JPH0525120B2 - - Google Patents

Description

[Detailed description of the invention]

BACKGROUND OF THE INVENTION The present invention generally relates to a method and apparatus for modifying the audio characteristics of synthesized speech to obtain a modified synthesized speech of any one of a plurality of speech sounds from a single, provided synthetic speech source. The device generates an audible synthesized speech from an original source of synthesized speech whose quality varies considerably and is affected by apparent age and/or gender differences of the intended speaker. Ru. In particular, the sampling period for digital audio data obtained from a synthesized audio source that includes imaginary voices that are obviously non-human voices, such as talking voices of animals, birds, monsters, etc., or multiple audio sounds with unusual sound quality. Created from a single synthesized voice source by simulating the adjusted state and changing the vocal tract model of digital voice data to a predetermined degree without changing the bit period and voice rate inherent in the original synthesized voice source. be able to. Speech analysis researchers generally expect to be able to vary the acoustic features of a speech signal by changing the obvious speech features associated with the speech signal. Accordingly, there is a related article entitled "Speech Analysis and Synthesis by Linear Prediction of Speech Waveforms" by Attal and Harner, Annual Report of the Acoustical Society of America, Vol. 50, No. 2 (Part 2), pp. 637-650 (published in April 1971). In the publication,
A method for simulating a female voice from an audio signal obtained from a male voice is described. In this method, certain acoustic features of the original speech model, such as pitch, formant frequency and its band, are changed. In the publication ``Speech Sounds and Characteristics'' (1973), published by MIT Press, Cambridge, Mass., on pages 84-93, phanto is referred to as k-factor or ``gender factor'' for male and female formants. The extracted correlations are explained and it is suggested that these k factors are correlated with vowels of a particular class. Additionally, U.S. Pat. No. 4,241,235 to Matskiany, issued December 23, 1980, discloses a speech modification system based on real human speech as opposed to synthetic speech. In this, the original audio sound is changed in such a way that another audio sound that is distinctly different from the original audio sound is generated. In this audio modification system, the audio signal source is a microphone or a connection to any source of live or recorded audio or audio signal. This type of audio modification system requires less storage space when the spoken or recorded audio may be modified directly and when the audio is recorded since the entire content of the audio is contained within a relatively short period of time. The application is limited to cases where the One of the speech synthesis techniques that has recently attracted attention is linear predictive coding. Linear predictive coding methods provide an acceptable trade-off between sound quality and the amount of data required for analysis and synthesis, while still maintaining an acceptable degree of flexibility in independently controlling acoustic parameters. I understand. Text-to-speech conversion systems based on speech synthesis have the ability to provide synthesized speech with virtually no vocabulary restrictions, since they are drawn from a pre-stored component sound library consisting of, for example, allophones or phonemes. I have it. Typically, a component sound library consists of a read-only memory containing digital audio data representative of audio components extracted from an adult male voice and by which words, phrases, and sentences can be formed. The reason for selecting male voices for this purpose is that adult male voices typically present a low pitch profile that is considered optimal for currently used speech analysis software and speech synthesis equipment. be. Text based on synthetic speech created from a male voice by changing the audio characteristics of the original synthetic audio source and producing audio sounds with multiple different audio characteristics depending on the identity of the characters in the sentence. The voice conversion system changes the voice characteristics depending on whether the characters in the text are the same or not.
Providing audible synthesized speech can further increase flexibility without increasing memory capacity. In this regard, co-pending U.S. patent application Ser. There is. The technology for converting the voice characteristics of synthesized speech disclosed in the latter U.S. application separates the pitch period, vocal tract model, and speech rate contained in the source of synthesized speech into their respective voice parameters, and separates the vocal tract model from the original. The values of pitch and audio data rate are then varied in a predetermined manner determined by varying the sample rate by a predetermined amount while remaining in the form of . The altered speech data parameters are further recombined with the original vocal tract model to create modified synthetic speech data formants having different speech characteristics than the synthesized speech obtained from the speech source. Therefore, in the above-mentioned U.S. Patent Application No. 375,434 filed on June 6, 1982,
One form of the preferred embodiment includes techniques that actually vary the sample rate. In this, the modified sample rate is used for the original pitch period data and the original speech rate data, and the modified pitch period and modified speech rate are created and recombined with the original vocal tract speech parameters. A modified audio data format is then created. From this format, an audible synthetic human voice is produced by a speech synthesizer and an audible means that has different audio characteristics than the synthesized human speech that would have been obtained from the original synthesized speech source. . SUMMARY OF THE INVENTION In accordance with the present invention, a method and apparatus are provided for modifying the audio characteristics of synthesized speech to obtain a modified synthesized speech of any one of a plurality of speech sounds from only one provided synthetic speech source. be done. The above method differs from the above-mentioned 1982 method in that it modifies individual speech parameters, including pitch period, vocal tract model, and speech rate, related to the source of the original synthesized speech separately, without separating them, and does not actually adjust the sample period. This is quite different from the technical measures taken in US Patent Application No. 375,434, filed on June 6th. Instead, the invention works by setting first and second reference factors of unequal magnitude. A first reference factor therein is based on the desired modified synthesized speech to be produced, and an adjustment of the sample period of digital audio data obtained from the source of the synthesized speech is based on the first reference factor and the second reference factor. Simulate based on the difference between By simulating the adjustment of the sampling period of the digital data obtained from the original synthesized speech source, it is possible to effectively change the vocal tract model of the digital audio data to a certain extent without changing the pitch period or voice rate. can. The modified digital audio data produced by simulating the sample period adjustment has altered audio characteristics compared to the synthesized speech produced from the synthesized speech source.
