GB2059726A - Sound synthesizer - Google Patents

Abstract

A sound synthesizer in which pulses of a period indicated by a fundamental period parameter are produced by a fundamental period sound source, the output from the fundamental period sound source or the output from a noise source is selected depending on whether a sound to be synthesized is a voiced or unvoiced sound, and the selected pulses are applied to a sound synthesis filter section 16 to synthesize the sound. The sound synthesis filter section is composed of a second- order filter means (57-61, 65-69) which serves as a second- order filter having the zero on a unit circle in a complex plane, means for cascade-operating second-order filter means of different coefficients, and feedback means 41, 42 for feeding back the output from the synthesis filter section 16 to the input side through two kinds of such cascade-operating means, and coefficients of the second-order filters are controlled by control parameters. <IMAGE>

Description

SPECIFICATION Sound synthesis The present invention relates to a method of sound synthesizing and to a sound synthesizer with which it is possible to reconstruct a sound of substantially the same quality as an original sound from its features transmitted or stored in a memory in a small amount of information.

For example, in the case of reconstructing speech from feature parameters of original speech, according to the prior art the output of a pulse generator simulating the vibration of the vocal cord and the output of a noise generator simulating turbulence are changed over or mixed together depending on whether the speech is voiced or unvoiced and the resulting output is amplitude-modulated in accordance with the speech amplitude to produce an excitation source signal which is applied to a filter simulating the resonance characteristics of the vocal tract to obtain synthesized speech. A synthesis system using partial auto correlation (PARCOR) coefficients and a formant synthesis system are examples of such speech synthesis system enploying the feature parameters. The former is set forth, for example, in J.D.Markel et al., "Linear Prediction of Speech", pages 92-128, Springer-Verlag, 1976, in which the partial auto correlation coefficients or the so-called PARCOR coefficients of a speech waveform are used as the feature parameters. If the absolute values of the PARCOR coefficients are all smaller than unity, the speech synthesizing filter is stable. The PARCOR coefficients may be relatively small in the amount of information for speech synthesis and the automatic extraction of the coefficients is relatively easy, but the individual parameters differ widely in the spectral sensitivity. Accordingly, if all the parameters are quantized using the same number of bits, spectral distortions caused by quantization errors for the respective parameters largely differ from each other.Further, the PARCOR coefficients are poor in their interpolation characteristics and, by the interpolation of the parameters, there are produced noises, resulting in an indistinct speech. Especially at a low bit rate, the speech quality is deteriorated by the spectral distortion and no satisfactory synthesized speech quality is obtainable. In addition, since the PARCOR coefficients do not directly correpond to spectral properties such as formant frequencies, and hence the PARCOR coefficients are not suitable for speech synthesis by rule.

The formant synthesis system is disclosed, for example, in J.L. Flanagan, "Speech Analysis, Synthesis and Perception", pages 339-347, Springer-Verlag, 1 972. This system is one which synthesizes speech using the formant frequencies and their intensity as parameters and which is advantageous in that the amount of information for the parameters may be small and in that the correspondence of the parameters to spectral quantities is easy to obtain. For the extraction of the formant frequency and the intensity thereof, however, it is necessary to make use of general dynamic characteristics and statistical properties of the parameters, and complete automatic extraction of the formant frequency and the intensity thereof is difficult.Accordingly, it is difficult to automatically obtain synthesized speech of high quality and it is likely to markedly degrade the quality of the synthesized speech by an error in the extraction of the parameters.

It is an object of the present invention to provide a sound synthesizer which is able to synthesize a sound of high quality using a small amount of information.

Another object of the present invention is to provide a sound synthesizer which permits relatively easy extraction of the feature parameters and operates stably and in which differences in the spectral sensitivity among the parameters are small and the quantization accuracy of the parameters is the same in the case of the same quantization bits.

Another object of the present invention is to provide a sound synthesizer which is excellent in interpolation characteristics for parameters used and hence is able to obrain a synthesized sound of high quality with a small amount of information.

Yet another object of the present invention is to provide a sound synthesizer which can be produced in a relatively simple structure.

SUMMARY OF THE INVENTION In a linear predictive analysis, the speech spectral envelope is approximated by a transfer function of an all-pole filter which is given by the following expression (1): a a H(Z) = ---- = (1) Ap(Z) 1 + a,Z + a2Z2 + + apZP where Z = e ., U is a normalized angular frequency 2srfAT, AT is a sampling period, f is a sampling frequency, p is the degree of analysis, aj (i = 1, 2, .. p) are predictor coefficients which are parameters for controlling the resonance characteristic of the filter and a is the gain of the filter.Here, Ap(Z) is represented by the sum of two polynominals which can be expressed as follows: Ap(Z) = 1/2(P(Z)+ Q(Z)) (2) P(Z) = Ap(Z)Z ZPAp(Z- ') (3) Q(Z) = Ap(Z) + Z- ZPA (Z- 1) (4) (a) When the degree of analysis p is even, the expressions (3) and (4) are factorized as follows:

(b) When the degree of analysis p is odd, the expressions (3) and (4) are factorized as follows:

a,j and 9j in the expressions (5) and (6) are called a line spectrum pair (hereinafter referred to as LSP) and in the present invention, they are used as parameters for representing spectral envelope information.

Expressing Ap(Z) as given by the expression (2), the transfer function H(Z) becomes as follows: a a a - = = = (7) Ap(Z) 1 +(Ap(Z)1) 1 ++fP(Z)-1 + Q(Z)- 1) The transfer function H(Z) is also formed as a filter having two feedback loops whose transfer functions P(Z) - 1 and Q(Z)- 1, respectively.The transfer functions P(Z) and Q(Z) are antiresonance circuits and their output become 0 at xj and xj. The frequency characteristic of Ap(Z) becomes as follows:

where Z = e-i". It appears from the above expression (8) that in a region where adjacent line spectral frequencies are close to each other, jAp(Z)12 is small and the transfer function H(Z) exhibits a strong resonance characteristic. By changing the values of the LSP parameters q and 0 describing the resonance characteristic of the transfer functions, an arbitrary speech spectral envelope can be obtained.

