DE3938312C2 - Process for generating source sound data - Google Patents

Abstract

In the generating process, an analog or digital input signal is fed to a comb filter, passing only the basic frequency and its harmonic oscillations for generating analog or digital output signals. The reflating signal sections are separated from the output signal and recorded. To determine an interval of uniform analog signal portions, the analog signal is converted into a digital one by a Fourier transformation for generating second Fourier components. The phases of the latter are mutually matched, and another Fourier transformation is carried out. The periods of the peaks in the output data values are then determined.

Description

The invention relates to a method for generating source sound data according to the preamble of claim 1. In such a method it is required to generate various data from an input signal. Also data must be compressed and recorded. With musical sound data interval data must be determined. Compression takes place e.g. B. in blocks. Compressing and also winning repetition periods in one Signal occurs e.g. B. by a digital signal processor.

Sound sources as used in electronic musical instruments are z. B. analog sound sources z. B. consist of VCO, VCA and VCF, and digital sound sources that have a programmable sound generator or a ROM from which sound data are read out. As a digital sound source  also used one made from the tones of Musikin instruments source sound data determined by sampling and this Stores data in a memory.

There is usually a memory for storing source sound data large capacity is required, extensive loading exist efforts to design procedures so that storage space is saved. This particularly includes loop formation d. H. a procedure that determines the periodicity of musical tone signals exploits and bit compression z. B. by nonlinear Quantization.

The looping also serves to generate tones that are longer than the original sampled tone. A music ton typically consists of a non-tone component, such as they z. B. when striking a piano key or when blowing a wind instrument and a periodic Sound component. The non-tone component represents a formant part with not fixed periodicity. In the Tonkom component, the signal has a specific signal interval. By repeating n (integer) of the interval as Loop domain can use an extended tone wrestle of storage space are generated.

There is a problem with loop formation, that is as a loop fenrauschen is known. Loop noise occurs due to a special distribution of the frequencies. The rough is also noticeable when its level is lower than that of white noise. Several factors are likely responsible for the loop noise.

One of the factors is that the loop period does not exactly match the signal period of a source tone signal. If e.g. B. a source sound of 401 kHz is subjected to a loop formation with a period of 400 Hz, the signal obtained by loop formation only has frequencies that correspond to an integer multiple of the Schleifenpe period. The fundamental frequency of the source tone is shifted to 400 Hz, with harmonic harmonics with frequencies of 800 Hz, 1600 Hz,. . . It can be shown that if there is a 1% offset between the source tone frequency and the loop frequency, an nth order harmonic is formed as follows during loop formation, which is heard as loop noise:

C _n = (sin (n - 0.01)) / (π (n - 0.01)) (a)

Another factor is that non-even harmonics are generated, that is, kth order harmonics where k is non-even. The source signal, which appears periodically at the first glance, is strictly not periodic with several harmonics that are of an odd order. When creating loops, these harmonic harmonics are forcibly shifted to neighboring even harmonics. This distortion generated during loop formation can be heard as loop noise. If the harmonic loop harmonics have a frequency component that is a times the loop frequency, where a is not necessarily an integer, the distortion factor as formed in loop formation is a function of a, which is given by the following equation given is:

where m is the integer closest to a lies. The distortion factor becomes for a = 0.5; 1.5; 2.5; . . . largest and for 1.0; 2.0; 3.0. . . the smallest.  

It is believed that the two factors mentioned are: are the main ones for loop noise. In any Case the loop noise is generated when the loops period is an integer multiple of the source tone period.

The frequency components of the loop noise have one spectral distribution, which is undesirable in reproduction is so that it should be removed as much as possible.

On the other hand, the scanned and stored sources sound data those that were digitized directly so that the sound quality depends on the quality of the sampling. If the sound has many noise components during sampling, ent the read out and reproduced signal also holds it Noise components. If a musical tone with so-called vibrato is sampled, this tone is slightly frequency modulated. At this leads to the formation of loops in frequency modulation sideband produced odd harmonics what is reproduced as loop noise.

To start and end the loop formation are usual Usually two points of the same level, especially the Pe gel 0 used.

This selection of loop points is difficult and time consuming manoeuvrable, since it is done according to a trial and error procedure Points with approximately the same level are called the start and end points set.

For loop formation it is necessary to set the basic fre to record the source sound data. Conventionally, this is how gone that the musical sound data a low pass filter (LPF) to filter out high frequency noise components tern. It then becomes the number of zero crossings through the Low pass filter determined to determine the interval. At  This method requires that the musical tone for is maintained for a long time since the interval of the Fundamental tone can only be measured if a larger number is counted from zero crossings. The procedure can be so do not apply to fast fading tones.

As another method of determining the interval, mention should be made of one that consists in performing fast Fourier transform to detect the peak levels of the musical sound data. However, if the fundamental frequency is lower than the sampling frequency f _S , it is not possible to determine the fundamental frequency with this method, which leads to low fidelity. In addition, there are tones in which the fundamental tone is considerably weaker than the harmonics, which also makes it difficult to reliably determine the peaks of the fundamental tone.