A speech synthesizer receives the modified digital speech data and generates an audible signal representative of human speech. This is converted by an audible means, such as a loudspeaker, into an audible synthesized voice in which the voice characteristics have been modified from the synthesized voice input from the source of the synthesized voice. the first and second reference factors by simulating the adjustment of the sample period of digital audio data obtained from a source of synthesized speech depending on whether the first reference factor is greater or less than the magnitude of the second reference factor; The synthesized speech spectrum is reduced or expanded by a predetermined amount set by the difference in comparing the magnitudes of the reference factors. Therefore, when the first reference factor is greater than the magnitude of the second reference factor, the synthesized speech spectrum is reduced by simulating adjustment of the sample period of the digital speech data obtained from the source of the synthesized speech. Instead, if the magnitude of the first reference factor is small compared to the second reference factor, the synthesized speech spectrum is expanded. In either case, the plurality of prediction coefficients obtained by appropriately transforming the reflection coefficients containing the vocal tract model represented by the digital audio data of the first phase are first filled with a predetermined number of blanks. A value is added. The digital audio data is then converted from the first phase to a second phase that includes a plurality of added blank values therein. A series of digital signals depending on whether the difference between the first and second reference factors is negative or positive in simulating adjustment of the sample period after the series of digital signals has been changed from the time domain to the frequency domain. The signal may be either compressed or expanded. A digitized audio waveform is then created from the digital audio data present as a compressed or expanded audio spectrum in response to pulses. Pitch period and amplitude information is removed from these spectra by converting them from the frequency domain back to the time domain. This digitized audio waveform was further analyzed.
Modified digital audio data is provided having a modified vocal tract model having a plurality of digital values indicative of reflection coefficient parameters. At least some of the reflection coefficient parameters have been changed in magnitude when compared to digital values representative of the reflection coefficient parameters of the digital audio data obtained from the original synthesized audio source. Thus, by using the method and apparatus according to the invention it is possible to create all kinds of audio sounds from a single synthetic audio source. The sounds in this method and apparatus can generally be generated by interpreting the characteristics of the sounds as if they were spoken by an imaginary talking animal, such as a squirrel or a squirrel. In this case, the synthesized speech spectrum is expanded, thereby increasing the formant frequencies of the digital speech data. This simulates the reduction of the vocal tract model, and the audible synthesized speech generated therefrom can give the impression of being spoken by a small-sized creature or person. On the contrary, by spectrally compressing the synthesized speech spectrum, the formant frequencies of the digital speech data obtained from the original synthesized speech source are reduced, thereby widening the vocal tract and making the synthesized speech a monster. It can give the impression that the sound is generated by something large, such as a monster or a demon. In addition to changing the spectrum of the synthesized speech, the pitch parameter and the magnitude of the pitch integral path are modified to further increase the scale of modification of the speech features without actually changing the sample rate of the digital speech data. is also taken into account. DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION The invention will now be described in detail with respect to preferred embodiments with reference to the figures. Referring more particularly to the figures, the methods and apparatus disclosed herein produce a modified synthesized speech that is any one of a plurality of speech sounds that clearly differ in age and/or gender of the speaker. It is possible to modify the speech characteristics of synthesized speech from a given single synthesized speech source used within a fixed sample rate linear predictive coding (LPC) speech synthesis system by creating a fixed sample rate linear predictive coding (LPC) speech synthesis system. In particular, some speech sounds created from a single synthesized speech source in accordance with the techniques of the present invention may be non-human, e.g., sounds that appear to be spoken by an animal (e.g., a squirrel or a squirrel) that exhibits a high incidental pitch. It also includes unusual vocal sounds that seem to be the voices of things. Another purpose of the synthesized speech spectrum is that the speech sounds produced in accordance with the present invention are characterized in such a way that they are perceived to be of very low pitch and have an tonal quality similar to that spoken by demons, monsters, etc. It can be created. The central idea of the technology of the present invention is to modify the vocal tract model of digital voice data to a predetermined degree by simulating the adjustment of the sample period of digital voice data obtained from a source of synthesized speech, and thereby The audio characteristics of the audible synthesized speech are changed. This audible synthesized speech is generated by an audible transmission means in the form of a loudspeaker connected to the output of the speech synthesizer to which the modified digital speech data is sent. As can be seen, FIG. 1c shows the digital audio data of the source of the synthesized speech in which the synthesized speech spectrum is associated with normal speech features that have not been modified by either compression or expansion in accordance with the techniques disclosed herein. FIG. 3 is a diagram showing a synthesized speech spectrum obtained from FIG. Figures 1a and 1b each show expanded forms of the original synthesized speech spectrum shown in Figure 1c, and Figure 1c shows approximately the synthesized speech spectrum.
Shows 36% expansion and actual sample period is
A corresponding change in the spectrum occurs when changing from 125 microseconds to 80 microseconds. Figure 1b shows a 16% extension of the synthesized speech spectrum of Figure 1c, corresponding to a change in the sample period from 125 microseconds to 105 microseconds. It shows changes in the synthesized speech spectrum. FIG. 1d shows a compressed synthesized speech spectrum that is approximately 20% smaller than FIG. 1c. Among these, the synthesized speech spectrum changes to the same extent as when changing the sampling period from 125 microseconds to 150 microseconds. In general, the broadening of the synthesized speech spectrum of Figure 1c, the effect of which is illustrated in Figures 1a and 1b, respectively, increases the formant frequency and simulates a vocal tract of reduced size, and the audible synthesis produced therefrom increases the formant frequency and simulates a vocal tract of reduced size. The impression can be given that the voice is spoken by a relatively small person. Conversely, the compression of the synthesized speech spectrum shown in Figure 1c, the result of which is shown in Figure 1d, causes a reduction in the formant frequencies, thereby simulating an enlarged vocal tract, and the resulting audible synthesis. The impression can be given that the voice is spoken by a relatively large person or thing. A further description of the figures 1a to 1d is given below in the detailed description of the method and apparatus for modifying the audio characteristics of synthesized speech from only one given synthesized speech source in accordance with the present invention. I will continue to do so. The audio parameters, including the k audio parameters indicative of pitch, energy and reflection coefficients, such as the initial source of the LPC synthesized speech, can be retrieved from a single source, such as the read-only memory 10 (shown in Figure 5). can. . This source is
Stored therein are digital voice data and appropriate digital control data for selective use by voice synthesizer 11 to generate analog voice signals representative of human speech. In accordance with a preferred form of the invention in this regard, adjustment of the sample period of digital audio data is simulated by transforming the synthesized audio spectrum. In this case, the input and output LPC audio parameters are in the form of digital audio data representing reflection coefficients,
The order of the LPC model is N, F _OLD indicates the sample frequency inherent in the LPC parameters before transforming the synthesized speech spectrum, and F _New indicates the desired explicit sample frequency after transforming the synthesized speech spectrum. . The first reference factor P and the second reference factor Q are then used in the process of simulating the adjustment of the sampling period, so they should be the nearest whole numbers to the value of Q=P・F _NEW /F _OLD . selected.