The procedure to obtain the LSP parameters is as follows; in the first step, auto correlation coefficients of speech wave are obtained at intervals of, for example, 10 to 20 msec, in the second step, predictor coefficients a1 of the transfer function H(Z) are obtained from the auto correlation coefficients, and in the third step, the solutions of the two polynominals P(Z) and Q(Z) are obtained from the predictor coefficients on the basis of the relationship of the expression (2), thus obtaining the LSP parameters xj and Sj. By controlling coefficients of the synthesis filter through utilization of the parameters representing the speech spectral envelope information, there can be obtained a filter whose transfer function H(Z) is equivalent to the speech spectral envelope.The transfer function of the feedback loop in the synthesis filter is provided in the form of a cascade connection of second-order filters, whose zeros are on a unit circle in a plane Z, as indicated by the expressions (5) and (6). Since these two second-order filter are identical in construction, the construction can also be simplified by multiple utilization of one second-order filter using time shared operation or what is called a pipeline operation. It is also possible to perform the filter operation by the processing of an electronic computer without forming the second-order filters as circuits.

As described above, in the present invention the characteristics of the synthesis filter are controlled by the aforesaid parameters xj and Bi but, in addition to these LSP parameters q and 9;, a fundamental frequency parameter and an amplitude parameter are employed as is the case with this kind of speech synthesizers heretofore used.By the fundamental frequency parameter, a voiced sound source is controlled to generate a pulse or a group of pulses of the frequency indicated by the parameter; the output from the voiced sound source or the output from a noise source is selected depending on whether the sound to be reconstructed is voiced or unvoiced; the selected output is applied to the sound synthesis filter; and the magnitude of a signal on the input or output side of the synthesis filter is controlled by the amplitude parameters.The LSP parameters xj and B,are subjected to cosine transformation by parameter transforming means to obtain - 2cosy'1 and - 2cosy,, which are used as control parameters for controlling the coefficients of the second-order filters of the sound synthesis filter respectively corresponding to the parameters. The control parameters are interpolated by interpolating means in the form of the cosine-transformed LSP parameters - 2cosa'1 and - 2cosy,. Also the interpolating means may be employed for the interpolation of the amplitude parameter.The LSP parameters > j and i are excellent in interpolatability and the interpolation is conducted at time intervals equal to or twice the sampling period of the original sound for producing the parameters; for example, the LSP parameters q and 9j are updated every frame of 20 msec and the parameters in each frame are futher interpolated every 1 25 ysec. It is also possible to effect the interpolation in the state of the LSP parameters xj and 9j and convert them to the control parameters.

The LSP parameters a; and 9j are small in the amount of information per frame as compared with the control parameters for the synthesis filter for speech synthesis in the past and excellent in interpolation characteristic. Accordingly, it is suitable to transmit or store the LSP parameters aij and i9* as they are and it is also possible to convert the received or reconstructed LSP parameters ; and 9; to the control parameters for the synthesis filter employed in other speech synthesizing systems, i.e. the PARCOR coefficients or linear predictor coefficients. In this way, the LSP parameters q and 0 can also be used in the existing speech synthesizers.The sound synthesizer of the present invention is applicable to the synthesis of not only ordinary speech but also sounds such as a time signal tone, an alarm tone, a musical instrument sound and so forth.

BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 is a block diagram showing the fundamental construction of an embodiment of the sound synthesizer of the present invention; Figure 2 is a block diagram showing a specific operative example of the sound synthesizer of the present invention; Figure 3 is a circuit diagram showing an example of a first-order or second-order filter forming a synthesis filter section; Figure 4A is a diagram illustrating an example of the synthesis filter section in the case of the degree of analysis being even; Figure 4B is a diagram illustrating an example of the synthesis filter section in the case of the degree of analysis being odd; Figure 5 is a diagram showing the relationship between the LSP parameters xj and 8i and the speech spectral envelope;; Figure 6 is a circuit diagram illustrating a specific operative example of the synthesis filter section in the case of the degree of analysis being 4; Figure 7 is a circuit diagram illustrating a specific operative example of the synthesis filter section obtained by an equivalent conversion of the circuit shown in Fig. 6; Figure 8 is a circuit diagram showing a specific example of the synthesis filter section in the case of the degree of analysis being 5; Figure 9 is a circuit diagram showing a specific operative example of the synthesis filter section obtained by an equivalent conversion of the circuit shown in Fig. 8; Figure 10 is a block diagram illustrating an example of the synthesis filter section employing the pipeline calculation system;; Figures 1 lA to ill inclusive, show timing charts showing the variations of signals appearing at respective parts during the operation of the filter section depicted in Fig. 10; Figure 12 is a circuit diagram showing the case in which the filter operation achieved by the operation shown in Fig. 11 is provided by a series connection of filters; Figure 13 is a block diagram illustrating an example of the synthesis filter using a microcomputer; Figure 14A is a diagram showing the variations of power with the lapse of time in the case where a speech "ba ku o N ga" was made; Figure 14B is a diagram showing the fluctuations in the LSP parameters xj and 9j with the lapse of time. in the case where the speech "ba ku o N ga" was made;; Figure 15 is a diagram showing the relative frequency distributions of the LSP parameters x and 0, to frequency; Figure 16 is a diagram showing the relationship between the number of quantizing bits per frame and the spectral distortion by quantization; Figure 1 7 is a diagram showing the relationship of the spectral distortion by interpolation to the frame length in the case of the parameters having been interpolated; and Figure 18 is a diagram showing an example of synthesizing speech by converting the LSP parameters coj and Aj to a parameters.