In addition to creating loops, memory is saved capacity also the bit compression of the Source sound data used. In practice, this is done for. B. there through that from several given filters is chosen, which on a block basis to the highest comm pression leads, each block consisting of several samples consists.

For such a filter selection, each block with 16 samples head or parameter data posed such. B. area or filter data. The filters data is used to select a filter that is the highest Compression leads or where the compression is chosen is that it is optimal for coding. There are e.g. B. Filters used with three operating modes, namely directly with PCM, with first order and second order. The differential fil The first and second order act when decoding or Play back as an IIR filter so that when decoding or playing back  the first scan of a block, one or two scans that precede the block as initial values are required.

If first or second order differential filter for the first block of Source sound data are selected, but a previous scan is missing, which results in one or two dates as initial values in a memory must be filed. The hardware for the decoder is therefore extensive richer, what not for circuit integration and cost reduction is desired.

From US 4,044,204 a method for separating spoken and unspoken parts known from a speech signal. With this procedure becomes an input signal containing speech and noise via a Delay element fed to various filters, which a low-pass filter is connected downstream. Repeating intervals are with a special Circuit determined in which a peak detector peak values one of signals delivered to the amplifier. This amplifier is a Sound input signal supplied via parallel channels with appropriate filters.

The invention has for its object to provide a method based on simple way to generate source sound data with good sound quality.

This task is carried out in a method of the type mentioned at the beginning according to the invention in the characterizing part of claim 1 contained feature solved. Advantageous developments of the invention result from patent claims 2 and 3.

According to the solution of claim 1, a comb filter is used, the only Signals of a fundamental frequency and of harmonic harmonics passes through, thereby improving the signal / noise ratio, and there are in particular repetitive signal areas by means of loops start points and loop end points depend on an integer lot times the interval of the fundamental frequency of the input signal.

Claim 2 specifies a method such as the length of a regular sweeping pattern can be determined.

Claim 3 describes a method that is used when the length of a regularly recurring pattern, be it according to the The method of claim 2 or another method. Then it will namely a range of the determined length back and forth in the overall signal pushed until it is determined that neighboring areas are similar. The similarity is only by comparing the beginning or end areas determined.

The invention is illustrated below with the aid of figures illustrated embodiment explained. Show it:

Fig. 1 is a block diagram of an apparatus for generating Quelltondaten;

Fig. 2 is a musical tone signal;

Fig. 3 is a flowchart for explaining how an interval length is determined;

Fig. 4 is a block diagram for explaining how peak values are determined;

Fig. 5, together with a musical tone signal envelope;

Fig. 6 is a diagram for explaining the decay curve for a musical tone signal;

Fig. 7 is a block diagram for explaining an envelope formation;

Fig. 8 is a characteristic of an FIR filter;

Fig. 9 shows a signal after envelope correction of the musical tone signal;

Fig. 10 is a diagram illustrating the characteristics of a comb filter;

FIG. 11 is a flowchart for explaining a recording method using a comb filter;

Fig. 12 is a diagram for explaining how the start and end of a loop area are determined;

Fig. 13 is a flow chart for explaining how signals are shaped using the optimal loop position;

FIG. 14A and B, a musical tone correction before and after time base;

Compression 15A and B are diagrams for explaining of blocks by bit.

Fig. 16 shows a signal as obtained by multiple loop signals being added together;

FIG. 17 is a formant region by envelope correction;

FIG. 18 is a diagram for explaining the signal processing before and after the selection of a loop area;

Fig. 19 is a block diagram for explaining the function of bit compression and coding;

FIG. 20 is a diagram illustrating the contents of a data block, as is obtained by bit compression and consolidation Ko;

Fig. 21 is a diagram illustrating the content of the headers of a music signal; and

Fig. 22 is a block diagram of a system with audio processing unit.

Different functions are explained with reference to Fig. 1, which occur between the scanning of a music source signal and storing in a memory. The music source data signal is entered at a terminal 10 . It can be a signal that is picked up directly by a microphone or a signal that is supplied in analog or digital form from a recording medium for digital audio signals.

The source data formed by the device according to FIG. 1 has undergone a so-called loop formation, which is now explained on the basis of the musical tone signal according to FIG . Immediately after the start of the generation of a tone, non-tone components are contained in the tone, as they are e.g. B. from hitting a key or from blowing on a wind instrument. Therefore, a so-called formant range FR with inaccurate signal periodicity is first generated. There then follows a repeated repetition of the same signal form which corresponds to the music interval (distance or pitch) of the musical tone.

A number n of periods of this repetitive waveform is used as the loop domain LP. It lies between a loop start point LP _S and a loop end point LP _E. The formant area FR and the loop domain LP are recorded in a memory, and for playback, the formant area is first played, and then the loop domain LP is played repeatedly for a desired period of time.