Q must be a round number so as not to generate complex impulse responses during intermediate stages of the method. In the flowchart in Figure 2, first, the audio parameters k ₁ , k ₂ ... k _N indicating the reflection coefficient are set to 20
Markel and Gray's publication ``Linear Predictive Coding'', published by Springer Books, New York, Heidelberg, and Berlin.
Predictive coefficients can be calculated using established processing methods such as "step-up processing" described on pages 94-95.
Converted to a ₀ , a ₁ , ...a _o . Thereafter, P- _(N+1) artificial blank values or zeros are added to the prediction coefficient sequence at 21 to define a sequence designated as a ₀ , a _{1 .} . . a _N , o, o, . . . o. These are a ₀ , a ₁ , ...
It can be expressed as a _N , a _N+1 , a _N+2 ... a _P-1 . In order to determine the discrete Fourier transform (DFT) of a digitized speech waveform with a number of points corresponding to the first reference factor P, corresponding to k speech parameters and whose prediction coefficients, including the added blank values, correspond to the first reference factor P. use. In this case, as a means for simulating the adjustment of the sample period of digital audio data and changing the audio characteristics, the first reference factor P and the second reference factor Q are set as explained above, and the adjustment of the sample period is The magnitudes of these factors are determined based on the desired audio characteristics obtained from the modified digital audio data produced by the simulation. Therefore, the first reference factor, P, will be equal to a predetermined number of points determined by the type of speech desired to be produced, whereas the second reference factor, Q, will be equal to the inverse discrete Fourier transform (IDFT).
This is the number of points used for. In this case, the second reference factor Q affects the storage capacity limit of the memory and the processing speed of the device that changes the voice characteristics of the synthesized speech,
Increasing the amount of Q improves the quality of decomposition of the audibly generated modified synthesized speech. In order to transform the synthesized speech spectrum according to the invention, the first reference factor P and the second reference factor must not be equal in magnitude. In the special case where P and Q are equal, the synthesized speech spectrum is not transformed from that obtained from the original synthesized speech source, which means that P/Q is equal and the ratio is 1.00.
, the condition shown in the diagram of FIG. 1c, where the effective sample period is 125 microseconds. After setting the respective sizes of the first and second reference factors P and Q, the P-point DFT of the prediction coefficient sequence with the blank value added is calculated, and this results in the When DFT is performed to set digital signal data in the frequency domain in 22,
Blank values added in previous steps of the method are effectively absorbed or lost. at point P
Calculation of the DFT is performed using suitable techniques such as those described in "Digital Signal Processing" by Otzpenheim and Schieffer, published by Prentice Hall. At this stage, the individual audio parameters are represented as R ₀ , R ₁ . . . R _P-1 .
Here, the inverted value of R ₁ is the digital audio value R ₀ , R ₁ , etc. obtained by calculating the P-point DFT of the reflection coefficient.
...calculated in 23 by inverting R _P-1 . This essentially converts the digital audio data from that used in a reverse synthesis filter to one that can be used in a forward synthesis filter. Here, the digital audio data is expressed as values s ₀ , s ₁ . . . s _P-1 . At this stage, the transfer function H(z) of the digital filter is transferred into the frequency domain and the digital audio data is placed in a form comparable to the untransformed synthetic audio spectrum. According to the present invention, the method disclosed herein comprises:
Contains digital audio data representing reflection coefficients and can be used to create a transformed synthetic audio spectrum. To this end, the synthesized speech spectrum is then compressed or expanded at 24 in FIG. 2 based on a comparison of the magnitudes of the first and second reference factors P and Q. The difference in magnitude of P and Q can simulate adjustment of the sample period and change the audio characteristics for the synthesized speech. P=Q as illustrated in Figure 1c where the ratio of P/Q is = 1.00
In this case, no voice change occurs because the synthesized speech spectrum is unconverted and is the same spectrum as the original digital speech data obtained from the original synthesized speech source. For P>Q, where the P/Q ratio is greater than 1.00, the synthesized speech spectrum obtained from the original source is compressed, and this compression reduces the center frequency of the formant and its band, as shown in the graph of Figure 1d. Decrease. In this case, P of digital audio data
−Q samples are removed from the center of the spectral sequence denoted by the signals s ₀ , s ₁ ...s _P-1 , and the sequence s′ _i i
=0, Q-1 can be obtained. For example, if the magnitude of the first reference factor P is chosen to be 256 and the magnitude of the second reference factor Q is chosen to be 150, then s′ _i
The terms of the signal s _i modified to obtain s i take the form shown below, so the terms deleted from the sequence s ₁ to form the sequence s′ ₁ are taken from the center of the spectral sequence. removed.

[Table] Formally, the above changes can be shown as follows. s' _i = s _i , i=0, Q/2-1 s' _i = s _i+PQ , i=Q/2, Q-1. Q>P such that the ratio of P/Q is less than 1.00. If the synthesized speech spectrum is to be expanded, Q-P samples each with a value of zero are added to the center of the spectral sequence s _i to form the sequence s′ _i , i=
0, Q-1 can be obtained. For example, when P is assigned 256 and Q is 400 to the magnitudes of the first and second reference factors, a conversion of terms from the sequence s _i to s' _i occurs as shown below.