DESCRIPTION OF THE PREFERRED EMBODIMENTS Referring first to Fig. 1, the feature parameters of a speech to be synthesized are applied from an input terminal 11 to an interface section 1 2 every constant period of time (hereinafter referred to as the frame period), for example, every 20 msec and latched in the interface section 1 2. Of the parameters thus input, the LSP parameters xi and Aj indicating spectral envelope information are provided to a parameter transforming section 13; and, of parameters indicating sound source information, amplitude information is applied to a parameter interpolating section and the other parameters, that is, information indicating the fundamental period (pitch) of the speech information indicating whether the speech is a voiced or unvoiced sound are applied to a sound source signal generating section 1 5.In the parameter transforming section 13, the input LSP parameters xj and 9j are transformed into control parameters - 2cosy' and - 2cos8, for a synthesis filter section 16, which parameters are provided to the parameter interpolating section 1 4. In the parameter interpolating section 14, interpolation values for the control parameters and the sound source amplitude parameter are respectively calculated at regular time intervals so that the spectral envelope may undergo a smooth change.The control parameters thus interpolated are supplied to the synthesis filter section 16, and the sound source amplitude parameter is applied to the sound source signal generating section 1 5. In the sound source signal generating section 15, a sound source signal depending on the features of speech is produced on the basis of the pitch information and the voiced or unvoiced sound information, and the sound source signal thus obtained is applied to the synthesis filter section 1 6 together with the interpolated sound source amplitude parameter. In the synthesis filter section 16, a synthesized speech is produced from the sound source signal and the control parameters.The output from the synthesis filter section 1 6 is provided to a digital analog converting section 1 7 and derived therefrom as an analog signal at its output terminal 1 8. A control section 1 9 generates various clocks for activating the speech synthesizer correctly and supplies them to the respective sections.

Fig. 2 illustrates in a little materialized form each section of Fig. 1. Every frame period the information on the voiced or unvoiced sound of speech is applied from the interface section 1 2 to a voiced sound register 23 and an unvoiced sound register 24, and a voice frequency parameter indicating the voice pitch is stored in a pitch register 25. The content of the pitch register 25 is preset in a down counter 27. The down counter 27 counts down pulses of a sampling frequency from a terminal 26 and every time its content becomes zero, tke counter 27 presets therein the content of the pitch register 25 and, at the same time, supplies one pulse to a gate 31.To the gate 31 are applied the output from the voiced sound register 23 and an output pulse or pulses from a pulse generator 28, and when these inputs coincide, the content of a sound source amplitude register 34 is provided via the gate 31 to an adder 32. In other words, when the speech to be synthesized is a voiced sound, the amplitude information is applied to the adder 32 from the sound source amplitude register 34 every period of fundamental voice frequency of the pitch register 25, the amplitude information from the sound source amplitude register 34 being preset therein from the interpolating section 14.

In the case where the speech to be synthesized is an unvoiced sound, the output from the unvoiced sound register 24 and a pseudo random series pulse from a pseudo random signal generator 36 are provided to a gate 37, and upon every coincidence of the both inputs, the amplitude information in the sound source amplitude register 34 is provided via the gate 37 to the adder 32. A sound source signal thus derived from the adder 32 is amplified, if necessary, by an amplifier 39 and then applied to the speech synthesis filter section 1 6.

In the parameter transforming section 13, the LSP parameters ; and 9j and the amplitude parameter are set in a register 21 from the interface section 1 2 every frame period. The LSP parameters wi and 0 are applied to a parameter converter 22, wherein they are transformed to control parameters - 2cosq and - 2cosy,. The parameter converter 22 is formed, for example, by a conversion table of a read only memory (ROM), which is arranged so that when accessed with addresses corresponding to xj and 8,, - 2cosy' and - 2cosy, are read out.A shift register 20 receives alternately the output from the parameter converter 22 and the amplitude parameter stored in the register 21 and converts them to a series signal, which is applied to the parameter interpolating section 14.

In the illustrated example, the parameter interpolating section 14 is shown to perform a linear interpolation. Upon turning ON a switch 29, the parameters of one frame are supplied to a subtractor 30, wherein a difference is detected between the parameter and that of the previous frame from an adder 33. The difference is stored in a difference value register 38 via a switch 91. Thereafter, the switch 91 is changed over to the output side of the difference value register 38 and the content thereof is circulated. At this time, the content of the difference value register 38 is taken out from bit positions higher than a predetermined bit position and supplied to the adder 33, wherein it is added to the content of an interpolation result register 92.For example, in the case of the parameter update period being 1 6 msec, if it is necessary to provide interpolation parameters 1 28 times during a frame update period, then the interpolation step width is a value obtained by dividing the difference value by 1 28 and this is obtained by shifting the difference value in the difference value register 38 towards the lower order side by seven bits. The result of addition by the adder 33 is provided to the interpolation result register 92 and, at the same time, it is used as the output from the parameter interpolating section 14.

In this way, there are derived from the adder 33 the values that are obtained by sequentially adding values once, twice, three time, ... the shifted value of the difference register 38 to the parameter of the previous frame in the interpolation result register 92 every circulation of the difference value register 38.

In this example, the parameter interpolating section 14 is used for the control parameter and the amplitude parameter on a time-shared basis, so that, though not shown, the control parameter and the amplitude parameter are alternately interpolated and the interpolation result register 92 is used in common to the both parameters. The amplitude parameter interpolated in the parameter interpolating section 14 is provided to the amplitude information register 34 in the sound source signal generating section 15, whereas the control parameter interpolated as mentioned above is applied to the speech synthesis filter section 1 6 as information for controlling its filter coefficient. The parameter update period, that is, the frame period, is selected to be in the range of 10 to 20 msec, and the interpolation period is selected to range from one to two sampling interval.The interpolation method is not limited specifically to the linear interpolation but may be other types of interpolation. The point is to ensure smooth variations of the interpolated parameters.

The synthesis filter section 1 6 is provided with a loop for feeding back the output through filter circuits 41 and 42 parallelly connected each other. The filter circuits 41 and 42 are supplied with the interpolated control parameter from an input terminal 44 and the outputs from the filter circuits 41 and 42 are added together by an adder 43, the output from which is, in turn, added to the input to the filter section 16 in an adder 45. The added output therefrom is fed back to the filter circuits 41 and 42 and, at the same time, derived at an output terminal 55.

As each of the filter circuits 41 and 42, use is made of a circuit which has a plurality of zeros on a unit circle in a complex plane. The filter circuits 41 and 42 can be both formed by a multistage cascade connection of first-order and/or second-order filters. In the case of forming the filter circuits as digital filters, use can be made of a first-order filter such, for example, as shown in Fig. 3A which is composed of a delay circuit 51 having a delay of one sample period and an adder for adding the delayed output and a non-delayed input, a second-order filter such as shown in Fig. 3B which is composed of two stages of delay circuits 51 and the adder 52 for adding the delayed ouput and the non-delayed input, and a second-order filter such as shown in Fig. 3C in which the output from a multiplier 53 for multiplying the delayed output from one stage of delay circuit 51 by - 2cosy, the delayed output from two stages of delay circuits 51 and the non-delayed input are added together by the adder 52. The transfer functions of the filters shown in Figs. 3A , 3B and 3C are 1 + Z, 1 - Z2 and 1 - 2cosa'Z + + Z2, respectively. It is also possible to employ higher order filters.