The input musical tone signal is sampled in a sampling block 11 , e.g. B. at 38 kHz. A 16-bit signal is obtained before sampling. The sampling corresponds to the A / D conversion of analog input signals.

In a subsequent interval detector 12 , the fundamental frequency, ie the frequency f _{0 of} the fundamental, is determined.

The principle of the interval detector 12 will now be explained. The musical tone signal sometimes has a fundamental frequency that is significantly lower than the sampling frequency f _S , so that it is difficult to determine the interval by only sampling the peaks in the musical tone along the frequency axis. It is therefore necessary to use the spectrum of harmonic overtones.

The musical tone signal f (t), the interval of which is to be determined, can be expressed by Fourier expansion as follows:
where a (ω) and ϕ (ω) mean the amplitude or phase of the overtones. If the phase shift ϕ (ω) is set to zero for each overtone, the above formula can be rewritten as follows:

The peaks of the signal f (t) thus adjusted correspond entirely number multiples of the periods of all overtones of the signal at t = 0. The peaks only show the period of the fundamental.

Based on this principle, the sequence of steps for detecting the interval will now be described with reference to FIG. 3.

Musical sound data and a value "0" are fed to a real part input 31 and an imaginary part 33 of a transformation block 33 for fast Fourier transformation.

For the fast Fourier transform, if the interval is denoted by x (t):

a _n cos (2πf _n t + θ) (3),

x (t) can be given by

This can be written in complex notation as

where the following equation

cosθ = (exp (jθ) + exp (-jθ)) / 2 (6)

is used. The following equation is obtained by Fourier transformation:

in which δ (ω - ω _n ) describes a delta function.

In a following block 34 the absolute value, ie the root of the sum of the squares of the real and the imaginary part of the data after the fast Fourier transformation, is calculated.

By using the absolute value Y (ω) of X (ω), the phase component stands out, so that the following applies:

Y (ω) = [X (ω) X (ω)] ^1/2 = (1/2) a _n δ (ω - ω _n ) (9).

This becomes the phase adjustment for all high frequency compos elements in the musical sound data. The phase components can be adjusted to 0 by setting the imaginary parts.

The normalization value calculated in this way is supplied as a real data part to a further block 36 for fast Fourier transformation, which in this case is designed as a block for inverse fast Fourier transformation. While this data reaches an input for real data, an input for imaginary data is supplied with the value "0". The inverse fast Fourier transform can be expressed by

The ge after the inverse fast Fourier transformation Won music data are output as a signal that the Synthesis of the cosine waves corresponds to the proportions higher frequency are adjusted in phase.

The peak values of the source sound data thus produced are determined in a peak detector 37 . The peak points are those points at which the peaks coincide for all frequency components. In a subsequent block 38, the peak values are sorted according to descending values. The interval follows by measuring the periods of the peaks determined.

FIG. 4 illustrates an arrangement of the peak detector block 37 from FIG. 3 for determining the peaks, ie the maximum values of the musical tone data.

It should be noted that within musical sound data a large number of peaks with different values  is available. An interval can be found that the distance between peaks is determined.

FIG. 4 is a sequence of musical tone data that result from the inverse Fourier transform, connecting through an input 41 a (N + 1) -stage shift register 42 is supplied and there is a register _{-N, 2.} . ., a _0,. . ., a _{N / 2} supplied to an output terminal 43 . The shift register 42 acts as a window of wide (N + 1) samples with respect to the sequence of musical sound data. The sample values are fed to a maximum value detector 44 via this window. Since the musical sound data first reaches the register a _{-N / 2} and is then transported to the register a _{N / 2} , (N + 1) scanning musical data is transferred from the registers to the maximum value detector 44 .

The maximum value detector 44 is designed so that when z. B. the value from the middle register a ₀ of all values became maximum, gives this value as maximum value to the output terminal 45 . The width (N + 1) of the window can be set as desired.

The envelope of the sampled digital musical tone signal is obtained in an envelope detector 13 using the interval data. The envelope signal, designated B in Fig. 5, is obtained by connecting the peaks in the musical tone signal A in sequence. It indicates a change in volume since the time the sound was generated. The envelope is described by parameter ADSR (Attack Time / Decay Time / Sustain Level / Release Time). In the case of a piano tone, part A (Attack Time T _A ) describes the time that elapses between pressing a key and the time until the desired volume is reached. The second part D (Decay Time T _D ) is the time that elapses between the time the desired volume is reached and the time a volume is reached that is maintained. The third portion S (Sustain Level L _S ) indicates the volume held, while the last portion R (Release Time T _R ) indicates the time that elapses between releasing a key and decaying a tone. The times T _A , T _D and T _R can also be understood as changes in the volume. Shell parameters other than the above four parameters can also be used.

It should be noted that the overall decay rate of the signal is obtained in the envelope detector 13 , at the same time as envelope data, as it is e.g. B. are given by the above-mentioned parameters ADSR. This fall time data takes a reference value "1" from the time of sound generation during the attack time T _A and then falls smoothly as illustrated by FIG. 6.

