any one of Claims 2 to 4, characterized in that it consists in (4) performing a weighted summation of the long-term correlation rate, of the order 1 correlation rate and of the zero crossing rate to obtain a score increment "z" and calculate from the possible scores of each frame preceding a current frame, the scores of the possible states of arrival. Method according to claim 5, characterized in that it consists in giving infinite score values to the prohibited states. Method according to either of Claims 5 and 6, characterized in that the increment value is equal to the weighted sum of the long-term correlation rate, of the correlation rate at order 1 and the zero crossing rate minus a predetermined threshold value Z₀. Device for implementing the method according to any one of Claims 1 to 7, characterized in that it comprises a device (6) for calculating the score increment value, coupled to a final decision device (8) comprising a device for calculating the scores, feedback of the states and elimination of the prohibited states by means of a standardization device (7). Device according to Claim 8, characterized in that the device (6) for calculating the score increment value comprises a device for calculating the long-term correlation (9), a device for calculating the correlation at order 1 (10 ) and a device (11) for calculating the zero crossing rate, coupled to a summing circuit (16). Device according to Claim 9, characterized in that the device (6) for calculating the score value is formed by a signal processing processor.

Images