The combination and the number of such filters depend on the degree of analysis and selected as shown in Fig. 4A or 4B depending on whether the degree of analysis is even or odd.

In Fig. 4A, the degree of analysis is 10, namely, an even number and the filter circuit 41 is constituted by a series connection of a first-order filter 56 having the transfer function 1 - Z and second-order filters 57 to 61 each having the transfer function 1 - 2cosa'Z + + Z2, and the output at the output terminal 55 is multiplied by + 1/2 in a multiplier 63 and applied to the series circuit. The output from the second-order filter 61 of the last stage and the output from the multiplier 63 are added together by an adder 62 and the added output therefrom is provided to the adder 43.In the filter circuit 42, the output from the multiplier 63 is supplied to a series circuit of a first-order filter 64 having the transfer function 1 + Z and second-order filters 65 to 69 each having the transfer function 1 - 2cos01Z + Z2, and the output from the series circuit and the output from the multiplier 63 are added together in an adder 71, the added output from which is applied to the adder 43. The multipliers 53 of the second-order filters 57 to 61 are respectively given control parameters a, = - 2cosy, to a5 = - 2cosy'5 and the multipliers 53 of the second-order filters 65 to 69 are respectively given control parameters b, = - 2cosy, to b5 = - 2cos05.

Fig. 4B shows the case where the degree of analysis is 11, namely, an odd number. In the filter circuit 41, the first-order filter 56 employed in the case of Fig. 4A is omitted but instead a second-order filter 72 having a transfer function 1 - Z2 iS used. In the filter circuit 42, the firstorder filter 64 is omitted but instead a second-order filter 73 given a parameter b6 = - 2cos06 is used.

In the filter circuits 41 and 42 the control parameters xj and i9j represent anti-resonance frequencies, at which the outputs from the filter circuits 41 and 42 become 0.5. Accordingly, in the case where the anti-resonance frequency applied to the filter circuits 41 and 42 are close to each other, the output from the adder 43 becomes close to unity and the feedback loop gain approaches unity. As a consequence, a high resonance characteristic appears at the output terminal 55. Here, , to w, and 9, to 85 are anti-resonance frequencies, which are characteristic of the speech spectral envelope information.Thses parameters and the spectral envelope characteristic bear such relationship as depicted in Fig. 5, from which it appears that the resonance characteristic of the spectrum can be expressed by the spacing between adjacent parameters. These parameters have the following relationship of order: O < 9, < , < 92 < co2 < 9i < i < sT (8') The synthesizing filter has the feature that it is stable when the above condition is fulfilled.

Next, a description will be given of a specific operative example of the synthesis filter section 1 6. Corresponding to the term in the braces of the denominator in the expression (7), the following identical equations are obtained from the expressions (5):

A digital filter is formed which has an all pole transfer function approximating the speech spectral envelope given by the expression (1) using the relationships given by the expressions (7), (9) and (10). Fig. 6 shows the case where P = 4. In Fig. 6, parts corresponding to those in Figs. 3B to Fig. 4 are identified by the same reference numerals. The input from the terminal 54 is added by the adder 45 to the output from the adder 43, and the added output is provided to the output terminal 55 and, at the same time, multiplied by + 1/2 in the multiplier 63.This 1/2 multiplication corresponds to that in the denominator in the expression (7). The output from the multiplier 63 is applied to delay means 74 whose delay time is one sampling period, i.e. the unit time. The delayed output is applied as the input to each of the second-order filters 57 and 65, in which it is applied to the delay means 51, the multipliers 53 and the adders 52.

In the both multipliers 53, the inputs thereto are respectively multiplied by a, and b,, and the multiplied outputs are each applied to an adder 94 for addition with the output from the delay means 51 in each of the filters 57 and 65. The outputs from the both adders 94 are provided to a common adder 81 and, at the same time, applied to the adder 52 via delay means having delay time of one sampling period in each of the filters 57 and 65. The outputs from the both adders 52 are respectively applied as the outputs from the filters 57 and 65 to the second-order filters 58 and 66 of the next stage. The filters 58 and 66 are identical in construction with the filters 57 and 65, but the coefficients for the multipliers 53 are a2 and b2, respectively. The output from the adder 94 of each filter is applied to an adder 82 for addition with the output from the adder 81. The outputs from the adders 52 of the both filters 58 and 66 are supplied to the adder 43 for subtraction from each other, and the adder 43 is further supplied with the output from the adder 82.

The delay means 74 corresponds to Z outside the braces in the expressions (9) and (10), and the filters 57 and 58 each constitute a second-order filter having a transfer function 1 + Z(aj + Z), and similarly the filters 65 and 66 each constitute a second-order filter having a transfer function 1 + Z(bj + Z). Accordingly, the series connection of the second-order filters 57 and 58 realizes the third term in the braces in the expression (9), and the delay means 51, the multiplier 53 and the adder 94 in the filter 58 realize (at+, + Z); consequently, by this circuit and the second-order filter 57, the second term in the braces in the expression (9) is realized, and the output is provided via the adder 82 to the adder 43.The delay means 51, the multiplier 53 and the adder 94 in the second-order filter 57 realize (a, + Z) and the output is supplied to the adder 43 via the adders 81 and 82. In this way, the terms in the braces in the expression (9) are realized by the second-order filters 57 and 58 and the adders 43, 81 and 82.

Likewise, the terms in the braces in the expression (10) are realized by the second-order filters 65 and 66 and the adders 43, 81 and 82. The expressions (9) and (10) differ in form only in that the signs of the third terms in the braces are different from each other, and on account of this difference, the sign of the input to the adder 43 differs. Accordingly, the adder 43, the second-order filters 57, 58, 65 and 66, the multiplier 63 and the delay means 74 realize the' expression (2), and the circuit arrangement of Fig. 6 materializes the expression (1) as a whole.