The function of the envelope detector 13 according to FIG. 1 will now be illustrated with reference to the flow diagram of FIG. 7.

The principle used is similar to that used for the Er average of the envelope of an amplitude modulated (AM) sig nals is used. The envelope is determined by it is assumed that the determined interval of the carrier fre corresponds to an AM signal. The envelope data who used when playing musical tones, the reason position of the envelope data and the interval data the.

The musical tone data are fed via an input connection 51 to an absolute value output block 52 , which determines the absolute value of the signal height data of the musical tones. This absolute value data is fed to a digital FIR (Finite Impulse Response) filter block 55 . This FIR block 55 serves as a low-pass filter, the clipping characteristic of which is determined by the fact that filter coefficients are supplied to it, which were first generated in an LPF coefficient generator 54 on the basis of interval data which are supplied via an input terminal 53 .

The filter characteristic according to the example of FIG. 8 has zero points at multiples of the fundamental frequency (f ₀ ). The envelope data illustrated with B in FIG. 5 can be obtained from the musical tone data signal A in FIG. 5 by weakening the frequencies of the fundamental and the overtones by the FIR filter. The formula applies to the filter characteristics

H (f) = k. (sin (πf / f ₀ )) / f (11)

in which f _{0 represents} the fundamental frequency, ie the interval of the musical tone signal.

Referring to Fig. 1 will now be explained how the pitch data for the formant portion FR and the loop LP domain from the tone height data of the sampled musical tone signal to be generated.

In a first block 14 for generating the loop data, the sampled musical tone data is divided, which was obtained with the aid of the previously determined envelope curve B according to FIG. 5 (or it is multiplied by the reciprocal value of the data).

This serves to generate an envelope correction which ensures a constant amplitude, as is shown in FIG. 9. This corrected signal, more precisely the corresponding signal height data, is filtered in order to also generate pitch data in which other components besides the sound components are weakened, or in other words in which the sound components are emphasized. Sound components are frequency components with integer multiples of the basic frequency f ₀ . The data is passed through a high pass filter (HPF) to low-frequency components such. B. remove vibrato frequencies that are present in the signal corrected with the envelope. This is followed by a comb filter with a frequency characteristic as indicated by dash-dotted lines in FIG. 10, ie with a frequency characteristic with frequency bands that are integer multiples of the fundamental frequency f ₀ . As a result, only the signals present in the HPF signal are passed through where they are weakened by non-tone components or noise components. If necessary, the data is also sent through a low pass filter (LPF) to remove noise components that are superimposed on the output signal from the comb filter.

Since musical sound data generally have a constant interval, ie a constant fundamental frequency, energy concentration occurs in the vicinity of the fundamental frequency f ₀ and integral multiples thereof. In contrast, noise components have a uniform frequency distribution. When the inputted musical sound data is processed with the aid of a comb filter with the frequency characteristic, as shown in broken lines in FIG. 10, only frequency components of the fundamental frequency or integer multiples thereof pass through the filter essentially unattenuated, while other components are weakened, which improves the signal-to-noise ratio. The frequency characteristic of a comb filter shown in FIG. 10 can be represented by the formula

H (f) = [(cos (2πf / f ₀ ) + 1) / 2] ^N (12)

are illustrated, in which f _{0 is} the fundamental frequency of the input signal, ie the frequency of the fundamental tone with a certain interval. N is the number of stages of the comb filter.

The musical tone signal, the noise components of which have been reduced in the manner mentioned, is fed to a circuit which determines repeated waveforms. Such waveforms, such as. B. the loop domain LP of FIG. 2 is separated and the remaining part is recorded, for. B. in a semiconductor memory. The stored musical sound data signal contains the non-sound component and part of the noise. Since it is ensured that the loop noise is reduced when playing the repeat th signal part.

The frequency characteristics of the HPF, the comb filter and the LPF are determined on the basis of the fundamental frequency f ₀ , ie on the basis of the interval data as obtained from block 12 .

A method of signal recording using said filtering will now be illustrated with reference to FIG. 11. In a step S1, the fundamental frequency f _{0 of} the given analog signal or the associated digital input signal is determined. In a step S2, the input analog signal is filtered by a comb filter of the type described above in order to generate an analog or digital output signal. In a step S3 it is ensured that only data from the fundamental frequency band and frequency bands of harmonic multiples are passed through. The output signal is stored in a step S4.

In the recording method described, that is Musical tone signal sent through a comb filter that only the Transmits fundamental and harmonic harmonics, attenuates non-tone components and noise components, to improve the signal-to-noise ratio. With loops education become musical sound data with weakened noise composers repeated to suppress the noise.

In a block 16 for recognizing the loop domain, the musical tone signal is examined for repeated waveforms in order to determine from tone components that differ from the aforementioned tone components that were emphasized by the filter method described. With these, the loop start point LP _S and the loop end point LP _{E are} formed.