[Table] Formally, this can be shown as follows. s' _i =s _i , i=0, P/2-1 s' _i =0, i=P/2, Q-1-P/2 s' _i =s _i+PQ , i=Q-P/ 2. Q-1. This technique explicitly changes the speed of the signal containing digital audio data without actually changing the processing speed, and thereby reduces the speed of such samples rather than actually changing the sample rate. It simulates the state where the rate is changed. At this stage, the number sequence in step 25 in Figure 2
s′ ₀ , s′ ₁ , s′ ₂ ... Q-point individual inverse Fourier transform (IDFT) is performed on s _Q-1 and calculated, and a series of signals h ₀ ,
h ₁ , h ₂ ... h _Q-1 is set. This series of signals provides the desired impulse response to a speech synthesis filter in which the linear predictive coding speech parameters are modified to simulate changes in sample rate.
Thereby, the synthesized speech spectrum can be returned from the frequency domain to the time domain where the speech data exists as a digitized speech waveform that does not have pitch information or energy information. The digitized audio waveform described above is similar to the digitized audio used in the audio analysis part. In a preferred embodiment, the magnitude of Q can be defined as a power of two. This is because it allows the use of a special form of IDFT (Inverse Fast Fourier Transform) instead of the more general IDFT used in the compression or expansion process of the synthesized speech spectrum shown at 24 in Figure 2. IFFT). The speed of performing signal processing techniques is significantly increased when IFFT is performed. In this example, P is equal to the nearest full integer value to QF _OLD /F _NEW . A device that changes the characteristics of audio by using the IFFT format is
It is possible to have a processing speed approximately proportional to Q, logQ. On the other hand, the processing speed when IDFT is used is proportional to Q ² . The series of signals h′ ₀ , h′ ₁ , h′ ₂ ……h′ _Q-1 is represented by 2 in Fig. 2.
As shown in 6, the reflection coefficients are analyzed and changed by substituting them into the Nth-order linear predictive coding formula.
By obtaining digital voice data indicating k' ₁ , k' ₂ , k _{' 3} . By fitting the series of signals h′ ₀ , h′ ₁ , h′ ₂ ……h′ _Q-1 with the N-th order LPC equation, we obtain k′ ₁ , k′ ₂ , k′ ₃ ……k′ _N. The technique described in Markel and Gray's publication "Linear Prediction of Speech", pp. 10-15, indicated above, is performed in setting the digital values representing the modified vocal tract model, and the prediction coefficient a′ Obtain digital audio data indicating _i . These data are further
At step 27 in the figure, the signal is converted into a digital audio value indicating the reflection coefficient k' _i as described on pages 95-97 of the above-mentioned publication. Therefore, Figures 1a and 1b are diagrams showing expanded versions of the original speech synthesis spectrum shown in Figure 1c, where the amount of Q is greater than the amount of P, and Figure 1d is a diagram showing the extension of the original speech synthesis spectrum shown in Figure 1c, where the amount of Q is greater than the amount of P. FIG. 4 is a graph showing a compressed synthesized speech spectrum where Q is larger than the amount of Q. Referring to FIG. 3, the first reference factor and the second
Logic for explaining in more detail the step of determining the sequence of numbers shown at 24 in FIG. A diagram is shown. For this _purpose , the series of signals determined by the ²³ phases of _FIG . is received as an input to a comparison device 30 on which a threshold value is set based on. If such an inequality sign is true, the comparison device 30 gives an output to the control circuit 31, and the control circuit 31
performs the steps of removing samples P-Q from the central portion of the signal series to produce a series of signal outputs: {s' _i } ^Q-1 _i=0 . If, on the other hand, comparison unit 30 determines that the proposition that P is not greater than Q is false, comparison unit 30 provides a selection output to a second comparison unit that has a threshold based on P being less than Q. If such inequality is true, the comparison unit 32 provides an output to the control circuit 33 which adds the blank value of Q-P as a complex zero to the center of the series of signals, {s' _i } ^Q-1 Create a series of signals by converting these signals indicated by _i=0 . If the inequality that P is greater than Q is false, since this means that P and Q are equal, the second comparator circuit 32 generates a series of unconverted signals as a selection output. As explained in connection with FIGS. 2 and 3, the compression or expansion of the synthesized speech spectrum obtained from the original synthesized speech source, as the case may be, is performed from the center of the series of spectra s _i to This is done by obtaining a transformed synthetic speech spectrum by removing sample values or adding a blank value of Q-P to the center of the series of spectra s _i . In this case, the complete series of spectra s _i characteristically comprises the spectral parts of the first and second series. Therein, the second series of spectral portions is a "mirror image" of the first spectral portion. Therefore, carrying out the method according to the invention only on the first series of spectral parts,
It is possible that the second series of spectra portions of the complete series of spectra s _i are not performed. This method combines the steps of adding and subtracting sample values from the original synthesized speech source to simulate sample period adjustments by compressing or expanding the synthesized speech spectrum into a complete series of spectra s _i This eliminates the need for the complicated process of performing this step on the center of the spectrum, and provides a practical method of performing it on the tail end of the first series of spectra. Therefore, if the above steps are used to operate only on the first series of spectral parts of the complete series of signals s ₁ as a series of signals, the method disclosed herein can be used to change the phonetic characteristics of the synthesized speech. The circuit of the device for this purpose can be simplified. It can be seen that if the first series of spectral parts is used as the series of signals s _i , the number of sample values removed or blank values added is halved. Therefore, in FIG. 3, for example, comparison unit 3
When 0 indicates that the inequality P>Q is true, control circuit 31 commands the deletion of P-Q/2 sample values from the end of the series of signals s _i . Instead, if the inequality P<Q is true, then
Control circuit 33 commands the addition of a blank value of Q-P/2 to the end of the series of signals. In connection with the method presented later, FIG. 4 shows an apparatus for modifying the speech characteristics of synthesized speech obtained from only one given synthesized speech source according to the invention. The device in this figure operates on the last part of the series of signals defined by the first series of spectral parts of the complete series of spectra s _i . Therefore, when the first reference factor P is greater than the second reference factor Q, the device shown in FIG. When P is less than the second reference factor Q, a blank value of Q-P/2 is added to the end of the series of signals. Referring to the apparatus of FIG. 4, the apparatus receives P point individual Fourier transform values and outputs Q point individual Fourier transform values as output. If the first reference factor P is greater than the second reference factor Q, the series of inputs can be truncated to obtain a series of outputs. On the other hand, if P is smaller than Q,
An artificial sample value with a value of zero is added to the end of the series of inputs to produce a series of outputs. The first and second reference factors P and Q are determined in relation to only the first series of spectral parts of the complete series of spectra (and therefore are determined for P and Q in the entire complete series of spectra). ), P-Q sample values are removed from the end of the series of inputs, or Q-P blank values are added to the end of the series of inputs. As shown in the diagram, each value in the series is 16-bit data so that two identical 8-bit component devices can be used in pairs as needed to perform an equivalent 16-bit function within the device circuitry. expressed. It will be appreciated that a single component having the required bit capacity may be used in place of the pair of components shown. For example, the single comparison unit 30 (shown in FIG. 3) could be replaced by comparison units 30a, 30b set to threshold Q-1. The apparatus of FIGS. 4a-4c includes a switching device which may be in the form of a JK flip-flop, such as the SN7474 integrated circuit available from Texas Instruments, Inc. of Dallas, Texas. The J-K flip-flop 40 selectively controls the device with a reciprocal generator and a reciprocal generator operated at step 23 of the method shown in FIG.