In this circuit arrangement, the expressions (9) and (10) are materialized by forming the filter circuit 41 with a series connection of (P/2)'s second-order filters 57 and 58 and the filter circuit 42 with a series connection of (P/2)'s second-order filters 65 and 66 in the feedback loop, by taking out the nodes of the second-order filters of the filter circuit 41, that is, taps 96 and 97, from the output sides of the adders 94 to obtain the total sums with the adders 81, 82 and 83.

The arrangement for taking out outputs from the taps of the filter circuits will hereinafter referred to as the tap output type.

In Fig. 6, the second-order filters are arranged towards the adder 43 in an increasing order of the value j but they may also be arranged in a decreasing order of the value j. In such a case, for example, as shown in Fig. 7, the output from the delay means 74 is provided to the secondorder filters 58 and 66, the ouputs from which are applied via the second-order filters 57 and 65 to the adder 43. In Fig. 7, the preceding stage of each second-order filter in Fig. 6 is exchanged with the succeeding stage; namely, the circuit 94 for adding together the outputs from the delay means 51 and the multiplier 53 is exchanged with the delay means 95. The output from the delay means 74 is provided via the taps 96 and 97 to the nodes of the secondorder filters 57 and 58.In other words, the circuit arrangement of Fig. 6 is'the tap output type, whereas the circuit arrangement of Fig. 7 is a tap input type. The circuit beginning with the tap 96 and ending with the adder 43 constitutes the first term in the braces of the expression (9), and the circuit from the tap 97 to the adder 43 constitutes the second term in the braces of the expression (9). The second-order filters 65 and 66 of the filter circuit 41 are also similarly formed. In connection with the filter circuit 41, the output from the delay means 74 is multiplied by - 1 in a multiplier 98 to materialize the minus sign for the third term in the braces of the expression (9).

In the case where p is odd, the following identical equation is obtained from the expression (8) corresponding to the term in the braces of the denominator in the expression (7).

As in the case of p being even, two types of digital filters respectively called the tap output type and the tap input type are materialized in such forms as shown in Figs. 8 and 9 from the relations of the expressions (7), (12) and (13). In Figs. 8 and 9, it is assumed that p is 5. In Figs. 8 and 9, the first-order filter 72 corresonds to Z in the third term in the braces of the expression (13) and the second-order filter 73 is to obtain such a characteristic that the products of the transfer functions (1 + b,Z + Z2) and (1 + b2Z + Z2) Of the filters 65 and 66 is multiplied by (b3 + Z).

As will be understood from Figs. 6 to 9, the + 1/2 multiplier 63 and the delay means 74 may also be disposed at any places in the feedback loop. Since the second-order filters are of the same type, it is possible to simplify hardware by forming the circuit arrangement so that the so-called pipeline operation is effected by using, on a time-division multiplex basis, one multiplier 53, the plurality of adders 52 and 94 and the plurality of delay means 51 and 95 making up one second-order filter. Fig. 10 illustrates the case where the example of the filter shown in Fig. 1 2 is arranged to conduct the pipeline operation. In this example, p = 10, and an operation of a set of parameters applied from the interpolating section is completed with a period of 1 76 clocks.In Fig. 10, parts corresponding to those in Fig. 1 2 are marked with the same reference numerals. The input side of a 16-bit static shift register 74, which performs the function of the delay means 74 is changed over by a switch S, between the ouput side of the shift register itself and the output side of the adder 45. A multiplicand input side of the multiplier 53 and the input side of the adder 52 are changed over by a switch S2 to the output side of the shift register 74, the output side of a (27-d)th shift stage counted from the input of the shift register 74 and the output side of a 31-bit shift register 101, d being an operation delay of the multiplier 53.The multiplier 53 is connected at one end to the output terminal 55 and the input side of the adder 94 and derives at the other output end the multiplicand input delayed by 22 clocks, which is provided to the (154 + d)-bit shift register 51. The output from an adder 81 is fed back to the input side thereof via a gate 102 and a 16-bit shift register 103, performing a cumulative addition through the adders 81 and 82 in Fig. 12. The gate 102 is opened only in the time interval between d + 2 and 145 + d. One input side of the adder 43 is changed over by a switch S3 between the output sides of the adders 52 and 81, and the other input side of the adder 43 is changed over by a switch S4 between the output sides of a 16th and a (d + 1 )th shift stages of the shift register 101.The input side of the shift register 101 is changed over by a switch 85 between the output sides of the adders 43 and 52.

The switches S, to S5 are each connected to the fixed contact side, during one operation period, that is, 1 76 clocks, for a clock period indicated by numerals labelled at the fixed contact. The shift registers 51, 95, 101 and 103 are respectively (154 + d)-bit, (175 -- d)-bit, 31-bit and 16-bit dynamic type ones and are always supplied with shift clocks. The broken line input to each of the adders 43, 45, 52, 81 and 94 indicates the timing of the operation boundary of each parameter; for example, 00 indicates a repetition every 1 6 clocks and an operation delay of each adder is selected to be one clock.Fig. 11 is a timing chart of the operation of each part in Fig. 10, Fig. 11A showing the timing of the clock, Fig. 11 B the inputs of the coefficients a,, bj and A to the multiplier 53 from the input terminal 44, Fig. 1 1C the multiplicand of the multiplier 53, Fig. 11 D one input to the adder 94 from the multiplier 53, Fig. 11 E the other input to the adder 94, Fig. 11 F the output from the adder 94, Fig. 11 G the output from the adder 81, and consequently the content of the register 103, Fig. 11 H the input to the adder 52 from the shift register 95, and Fig. 111 the output from the adder 52. Fig. 12 shows these inputs and outputs in the form of signals appearing at the respective parts in the case where the second-order filters are cascade-connected.

As shown in Fig. 11, in the period between clocks 0 and 16, a coefficient a,(t) and a multiplicand x,(t) are multiplied in the multiplier 53 to effect the multiplication in the secondorder filter 57 in Fig. 12, and the result of multiplication is obtained from a dth clock. In the period between clocks 1 6 and 32, as shown in Figs. 11 B and 11 C, a coefficient b,(t) and a multiplicand y1(t) are multiplied to perform the multiplication in the second-order filter 65.The multiplicand x,(t) is delayed by the shift register 51 along with 22 bits of the multiplier 53 by (176 + d) clocks, so that as shown in Fig. 11 E, a multiplicand x,(t -- 1) is applied to the adder 94 from the dth clock and added with the output a,x, derived from the multiplier 53 at that time, and the added output x,'(t) is provided via the adder 81 to the shift register 103 for accumulation. That is, the output from the adder 81 is supplied to the signal system of the adders 81, 82, ... in Fig. 12.