In the determination block 16 , loop points are selected which are separated from one another by an integer multiple of the interval of the musical tone signal. The selection principle is explained below. As just mentioned, the loop point spacing must be an integer multiple of the interval as given by the fundamental frequency of the musical tone signal. By exactly determining the interval, the loop point distance can be determined.

Since the loop spacing is predetermined after the interval has been determined, two points are selected which are separated from one another by the named spacing. The signals in the vicinity of these points are then examined and a correlation is formed. An evaluation function will now be explained which uses the convolution or sum formation of products with respect to sampling points in the area of the two points mentioned. The convolution is carried out several times for sets of all points in order to determine the correlation of the signals. The musical sound data are successively fed to a unit for adding up products, e.g. B. a digital signal processing device, which is described below. The convolution is calculated and issued there. The set of two points for which the convolution becomes maximum is regarded as the set of the correct loop start point LP _S and the correct loop end point LP _E.

In Fig. 12, a _{0 is} a candidate point for the loop starting point LP _S. b ₀ is accordingly a candidate for the loop end point LP _E. Further amplitude data a _-N,. . ., a _-2 , a _-1 , a ₀ , a ₁ , a _2,. . ., a _N in several points, here (2N + 1) points before and after the candidate point a ₀ are also shown. Correspondingly, amplitude points b _-N,. . ., b _-2 , b _-1 , b ₀ , b ₁ , b ₂ ,. . ., b _N before and after the candidate point b ₀ . The estimation function E (a ₀ , b ₀ ) then becomes:

The convolution around points a ₀ and b ₀ follows from equation (13). The sets of candidates a ₀ and b ₀ are gradually changed until all loop point candidates are examined. Those for whom the evaluation function E becomes maximum, who are regarded as the loop points.

In addition to the convolution method, e.g. B. the method of mean squares can also be used to find loop points. The following equation then applies:

In this case, the correct points are available if the Evaluation function becomes minimal.

The selection method for optimal loop points described above can always be used when analog signals are digitized with repeated periods in order to form loop data. The method is further illustrated below using the flow chart of FIG. 13.

Referring to FIG. 13, an analog signal with repeated Sig is nalform converted in step S11 into a digital signal containing more reren sampling points. In a step S12, sampling points are selected which are separated from one another by the repetition period. In a step S13, values for a predefined evaluation function are calculated on the basis of sampling points which are located in the vicinity of the two selected points. The selected points are then shifted in a step S14 with the distance maintained until a predetermined number of shifts has been reached. If the latter is the case, the set that shows the greatest correlation is selected from the calculated values. Then, in a step S16, those sampling data between the selected points are extracted as sampling data for the repeated signal.

In the loop point determination block 16 according to FIG. 1, the loop start point LP _S and the loop end point LP _{E are determined} on the basis of the interval. From the spacing of the points, which can comprise several intervals of the fundamental tone, there follows an interval conversion ratio, which is used in a time base correction block 17 .

The time base correction is carried out in order to adapt the intervals of different source sound data to one another when this data is stored in a memory. Instead of the interval change ratio, the interval data as supplied by the interval detector 12 can be used for this purpose.

The interval normalization process carried out in the time base correction block will now be explained with reference to FIG. 14.

Fig. 14A shows the musical tone signal before companding the time base, while Fig. 14B shows the corresponding signal thereafter. The time axes are divided into blocks for quasi-direct bit compression and coding, as explained below.

With signal A before the time base correction, the loop domain LP is generally not in direct relation to a block. In the signal of FIG. 14B, the loop LP domain is companded in the time base so as to (m-fold) is an integer multiple of a block length or block period. The loop domain is also shifted along the time axis, so that the loop start point LP _S and the loop end point LP _E match the boundaries of blocks. The time base correction, ie companding the time base and shifting in the manner mentioned, ensures that loops can be formed for an integer number (m) of blocks, thereby normalizing the interval of the source sound data during recording.

With an offset ΔT from the foremost block limit of the musical signal, a pitch signal value "0" can be inserted. This data value "0" serves as a pseudo data value, so that filters which require an initial value for the data compression can be selected. This is explained below using a blockwise compression illustrated by FIG. 21.

Fig. 15 shows the structure for one block for Signalampli tudenwerte after the time base correction, which data is then subjected to a bit compression and encoding as later explained. There is h amplitude data in one block (number of samples or words). In the present case, the interval normalization consists of basic companding in a time in which the number of words within n periods (the constant time period T _W according to FIG. 2), ie within the loop period LP, is an integer multiple (m times) of the number of the words h are in the block. The interval normalization is preferably that the loop start point LP _S and the loop end point LP _{E are} each shifted to block boundaries. If this is the case, it becomes possible to avoid errors when switching blocks, which is the case when decoding with the aid of a bit compression and coding system.

In FIG. 15A, words WLP _S and WLP _E in one block each are the samples at the loop start and end points LP _S and LP _E , more precisely the samples in each case directly before a loop end point LP _E in the corrected signal. Unless shifted, the start and end points mentioned do not necessarily coincide with the block boundaries, so that, as shown in Fig. 15B, the words WLP _S and WLP _{E are} at any position within the blocks. In any case, however, the number of words between the words WLP _S and WLP _{E is} m times the number of words h in a block, where m is an integer, which means that the interval is normalized.