and an inverse separate Fourier transform processor operating at the output of the 24 stages of synthetic speech spectral transformation performed. When a control switch occurs between the reciprocal generator and the IDFT processor, comparators 30a and 30b generate pulses to clear counters 41a and 41b. When the control of the reciprocal generator shown in step 23 is taking place, memory means in the form of random access memories 42a, 42b are set for writing.
On the other hand, RAMs 42a and 42b are set for read-only access. The counters 41a and 41b are increment counters that calculate coefficients from 0 to Q-1, and store respective frequency values in accordance with the counts in the RAMs 42a and 42b.
If the count is less than the value of P, the comparison units 32a, 32b connect the control lines to multiple latch circuits 33a, 3 (e.g., corresponding to the control circuit 33 of FIG. 3).
Since it is set for 3b, the data output from the reciprocal generator is stored in the RAMs 42a and 42b. When the single count reaches the value P, the multiple latch circuits 33a and 33b pass a blank value of 0 and store it in the RAM 42 for each count.
a, 42b. The and inputs of JK flip-flop circuit 40 are both set to a logic "0" which pulses each CK input to tie the value to Q. Q is logic “0”
(=“1”) and the timing pulse given from the reciprocal number generator is used to control the device circuit.When Q has the logical value of “1” (=
0), the timing pulse given from the IDFT processing device is used to control the device circuit. As explained, two 8-bit counters 41a,
42b is configured to form a single 16-bit counter (by connecting between the RCO output of the least significant coefficient counter and the input of the most significant counter). Upon receiving the appropriate timing pulse from either the reciprocal generator or the IDFT processor, counters 41a and 41b are incremented by one as long as the inputs have a logic "1" value. If the input has a logic "0" value, the timing pulse resets the counters 41a, 41b. (Both 8-bit counters 41a,
41b is assumed to have a value of zero) comparator circuit 30a,
30b indicates the current values of counters 41a and 41b as Q-
Compare with the value of 1. When the counters 41a, 41b reach this value, the comparison devices 30a, 30b =
The Q output has a logic "0" value and is thereby connected to the inputs of counters 41a, 41b.
The output of OR gate 43 becomes logic "0". The next timing pulse therefore resets the calculators 41a, 41b. RAMs 42a and 42b are constructed by two pairs of electrostatic RAMs available as part number TMS-4016 integrated circuits from Texas Instruments, Inc., Dallas, Texas, each capable of providing a storage capacity of 2048 x 8 bits. It has the ability to store 2046 x 16 bit values in total. Counters 41a and 41b are used as RAM address circuits. RAM42a, 42b
is connected to a logic inverter 44, which further inverts the reciprocal generator timing pulses and J-
It is connected to an AND gate 45 for producing an AND with the output of the K flip-flop device. When Q has the value of logic "1" (and the timing pulse of the reciprocal generator has the value of logic "1"), the value received from the reciprocal generator is stored in RAM 42a, 42b.
is stored in the . When has a logic “0” value, the value is stored in RAM4 for use by the IDFT processing unit.
2a and 42b. Comparing devices 32a, 32b are counters 41a, 41
The current value of b is compared with the value P-1. If the current value of the counters 41a, 41b is less than or equal to the value of P-1, the multiple latches 33a, 33b
The A/input of is set to logic "1", so that the Y output of multiplex latch 33a, 33b becomes
The Y outputs of multiple latches 33a and 33b are set to the data value obtained from the reciprocal generator, and the Y outputs of multiple latches 33a and 33b are output to RAM 42.
a, 42b as data. If the value of the counter is greater than the value of P-1, the multiple latch 33
The A/input of multiplex latches 33a, 33b is set to logic ``0'', which causes the Y input of multiplex latch 33a, 33b to
The output is set to a logic ``0'' value. The CLK (clock) input of multiple latches 33a and 33B is
It is connected to an AND gate 45 which ANDs the reciprocal generator timing pulse and the output of the JK flip-flop device 40. has a logic "1" value and the timing pulse of the reciprocal generator is generated, the multiplex latches 33a, 33b transfer the blank value of zero to the RAMs 42a, 4.
2b and continues to transfer to each counter until the value of the counter reaches the value of Q-1. On the other hand, since the Y outputs of multiplex latches 33a and 33b are set to a high impedance state, data can be read from RAMs 42a and 42b when the IDFT processor is controlled. Counters 41a and 41b are integrated circuits SN74LS592
It has a pair of 8-bit counters available as 8-bit counters. On the other hand, a pair of 8-bit comparators can be provided by the integrated circuit SN74LS684, and a pair of multiple latches can also be provided by the integrated circuit SN74LS606.
All of these integrated circuits are available from Texas Instruments, Inc., Dallas, Texas. The apparatus shown in FIG. 4 changes the audio characteristics of the synthesized speech disclosed herein by converting the synthesized speech spectrum to simulate the state in which the sample period of digital speech data obtained from the source of the synthesized speech is adjusted. Although specifically described with reference to a suitable circuit system for implementing the method, it will be appreciated that any suitable general purpose computer may be used for this purpose. FIG. 5 is a functional block diagram of the speech synthesis system. Therein, the audio feature changing device of FIG. 4 is combined in accordance with the invention. It can be seen that FIG. 5 illustrates a general purpose speech synthesis system. This system, for example,
may be part of a text-to-speech system such as that disclosed in copending U.S. patent application Ser. In some cases, it is a complete speech synthesis system without any functionality. For this purpose, the speech synthesis system of FIG.