The output from the adder 94 is also provided to the (175-d)-bit shift register 95, as shown in Fig. 11 H. Accordingly, in the period between the clocks 0 and 16, the output from the shift register is x1,(t - 1), as shown in Fig. 11 H, and this output is added with the multiplicand x,(t) in the adder 52, the output x2(t) from which is applied as the input to the second-order filter 58 in Fig. 12. The output x2(t) from the adder 52 is provided via the shift register 101 to the multiplier 53. As shown in Fig. 11 C, the output x2(t) is multiplied by the coefficient a2(t) in the multiplier 53 in the period between clocks 32 to 48.Prior to this multiplication, b,(t) and y1(t) are multiplied, as described previously, and the multiplied output is similarly processed, thereby to obtain the output y2(t) from the second-order filter 65 in the period between clocks 48 and 64. In this way, the multiplication of the coefficent a and the multiplicand x and the multiplication of the coefficient b and the multiplicand y are carried out alternately every 16 clocks, and the multiplied results are applied to the shift register 51, as indicated by a,x" b1y1, a2x2, b2y2, ... in Fig. 11 D. Further, the second-order filters 57, 58, 59, 60 and 61 respectively derive therefrom x,'(t), x2'(t), x3'(t), x4'(t), x5'(t) and x2(t), x3(t), x4(t), x5(t), x6(t), which are provided to the shift registers 95 and 101. Similarly, Yr'(t) to y5'(t) and y2(t) to y6(t) are respectively obtained from the second-order filters 65 to '9, and these outputs are applied to the shift registers 95 and 101 alternately with x'(t) and xtt), respectively.In the period between clocks 145 and 161, the output y6 derived from the adder 52 at that time and x5 in the shift register provided previously are subtracted one from the other in the adder 43, and (x6 - y6) is supplied via the switch S5 to the shift register 101, wherein it is delayed by (d + 1) clocks. The delayed output is taken out from the switch S4 for input to the adder 43 in the period between clocks 147 + d and 1 63 + d. The output yielded from the shift register 103 at that time is provided to the adder 43 via the adder 81 and the switch S3.The output from the adder 43 at that time becomes the output from the adder 43 in Fig. 1 2 and this output is applied to the adder 45, wherein it is added with the input at the terminal 54 to provide Z(t). The added output Z(t) is supplied to the register 74, wherein it is delayed by the delay means 74 in Fig.

1 2. The delayed output is applied to the multiplier 53 and at that time the coefficent A is provided as an amplitude interpolation output at the terminal 44 and A Z(t) is derived from the multipier 53 at the output terminal 55. This multiplication is performed in the case where the output from the synthesis filter section 1 6 is multiplied by the amplitude information A in a multiplier 104 in Fig. 1 2. From the shift register 74 is taken out an output Z(t)/2 having shifted down by one bit and this is taken out via the switch S2 to the multiplier 53 as Z(t - 1 )/2, that is, x(t) and y(t), in the next subsequent operation period for a new set of parameters.The output of the output terminal 55 can also be obtained as parallel outputs through an output buffer 105 of a static shift register.

The pipeline operation described above is also applicable to other types of synthesis filter section 1 6. Furthermore, as will be appreciated from the arrangement of Fig. 10, the filter operation can be achieved by the addition, multiplication and delay, so that this filter processing can also be effected using a microcomputer. For example, in Fig. 13, by successively reading out, interpreting and executing programs in a program memory 107, a central processor unit 106 loads therein from an input port 111 a sound source signal and control parameters respectively applied from the sound source signal generating section 1 5 and the interpolating section 14 to terminals 108 and 109, and the central processor 106 sequentially performs the operations described previously with regard to Fig. 11.A read-write memory 11 2 is used instead of the registers 51, 74, 95, 101, 103 and 105 in Fig. 10. The results of the operations are written in the read-write memory 11 2 and read out therefrom at suitable timing to perform operations. The output thus obtained is applied from an output port 11 3 to the output terminal 55. The central processor 106, the memories 107 and 11 2 and the ports 111 and 11 3 are connected to a bus 114.

By any one of the abovesaid methods the output from the synthesis filter section 1 6 is obtained. The output is converted by the D-A converting section 1 7 in Fig. 2 to an analog signal to provide a speech output. In the D-A converting section 17, if the input thereto is a serial signal, then it is applied to a shift register 11 5 and the content of the shift register 11 5 is converted by a D-A converter 11 6 to analog form.

As described previously, the LSP parameters q and At in the speech feature parameters used in the present invention can be obtained by obtaining the solutions of the expressions (5) and (6). In Figs. 14A s 14B there are shown the results of analysis of a speech "bakuoNga" using the LSP parameters q and Aj. In Figs. 14A and 14B, the abscissa represents time t, in Fig. 14A the ordinate represents power, and in Fig. 1 4B the ordinate represents normalized angular frequency.Seeing instantaneous points in Fig. 14B, the frequency rises in the order of parameters Oi, a'i, a" z92, 6'i2, ... 05, a'5, this order does not change and the parameters 0 and a' do not coincide with each other in one frame. Accordingly, it is guaranteed that the synthesizing filter section 1 6 is always stable. The frequency distributions of the LSP parameters s9j and aij are shown in Fig. 15, in which the abscissa represents normalized angular frequency f and the ordinate the relative frequency D.As shown in Fig. 15, each parameter is not distributed over a wide frequency band but restricted to a relatively narrow frequency band, so that the LSP parameters , and 0 can be quantized in connection with the frequency range in which they are distributed.