Companding the time base of a music signal with the The normalization just mentioned can be done in different ways consequences. For example, it can be done that the amplitudes values of the sampled signals are interpolated, where a Parent scan filter is used.

If the loop period for a music signal is not an integer multiple of the sampling period, so that there is an offset in the amplitude values at the loop start point LP _S and the loop end point LP _E , the amplitude corresponding to the amplitude value at the start point can be found adjacent to the loop end point, and in that z. B. parent scanning is used to implement the loops period.

Such a loop period, which is not an integer multiple of the sampling period, can be set to be an integer multiple of the block period, which is done by the time base correction method described above. If time base companding by e.g. B. 256-fold higher-level sampling, the amplitude error between the values for the start point LP _S and the end point LP _E can be reduced to 1/256.

After the loop domain LP is determined and corrected in time basis, loop domains LP are appended to one another, as illustrated by FIG. 16, to generate loop data. The signal according to FIG. 16 is obtained by the fact that a single loop domain LP according to FIG. 14B is picked out and this one domain is connected to one another several times. The concatenation of loop domains takes place in a loop signal generation block 21 ( FIG. 1).

Since the loop data is formed by repeating a loop domain LP several times, a start block with the word WLP _{S is} immediately followed by an end block with the word WLP _E , which corresponds to the loop end point WLP _E , more precisely the point directly before this point. So that coding for bit compression can take place, at least the end block must be stored directly before the start block of the loop domain LP. More generally speaking, at the time of bit compression and block-based coding, the parameters for the start block, ie the data for bit compression and coding for each block, e.g. B. range or filter selection data, which will be described below, can only be formed on the basis of data of the start and end blocks. This technique can also be used in the case where music data consists only of loop data without formants.

With this procedure, the same data is available for several sampling points before and after the loop start point LP _S and the loop end point LP _E. Therefore, the parameters for bit compression and coding in the blocks immediately before these points are the same, which can reduce errors and noise when reproducing the loops after decoding. The music data during playback are then stable and free of transition noise. In the embodiment, there are about 500 sample data in a loop area LP before the start block.

For the signal generation for the formant region FR, an envelope correction is carried out in a block 18 ( FIG. 1), corresponding to the correction for loop data in block 14 . The envelope correction is carried out by dividing the sampled musical sound data by envelope data ( Fig. 6) consisting only of the data for falling signals, thereby obtaining amplitude values for a signal as shown in Fig. 17. When Ausgangssig nal according to Fig. 17 makes accordingly just the envelope of currency end of the time period T _A noticeable, while in the other rich Be constant amplitude is present.

If necessary, the envelope-corrected signal is filtered in a block 19 . This is done with a Kammfil ter with a frequency characteristic z. B. corresponds to that as shown by the dash-dotted line in Fig. 10. The comb filter data are formed on the basis of the interval data for the fundamental frequency, which data are supplied by the interval detector 12 . The data are used to generate signal data from the formant area in the source sound data, which are finally stored in a memory.

In a subsequent time base correction block 20 , a time base correction is carried out, as described above with reference to block 17 . The time base correction is used to normalize the intervals for the sound data by companding the time base based on the interval change ratio found in block 16 or based on the interval data provided by block 12 .

In a mixer 22 , the formant data and the loop data, which have been corrected with the same interval data or interval change ratios, are mixed. For this purpose, a Hamming window is applied to the formant area formation signal from block 20 . A fade-out type signal to be mixed with the loop data is generated which decays. A similar Hamming window is applied to the loop data from block 21 . A fade-out type signal to be mixed with the formant data is generated which grows. The two signals are mixed (or blended) to produce a musical tone signal that represents the source sound data. As loop data to be stored, data can be picked out which comes from a loop domain which is somewhat away from the cross-faded area, thereby reducing the loop noise when loop signals are reproduced. In this way, pitch data of a source tone signal is generated that has a formant area (non-tone components) and loop data (repeated, regular data).

The starting point for the loop data signal can also be with the Loop start point for the formant-forming signal verbun be that.

To determine the loop domain and mix the forman area and the loop data is first ge manually mixes. You then listen to the result and it becomes ver based on the loop start and end points is working.

Before the more precise determination of the loop area takes place in block 16 , this determination and the mixing are carried out by hand, and a trial listening takes place, as explained in the following FIG. 18. Thereafter, the more detailed process described above is executed from step S26 in FIG. 18.