0, this memory has digital audio data and digital control data stored therein which are controlled by a controller 12 in the form of a microprocessor and selectively accessed by the speech synthesizer 11. ing. As explained herein, the digital audio data contained in the audio ROM 10 represents reflection counts and has a single synthesized audio source. This synthesized speech source is the speech synthesizer 1
Audio data can be processed using the linear predictive coding method used by No. 1 to obtain an analog audible signal representative of human speech. The digital audio data stored in the POM 10 is indicative of complete words or parts of words, such as allophones or phonemes connected in a series controlled by the microprocessor 12, and the storage capacity of the ROM 10 is It is possible to form a series of audio data representing a very large number of words compared to the following. The voice ROM 10 is directly connected to the voice synthesizer 11, but is selectively determined by the operation of the control device 11, and is connected directly to the voice synthesizer 11.
It will be appreciated that configurations are possible that also include digital data received by the synthesizer.As shown in FIG. 11. The controller 12 is programmed for word selection and audio feature selection for each word, and the digital audio data accessed by the controller 12 from the audio ROM 10 is output from the ROM for a given word (which is a series of (including strings of allophones or phonemes). A predetermined audio feature profile is applied to this output by setting the difference in magnitude of the first and second initial factors P and Q. When P=Q, the voice
The audio characteristics of the digital audio data stored in the ROM 10 do not change at all, and the digital audio data is selectively accessed by the audio synthesizer 11 under the control of the controller 12 via the conductive part 12a. Ru. Suitable audio means 13, such as a suitable bandpass filter, a preamplifier 14, and a speaker 15 are connected to the output of the speech synthesizer 11 so that the analog audio signal produced by the speech synthesizer 11 can be audible. A synthetic human voice is created. The microprocessor forming controller 12 is of any suitable type, such as the TMS7020 manufactured by Texas Instruments, Inc. of Dallas, Texas. This microprocessor is optionally available as an audio ROM component TMS6100 from Texas Instruments, Inc., Dallas, Texas.
Voice data and digital command data are accessed from 10. Speech synthesizer 11 processes digital audio data using linear predictive coding and generates an analog signal output representative of synthesized human speech. This speech synthesizer 11 is available as component TMS5100 from Texas Instruments, Inc., Dallas, Texas, disclosed in U.S. Pat. You can use any type. According to the invention, a cooperating audio feature modification device 1
7 is located between the control device 12 and the speech synthesizer 11. The audio feature changing device 17 of the signal processing device 16 corresponds to the device circuit shown in FIG.
As explained above, when the digital audio data read from the ROM 10 is sent to the inside of the signal processing device 16 via the conductive section 12b, and the output from there is sent to the speech synthesis device 11 through the conductive section 12c, the audio A transformation of the composite spectrum is performed. As previously explained, depending on the magnitudes assigned by the microprocessor 12 to the first and second reference factors P and Q, the audio feature modifying device 17 changes the audio characteristics from the audio ROM 10 by the microprocessor 12. A modified K′ audio parameter indicating the reflection coefficient is generated with reference to the initially accessed K audio parameter. The modified K' voice parameter input to the voice synthesizer 11 is transmitted to the speaker 15.
The function is to change the characteristics of the audible synthesized speech generated by the . In this example, the modified vocal tract model based on the digital audio data determined by the modified K' audio parameters is combined with the original pitch period and audio rate of the synthesized audio source to be heard by the loudspeaker 15. The predetermined pitch period and predetermined speech rate remain unchanged to produce a synthesized speech with modified speech characteristics as an output that can be heard. Regarding the later embodiments, the K speech parameters defining the vocal tract model of the original synthesized speech source are passed through the signal processing device 16 and the speech feature modification device 17 via the conductive portion 12b as modified K′ speech parameters. The K audio parameters are input to the speech synthesizer 11 via the conductive section 12c, while the pitch and energy parameters are input to the speech synthesizer 11 via the conductive section 12a, bypassing the signal processing device 16. 12 from the pitch and energy parameters that cooperate within each frame of audio data. An alternative method is that the pitch and energy parameters are passed through the signal processing device 16 without being manipulated by the conductive part 12b and input into the speech synthesizer 11 along with the modified K' speech parameters. You can also make it so that However, if the pitch parameter is encoded within a unit period of the sample period, to simulate the adjustment of the sample period and transform the synthesized speech spectrum, it is necessary to maintain the same pitch frequency as before the conversion of the synthesized speech spectrum. It is necessary to adjust the encoding pitch value. This adjustment is performed by multiplying the original encoded pitch value by the ratio Q/P. For example, in Dallas, Texas,
The TMS5100 speech synthesis component available from Instruments, Inc. requires the step of weighting such encoding pitch parameters. No weighting is required if the pitch parameter is encoded in units of frequency or other units, such as units of time between pitch pulses that occur intermittently in milliseconds. It can also be interpreted as a voice created by the difference in the age or gender of the person speaking, but the altered vocal characteristics of the synthesized voice created using this method can be interpreted as the voice of an animal, bird, animal, bird, etc.
The sound quality tends to be such that it seems that something other than human beings is generated, such as a monster or demon, which makes one imagine that it is generated from an imaginary or unusual source. As mentioned earlier, the magnitude of audio feature changes that can be performed without changing the sample period on digital audio data is distinct from the transformation of the synthesized audio spectrum by simulating sample rate adjustments. It can be adjusted by independently modifying the amount of the pitch parameter and the pitch integral path. In this connection, the method of the present invention is a continuation of the above, since it is possible to modify the vocal tract model pitch parameters pitch integral path separately and produce speech with several phonetic features from a single given input source. U.S. Patent Application No.