The LSP parameters cXsj and 8, are little in quantizing distortion. Fig. 1 6 shows a spectral distortion Ds of a synthesized speech when various parameters were quantized variously, the abscissa representing the number of quantizing bits B per frame and the ordinate the spectral distortion Ds.The line 11 7 shows the case where in consideration of only the parameter distribution, the PARCOR coefficient is quantized linearly only in the coefficient was distributed; the line 11 8 shows the case where the number of quantizing bits for the PARCOR coefficient was increased in consideration of the spectral sensitivity in addition to the parameter distribution in the case of the line 117, especially in the case of markedly affecting the spectrum; the line 11 9 shows the case where the LSP parameters oi and Aj were quantized in consideration of only the parameter distribution; and the line 121 shows the case where the LSP parameters aij and A were quantized in consideration of the parameter distribution and the spectral sensitivity. It will be seen from Fig. 1 6 that in the case of using the same number of quantizing bits, the spectral distortion becomes smaller in the order of the lines 117, 118, 119 and 121. Since the lines 11 9 and 1 21 are close to each other, the LSP parameters zj and Aj are not so much affected in spectral distortion even if the spectral sensitivity is not taken into account. Accordingly, since it is sufficient to perform the quantization taking into consideration the parameter distribution range alone, the quantization is easy.The value that the number of quantizing bits per frame at which the spectral distortion is 1dB in the case of the line 11 9 is divided by that number of quantizing bits in the case of the line 11 7 is 0.7. Similarly, the ratio of the number of quantizing bits per frame at which the spectral distortion is 1 dB between the lines 11 8 and 1 21 is 0.8. From this, it will be understood that the LSP parameters xj and Aj are excellent. One dB is a difference limen of the spectral distortion of a synthesized speech.

Fig. 1 7 shows interpolation characteristics, the abscissa representing a frame length Tf and the ordinate the spectral distortion Ds. Fig. 1 7 shows the spectral distortion of a synthesized speech in the case where a frame in which an original speech was analyzed in 1 Omsec was used as the reference, the frame length was increased to 20 to 70 msec and parameters were interpolated every 1 Omsec.The line 1 22 shows the case where use was made of the PARCOR coefficients, and the line 1 23 shows the case where use was made of the LSP parameters a" and 01. As will be seen from Fig. 17, in the case of the same distortion, the frame length To can be made longer by the LSP parameters than the frame length To by the PARCOR coefficients, that is, the parameter update period can be increased, so that the entire amount of information can be reduced by that.In addition, since the LSP parameters are smaller than the PARCOR coefficients in the number of bits per frame, as seen from Fig. 16, the amount of information for the same distortion may be small by the product of the reduction ratios in Figs. 1 6 and 17; namely, in the case of the LSP parameters, the amount of information may be about 60% of that in the case of the PARCOR coefficients.

In the case of employing the LSP parameters, it is meaningless as in the case of other parameters that they are interpolated with a shorter period than the sample period of the original speech used in the making of the parameters. Experiments revealed that the interpolation period might be about twice or less the sample period of the original speech, but that when the former was about four times the latter, noises were introduced to make the synthesized speech indistinct. Accordingly, it is preferred that the interpolation period be equal to or twice the original speech sampling period.

As has been described in the foregoing, the LSP parameters are relatively easy to automati calls extract, and consequently can be extracted on a real time basis. Furthermore, the LSP parameters are excellent in the interpolation characteristic and small in deviation of the quantizing characteristic and permits transmission and storage of speech in a small amount of information. In the speech synthesis, speech of high quality can be reconstructed and synthesized with a small amount of information, and as long as the relationship of the expression (8) holds true, the stability of the synthesizing filter is guaranteed.

In Fig. 2, it is also possible to widen the spectrum by generating from the pulse generating section 28 a train of pulse groups, such as the Barker series, instead of the pulse train. The interpolating section 14 may also be provided at the preceding stage of the parameter transforming section 1 3. Namely, the LSP parameters from the interface section 12 may also be subjected to the cosine transformation in the parameter transforming section 1 3 after being interpolated. In this case, the use of a read only memory is uneconomical since its memory capacity need be enormous; accordingly, it is preferred to perform parameter conversion using an approximation operation of the cosine rather than using the read only memory as described in the example of Fig. 2.In Fig. 2, the information indicating whether speech is a voiced or unvoiced sound is entered and loaded in the voiced sound register 23 and the unvoiced sound register 24, but this information need not always be provided. That is, a detector circuit is provided for detecting whether the fundamental period parameter applied to the pitch register 25 is zero or not; in the case of detecting zero, the sound is decided to be an unvoiced sound and the gate 37 is opened; and in the case of other values than zero, the sound is decided to be a voiced sound and the gate 31 is opened. The control by the amplitude parameter may also be effected in connection with the output from the filter section 16, as described previously with respect to the embodiment of Fig. 1 2.

In the foregoing, as the synthesis filter, use is made of a filter which includes in the feedback circuit the means for connecting in series a plurality of first order and second-order filters of different coefficients, each having the zero on a unit circle, through utilization of the LSP parameters. However, the synthesis filter need not always be limited specifically to such a filter and the speech synthesis may also be effected by transforming the LSP parameters to some other types of parameters and using other filters. For example, as shown in Fig. 1 8 in which parts corresponding to those in Fig. 1 are identified by the same reference numerals, the fundamental period parameter in the feature parameters applied to the interface section 12 is provided to the sound source signal generating section 15, and the amplitude parameter is supplied to the interpolating section 14. The amplitude parameter thus interpolated is applied to the sound source signal generating section 15, in which it is processed as described previously in respect of Fig. 2, providing a sound source signal to the synthesis filter section 16. The LSP parameters are supplied to an LSP parameter transforming section 124, in which they are transformed to other types of parameters, such as an a parameter, PARCOR parameter or the like.For example, from the LSP parameters are obtained polynominals P(Z) and Q(Z) using the expression (5) or (6), and from the polynominals the predictor coefficients ai of the transfer function H(Z) are obtained using the expressions (1) and (2). By interpolating the thus obtained predictor coefficients a, in the interpolating section 14 as required, the characteristics of the sound synthesis filter section 1 6 are controlled.The filter section 1 6 is formed, for example, as a cyclic filter, in which, as shown in Fig. 18, the sound source signal from the sound source signal generating secdtin 1 5 is made a-fold by a multiplier 1 25 and applied to an adder 1 26 for subtraction from the output of an adder 1 27 and the ouput from the adder 1 26 is provided to the output terminal 55. The output thus derived at the output terminal 55 is applied to a series circuit of delay circuits D, to Dp, each having a delay time of one sample period. The outputs from the delay circuits D1 to Dp are respectively multiplied by coefficients oe, to ap from the interpolating section 1 4 in multipliers M, to Mp. The multiplied outputs are sequentially added and then added together in the adder 1 27.