In the flowchart according to FIG. 18, the loop points are determined with a low accuracy in a step S21 by determining zero point crossings or by optically checking the signal. In a step S22, the signal between the loop points is reproduced several times. In a step S23, it is judged by listening whether the loop formation is carried out properly. If this is not the case, step S21 follows in order to determine loop points again. This process continues until a satisfactory result is obtained. If this is the case, a step S24 follows in which the signal z. B. is mixed by fading with the For mantensignal. In a step S25, it is checked whether the auditory impression is good at the transition from the formant to the loops. If this is not the case, step S24 is repeated to mix again. If the hearing pressure is good, a step S26 follows in which the loop points are determined according to block 16 with high accuracy. There is scanning z. B. with a resolution of 1/256 over the sampling period with z. B. 256-fold superordinate scanning. In a subsequent step S27, the interval conversion ratio for the interval normalization is calculated. In a step S28, the time base correction is carried out in accordance with blocks 17 or 20 . In a step S29, the loop formation according to block 21 ( FIG. 1) is carried out. The mixing according to block 22 takes place in a subsequent step S30. The process from step S26 is carried out using the loop points as obtained in steps S21-S25. Steps S21-S25 can be omitted for full automation of looping.

The amplitude values thus obtained, which consist of the formant region FR and the loop domain LP, are bit-compressed and encoded in a block 23 following the mixer 22 .

Various systems can be used for compression and encoding. One is suitable as proposed in JP-62-008629 and JP-62-003516. There, a predetermined number of h sample words with amplitude values are grouped in a block, and bit compression is carried out block by block. This highly effective system is briefly explained below with reference to FIG. 19.

In the block diagram according to FIG. 19, an encoder 70 is present on the recording side and a decoder 90 on the reproduction side. The amplitude values x (n) of the source sound data signal are fed to an input terminal 71 of the encoder 70 .

The amplitude values x (n) of the input signal are fed to an FIR digital filter 74 , which is formed by a prediction element 72 and an adder 73 . A respective amplitude value x (n) of the prediction signal from the prediction element 72 is supplied to the adder 73 as a subtraction signal. There it is subtracted from the input signal x (n) in order to generate a prediction error signal or a deviation signal d (n). The prediction element 72 calculates the prediction value x (n) from a primary combination of the last p input values x (n - p), x (n - p + 1),. . ., x (n - 1). The FIR filter 74 is referred to below as the coding filter.

With the highly effective system for bit compression and coding specified above, the source sound data falling within a predetermined period of time, ie a predetermined number h of words grouped into blocks, are selected for each block. The coding filter 74 has optimal characteristics for this. It can be formed that several, in the case of four, different characteristics are defined in advance and then the optimal characteristic is selected, ie the one with which the highest compression ratio can be achieved. For this purpose, a set of coefficients is stored in the predictor 72 in a plurality of, in the example four, sets of coefficient memories. A set of coefficients is selected in time multiplex.

The deviation signal d (n), which represents a predicted error, is fed via a summing point 81 to a bit compressor, which has a level shifter 75 (G) and a quantizer 76 . There the compression takes place in such a way that the index part and the mantissa correspond to the gain G or the output signal from the quantizer 76 in the case of floating-point command. A requantization is thus carried out, in which the input data are shifted by the level shifter 75 by a number of bits, which number corresponds to the gain G. This shifts the range and a predetermined number of bits of the bit shifted data is taken out by the quantizer 76 . A noise shaping circuit 77 is used to form the quantized error between the output signal and the input signal of the quantizer 76 at a summing point 78 and to give this signal via a G ^-1 level shifter 79 to a predictor 80 . The predicted signal for the quantization error is given as a subtraction signal to the summing point 81 in order to obtain a so-called error feedback. After this requantization by the quantizer 76 and the error feedback by the noise shaping circuit 77 , an output signal (s) is obtained at an output terminal 82 .

The output signal d '(n) from the summing point 81 is the difference d (n) from the summing point 73 , minus the prediction signal (n) for the quantization error, as formed by the noise shaping circuit 77 . The output signal d "(n) from the G level shifter 75 corresponds to the output signal d '(n) from the summing point 81 multiplied by the gain G. The output signal (n) from the quantizer 76 is the sum of the output signals d" (n) from the G- Level shifter 75 and the quantization error signal e (n), as is given by the summing point 78 . After passing through the G ^-1 level shifter 79 and the prediction element 80 , using the primary combination of the last r input signals, the quantization error e (n) is converted into the prediction signal (s) for the quantization error.

After encoding as described above, the source sound data is converted into the output signal (s) from the quantizer 76 and is output at an output terminal 82 .

Mode selection data for the optimal filter selection are output by an adaptive range prediction circuit 84 and e.g. B. given to the prediction element 72 of the coding filter 74 and an output terminal 87 . Area data for determining the bit shift number or the gains G and G ^-1 who are also given to the level shifters 75 and 79 and to an output terminal 86 .

The playback-side decoder 90 has an input terminal 91 , to which a signal '(n) is supplied, which is formed by transmitting or recording and reproducing the output signal (n) from the output terminal 82 from the encoder 70 ge. This input signal is fed to a summing point 93 via a G ^-1 level shifter 92 . The output signal x '(n) from the summing point 93 is fed to a prediction element 94 and thereby converted into a prediction signal (n), which is then passed to a summing point 93 , where the output signal "(n) from the slide 92 is added to it. The sum signal is given as a decoder output signal '(n) to an output terminal 95 .