It has even more flexibility than No. 375434 (filed on May 6, 1982). Thus, the voice obtained from the synthetic voice source is modified to sound similar to other human voices. The vocal characteristics of human speech that convey impressions of a person's age, size, temperament, and even gender are
Thus, using the techniques disclosed herein, voices can be modified to produce voices with unnatural sound quality (eg, monotonous pitch).
For example, modification of the pitch parameter can be done by the method previously described in the aforementioned publication by Adal Haaner, "Speech Analysis and Synthesis by Linear Prediction of Speech Waves," which involves multiplying the pitch factor by a constant value. Although the present invention has been described in terms of changing the speech characteristics of synthesized speech by modifying the K speech parameters or reflection coefficients that define the vocal tract model, it is possible to Digital audio data can be used as digital audio data to define a vocal tract model. These data are modified by performing a transformation of the synthesized audio spectrum in the manner disclosed in simulating adjustment of the sample period. Thus, although the present invention has been specifically described with respect to the preferred embodiment thereof, changes and modifications of the invention will be apparent to those skilled in the art based on the description provided in this application of the invention and the present invention is incorporated herein by reference in the accompanying patents. It will be understood that we are limited only by the scope of the claims.
It is therefore intended that the appended claims are intended to cover any modifications or embodiments that fall within the original spirit of the invention.

[Brief explanation of the drawing]

Figures 1a to 1d are synthesized speech spectra created from digital audio data obtained from a single synthesized speech source identical to the spectra shown in Figure 1c, following a method that simulates sample period adjustment. 1a, 1b and 1d show the modified synthesized speech spectra; FIG. FIG. 2 is a flowchart schematically illustrating a method for modifying the speech characteristics of synthesized speech obtained from a single given synthesized speech source in accordance with the present invention. Third
The figure is a diagram illustrating in more detail the series of flowcharts in Figure 2, in which the sample period of digital audio data obtained from a synthesized voice source is adjusted by compressing or expanding the synthesized voice spectrum. FIG. 3 is a diagram showing a simulated process. Figures 4a to 4c are schematic circuit diagrams of an apparatus for modifying the audio characteristics of synthesized speech obtained from a single provided synthesized speech source in accordance with the present invention. FIG. 5 shows that the apparatus of FIGS. 4a to 4c can be combined to provide a plurality of different audio tones with very unique audio characteristics from a memory containing digital audio data of a single synthesized audio source. This figure shows a functional block diagram of a possible speech synthesis system.

Claims (1)

[Scope of Claims] 1. A method for changing the speech characteristics of synthesized speech to obtain a modified synthesized speech of any one of a plurality of speech sounds from a single given source of synthesized speech, comprising: A source of synthesized speech in the form of digital speech corresponding to each sample of an analog speech signal obtained at a time interval defined by a predetermined sample period, from which the synthesized speech can be retrieved, and from which the synthesized speech can be extracted. The audio consists of audio parameter frames provided at a predetermined audio rate;
Each of the speech parameter frames has a predetermined pitch period and a predetermined vocal tract model defined by a plurality of prediction coefficients, the plurality of prediction coefficients defining a predetermined vocal tract model for each frame of digital speech data. A predetermined number of blank values are added to , the digital audio data is changed from the first phase in the time domain to the second phase in the frequency domain by the operation of the first Fourier transform, and at that time, the predetermined number of blank values added are inverting the digital audio values of a plurality of prediction coefficients that are absorbed into the sequence of digital audio data signals and defining a synthesized audio spectrum and defining a predetermined vocal tract model for each frame of digital audio data in the frequency domain; and set the first reference factor P as the first integer,
This first integer is an integer equal to a selected number of predetermined points over the audio spectrum determined by the type of audio to be produced in the operation of the Fourier transform; A second integer whose size is unequal to the integer of simulating an adjustment in the sample period for digital audio data from a source of synthesized speech based on the difference between a second reference factor P and Q, the second reference factor Q provided by the second integer being P× is the closest even number to the product of F _NEW /F _OLD , where F _NEW is the desired apparent sample frequency of the simulated sample period and F _OLD is the sample frequency inherent in the given sample period. and modifying the predetermined vocal tract model of the digital audio data in response to a simulated adjustment in the sample period, the magnitude of the first integer providing the first reference factor P being the second reference factor. compresses the synthesized speech spectrum if the first integer providing the first reference factor P is greater than the second integer providing the second reference factor Q;
expand the synthesized speech spectrum and generate modified digital speech data as a digitized speech waveform that provides an impulse response value with the predetermined pitch period and amplitude data erased; For this purpose, the compressed or expanded synthesized speech spectrum is restored from the second phase in the frequency domain to the first phase in the time domain by a second Fourier transform operation, and the modified vocal tract model is used as a plurality of prediction coefficients. In providing the modified digital audio data, the digitized audio waveform is analyzed, and a plurality of prediction coefficients defining the modified vocal tract model are converted into reflection coefficients, which are represented by the reflection coefficients. generating an audio signal representative of human speech obtained from the modified digital speech data; converting the audio signal into an audible synthesized speech having altered speech characteristics from the synthesized speech obtained from the source of the synthesized speech; Method. 2. In the method according to claim 1, only the vocal tract model of the digital audio data is changed by a simulated adjustment in the sampling period of the digital audio data, and the predetermined pitch period of the source of the synthesized speech is changed. and a method in which the predetermined audio rate does not change. 3. With the method set forth in claim 2, the synthesized speech spectrum is compressed such that the magnitude of the first reference factor P is larger than the magnitude of the second reference factor Q. , a simulated adjustment in the sample period of the digital audio data from the source of the synthesized audio is applied to the first one from the sequence of spectral signals representing the digital audio data.
and a second reference factor P and Q by erasing a plurality of samples corresponding to the difference in magnitude between them, followed by generating modified digital audio data with altered audio characteristics. how to. 4. A method as claimed in claim 3, in which a plurality of samples are removed from the center of a spectral signal sequence in performing a simulated adjustment in the sample period of digital audio data from a source of synthetic audio. 5. A method as claimed in claim 3, in which a plurality of samples are removed from the end of a spectral signal sequence in performing a simulated adjustment in the sample period of digital audio data from a source of synthetic audio. 6. By the method described in claim 2, the synthesized speech spectrum is expanded such that the magnitude of the first reference factor P is smaller than the magnitude of the second reference factor Q. and a simulated adjustment in the sample period of the digital audio data from the source of the synthesized audio is applied to the sequence of spectral signals representing the digital audio data.