It will be apparent that many modifications and variations may be effected without departing from the scope of the novel concepts of this invention.

Claims (18)

1. A sound synthesizer in which a sound source signal and control parameters for controlling the characteristics of a filter are applied to a synthesis filter section and filter coefficients of the synthesis filter section are controlled by the control parameters to obtain a synthesized sound signal, characterised in that the synthesis filter section is composed of second-order filter means serving as second-order filters, each having the zero on a unit circle in a complex plane, means for cascade-operating such second-order filter means of different coefficients, and feedback means for feeding back the output from the synthesis filter section to the input side thereof through two kinds of such cascade-operating means.

2. A sound synthesizer according to claim 1, wherein the sound source signal source is composed of a fundamental period sound source controlled by a fundamental period parameter to generate a pulse or a pulse group of the period indicated by the parameter, a noise course for generating random pulses, and select means for selecting taking out the output from the fundamental period sound source or the output from the noise source depending on whether speech to be synthesized is a voiced or unvoiced sound.

3. A sound synthesizer according to claim 1 or 2, further comprising amplitude control means for controlling the magnitude of a signal at the input or output side of the synthesis filter section by an amplitude parameter.

4. A sound synthesizer according to any one of claims 1 to 3, wherein the second-order filter means is composed of first delay means for delaying the input for a unit time, first adder means supplied with the delayed output and the output from the synthesis filter section, second delay means for delaying the output from the first adder means for a unit time, multiplier means for multiplying the output from the first adder means and the coefficient, and second adder means for adding together the multiplied output, the output from the second delay means and the input to the second-order filter means to provide the output from the second-order filter means.

5. A sound synthesizer according to any one of claims 1 to 3, wherein the second-order filter means is composed of first delay means for delaying the input to the second-order filter means for a unit time, multiplier means for multiplying the input to the second-order filter means and a coefficient thereof, first adder means for adding the multiplied output and the output from the first delay means, second delay means for delaying the added output for a unit time, and second adder means for adding the output from the second delay means and the input to the second-order filter means to provide the output from the second-order filter.

6. A sound synthesizer according to any one of claims 1 to 5, wherein the second-order filter means is formed as a second-order filter circuit; a plurality of such second-order filter circuits of different coefficients are cascade-connected to form the cascade-operating means; and a pair of cascade-connected second-order filter circuits of different filter coefficients form the two feedback means.

7. A sound synthesizer according to claim 4 or 5, wherein the second-order filter means is formed as a second-order digital filter circuit; and the second-order digital filter circuit is used on a multiplex basis by a pipeline operation system by operating the filter circuit a plurality of times within a unit time and changing the coefficient of the filter circuit for each operation.

8. A sound synthesizer according to any one of claims 1 to 5, wherein the filter means and cascade-operating means are formed by operating means for performing filter processing by interpreting and executing a program.

9. A sound synthesizer according to any one of claims 1 to 8, further comprising parameter transforming means for obtaining the control parameters by the cosine transformation of parameters for controlling the characteristics of the synthesis filter section.

10. A sound synthesizer according to any one of claims 1 to 9, further comprising interpolating means for interpolating the control parameters and supplying them to the synthesis filter section.

11. A sound synthesizer according to claim 9, which further comprises interpolating means for interpolating the parameters representing the characteristics of the synthesis filter section, and wherein the interpolated parameters are subjected to cosine transformation by the parameter transforming means.

1 2. A sound synthesizer according to claim 10 or 11, wherein the interpolation period in the interpolating means is equal to or twice the sample period of an original sound signal.

1 3. A sound synthesizer according to claim 10, 11 or 12, wherein the interpolating means is used on a multiplex basis for the interpolation of an amplitude parameter.

14. A sound synthesizer comprising: a sound source signal source for generating a sound source signal; a LSP parameter source for generating LSP parameters; parameter transforming means for transforming the LSP parameters to control parameters of a type different from the LSP parameters; and a sound synthesis filter section supplied with the sound source signal and controlled by the transformed control parameters in its characteristics.

1 5. A sound synthesizer according to claim 14, wherein the sound source signal source is composed of a fundamental period sound source controlled by a fundamental period parameter to generate a pulse or a pulse group of the period indicated by the parameter, a noise source for generating random pulses, and select means for selectively taking out the output from the fundamental period sound source or the output from the noise source depending on whether a sound to be synthesized is a voice or unvoiced sound.

16. A sound synthesizer according to claim 14 or 15, further comprising amplitude control means for controlling the magnitude of a signal at the input or output side of the sound synthesis filter section by an amplitude parameter.

17. A sound synthesizer according to any one of claims 14 to 16, wherein the parameter transforming means is means for transforming the LSP parameters to predictor coefficients, and wherein the sound synthesis filter section is a cyclic digital filter.

18. Sound synthesizing method, wherein sound source signal parameters representing a sound source signal and control parameters for controlling characteristics of filter means are supplied to a medium, the sound source signal is generated in accordance with the sound source signal parameters from the medium, the sound source signal is fed to the filter means while the characteristics of the filter means being controlled by the control parameters from the medium thereby producing therefrom a synthesized sound signal, characterized in that; ; a spectral envelope of original sound is approximated by a transfer function H(Z) of said filter means expressed by a = Ap(Z) 1 +a1Z+a2Z2+ eY2Z2 .. +aZP where Z = e i", a being a constant, a normalized angular frequency 2'rfAT, AT a sampling frequency, f frequency, p a degree of analysis, and a1 (i = 1, 2, ... p) predictor coefficients, said Ap(Z) is further expressed as a sum of two polynominals P(Z) and Q(Z) by Ap(Z) = + fP(Z) +0(Z)) where P(Z) = Ap(Z)Z Ap(Z)ZZPAp(Z1) Q(Z) = A,(Z) + ZZPA,(Z-1) said polynominals are each factorized and angular frequencies which render the polynominals zero are used as said control parameters.