This is done by adding a plurality of blank values corresponding to the difference in magnitude between the reference factor Q and the first reference factor P, followed by a modified digital voice with modified audio characteristics. How to generate data. 7. In the method of claim 6, a plurality of blank values are added to the center of the spectral signal sequence in performing a simulated adjustment in the sample period of digital audio data from a source of synthesized speech. How to do it. 8. In the method of claim 6, a plurality of blank values are added to the ends of the spectral signal sequence in performing a simulated adjustment in the sample period of digital audio data from a source of synthesized speech. How to do it. 9. In the method according to claim 1, the first reference factor P is a number equal to the number of predetermined points determined by the type of speech to be produced in the inverse discrete Fourier transform, and the second reference factor Q is an even number of points in an inverse discrete Fourier transform, and the second Fourier transform operation is an inverse discrete Fourier transform. 10. In the method of claim 1, a sum of P-(N+1) blank values is added to the plurality of prediction coefficients prior to the first Fourier transform operation, where N is a predetermined voice. A method that is the number of predictive coefficients that define a road model. 11 A method for modifying the speech characteristics of synthesized speech to obtain a modified synthesized speech of any one of a plurality of speech sounds from a single given source of synthesized speech, defined at a predetermined sample period. a source of synthesized speech in digital audio format corresponding to each sample of an analog audio signal obtained at a time interval, from which synthesized speech can be extracted, said digital audio being provided at a predetermined speech rate; the voice parameter frames each having a predetermined pitch period and a predetermined vocal tract model defined by a plurality of prediction coefficients; A predetermined number of blank values are added to the plurality of prediction coefficients that define the road model, the digital audio data is changed from the first phase in the time domain to the second phase in the frequency domain by the operation of the first Fourier transform, and then the addition of blank values is performed. a predetermined number of blank values defined in the digital audio data signal are absorbed into the sequence of digital audio data signals and define a synthesized audio spectrum; inverting the digital audio value of the prediction coefficient and setting the first reference factor P as the first integer;
This first integer is an even number equal to the number of predetermined points over the speech spectrum determined by the modified synthesized speech to be produced in the operation of the inverse fast Fourier transform, and the first integer provides the first reference factor P. A second integer whose size is unequal to the integer 1 is set as the second reference factor Q, and the second integer has a power of 2 at an even point in the inverse fast Fourier transform, and the voice spectrum simulating an adjustment in the sample period for digital audio data from a source of synthesized speech based on the difference between said first and second reference factors P and Q; The first reference factor P, provided by an integer of 1, is the integer closest to the product of Q×F _OLD /F _NEW , where F _OLD is the sample frequency containing the given sample period and F _NEW is , is the desired apparent sample frequency of the simulated adjusted sample period, and modifies the predetermined vocal tract model of the digital audio data in response to the simulated adjustment in the sample period, for which the first reference factor P is compressing the synthesized speech spectrum if the magnitude of the first integer provided is greater than the second integer provided by the second reference factor Q; a second integer providing a second reference factor Q
expand the synthesized speech spectrum and generate modified digital speech data as a digitized speech waveform that provides an impulse response value with the predetermined pitch period and amplitude data erased; For this purpose, the compressed or expanded synthesized speech spectrum is returned from the second phase in the frequency domain to the first phase in the time domain by a second Fourier transform operation using inverse fast Fourier transform, and the modified vocal tract model is created. In providing modified digital voice data having a plurality of prediction coefficients, the digitized voice waveform is analyzed, and the plurality of prediction coefficients defining the modified vocal tract model are converted into reflection coefficients. , generating an audio signal representative of human speech obtained from modified digital speech data represented by a reflection coefficient, and converting said audio signal to said synthesized speech having altered speech characteristics from said synthesized speech obtained from said source of synthesized speech. How to convert to audible synthesized speech. 12. The method of claim 11, wherein only the vocal tract model of the digital audio data is changed by a simulated adjustment in the sampling period of the digital audio data, and the predetermined pitch period and A method that does not change the predetermined audio rate. 13. With the method described in claim 12, the synthesized speech spectrum is such that the first reference factor P is set larger than the second reference factor Q. A simulated adjustment in the sample period of the digital audio data from the compressed and synthesized audio source determines the difference in magnitude between the first and second reference factors P and Q from the sequence of spectral signals representing the digital audio data. and, following this, generate modified digital audio data with altered audio characteristics. 14. In the method of claim 13, a plurality of samples are removed from the center of the spectral signal sequence in performing a simulated adjustment in the sample period of digital audio data from a source of synthesized speech. Method. 15. In the method of claim 13, in performing a simulated adjustment in the sample period of digital audio data from a source of synthesized speech, a plurality of samples are removed from the end of the spectral signal sequence. How to do it. 16. With the method recited in claim 12, the synthesized speech spectrum is expanded such that the magnitude of the first reference factor P is smaller than the magnitude of the second reference factor Q. , a simulated adjustment in the sample period of digital audio data from a source of synthesized audio adds a second to the sequence of spectral signals representing the digital audio data.
This is done by adding a plurality of blank values corresponding to the difference in magnitude between the reference factor Q and the first reference factor P, followed by a modified digital voice with modified audio characteristics. How to generate data. 17. In the method of claim 16, in performing a simulated adjustment in the sample period of digital audio data from a source of synthesized speech, a plurality of blank values are located in the center of the spectral signal sequence. How to add. 18. In the method of claim 16, in performing a simulated adjustment in the sample period of digital audio data from a source of synthesized speech, a plurality of blank values are provided at the ends of the spectral signal sequence. How to add. 19. The method of claim 11, wherein a sum of P-(N+1) blank values is added to the plurality of prediction coefficients prior to the first Fourier transform, where N is a specific vocal tract model. is the number of predictive coefficients that define the method.