The range data and the mode selection data output from the output terminals 86 and 87 of the encoder 70 are input to the input terminals 96 and 97 of the decoder 90 , respectively. The range data from the input terminal 96 is transmitted to the slider 92 to set its gain G ^-1 , while the mode selection data is input from the input terminal 97 to the predictor 94 to set the prediction characteristic to be that of the predictor 72 of the encoder 70 matches.

In the decoder 90 thus constructed, the output signal "(n) from the slide 92 is the product of the input signal '(n) and the gain G ^-1 . The output signal' (n) from the summing point 93 is the sum of the said output signal from Slider 92 and from the prediction signal '(n).

Fig. 20 shows a block of output data from the encoder bitkomprimie leaders 70th There are 1-byte header data (parameter data relating to compression or sub-data) RF and 8-byte scan data D _A0 -D _B3 . The header data RF is 4-bit area data, 2-bit mode selection data or filter selection data and two 1-bit flag data, such as. B. a data value LI, which indicates the presence or absence of a loop, and a data value EI, which indicates whether the last block of the signal is negative. Each sample of the amplitude data is reproduced by four bits after bit compression, while 16 samples of 4-bit data D _A0H -D _{B3L are contained} in the data D _A0 -D _B3 .

FIG. 21 shows a block of bit compressed and encoded amplitude data corresponding to the beginning of the musical tone signal shown in FIG. 2. Only the amplitude data without the head are shown. For the sake of clarity, each block consists of only eight sample data. B. 16 sample data are available. This applies to the case of FIG. 15.

The quasi-instantaneous bit compression described above and encoding selects either immediate PCM mode that directly outputs the entered musical tone signal, or one First order differential filter mode or such second order, each of which filtered the musical tone signal output, resulting in output signals with the highest compression relationship leads.

When music tones are sampled and read from a memory starts the input of the musical tone signal in one tone generation start point KS. If a differential filter mode first or second order that needs an initial value selected in the first block after the tone generation starting point KS it would be necessary to deduce an initialization value to save. However, it is desirable that this not be required is. For this reason, in the initial period after the tone generation start point KS pseudo input signals testifies that lead to the immediate PCM mode ge is chosen. The data processing then takes place in such a way that with these pseudo signals the input signals are processed.

This is done e.g. B. in that, as shown in Fig. 21, a block with loud values "0" is arranged as a pseudo input signal block in front of the tone generation start point KS. These data "0" are treated as amplitude values and bit-compressed. This can be done by storing a block with the values "0" in a memory or by initially placing values "0" before the starting point KS when scanning musical tones, ie a quiet area before the sound is generated. At least one block of pseudo input signals is required.

The musical sound data including the pseudo input signals thus formed are bit-compressed and encoded by the system illustrated in FIG. 19 and then recorded, e.g. B. in a memory. The compressed signal is played back.

When playing music sound data with pseudo input signal the immediate PCM mode is selected first, which makes it is not necessary, initial values for differential filters to determine first or second order in advance.

The delay in the start of the sound signals due to of the pseudo input signal appears to interfere with the Tone signal because there is silence at first. However, this bothers not actually, since the sampling frequency is 32 kHz, with 16 samples each in a block. This makes the ver delay about 0.5 msec, which is not noticeable in the auditory impression.

The bit compression and coding and other signal processing described above are preferably carried out with the aid of a digital signal processor (DSP). Fig. 22 illustrates an embodiment for an audio signal processing unit 107 as a sound source unit, processes the Quelltondaten. Peripherals are also shown.

A host computer 104 , e.g. B. in the form of a conventional PC of a digital electronic musical instrument or a Spielge device is connected to the audio signal processing unit 107 as a sound source unit, so that the source sound data from the host computer 104 are loaded into the processing unit 107 . The audio signal processing device (APU) 107 has a CPU 103 , such as a microprocessor, a digital signal processor DSP 101 and a memory 102 for storing the source sound data. So at least the data just mentioned are stored in memory 102 . The DSP 101 performs a variety of operations, including readout control for the source sound data, but also loop bit expansion or restoration, interval conversion, envelope value addition, or echo generation. The memory 102 also serves as a buffer memory for values that are generated in these processes. The CPU 103 controls the operations performed by the DSP 101 .

The digital musical sound data, as they are finally available from the end of the various processes carried out by the DSP 101 using the source sound data from the memory 102 , are converted by a D / A converter 105 and sent to a loudspeaker 106 .

In the embodiment, the source sound data is verified by Ver bind the formant area and the loop domain other formed. The method described can also be based on Source sound data can be applied only from loop domains consist. The decoder-side devices and the memories can also be arranged externally, also pluggable. The Ver Driving is not only applicable to the creation of musical tones cash, but also on language production.

Claims (3)

1. A method for generating source sound data, in which an analog or digital input signal is fed to a comb filter, which only allows the fundamental frequency and harmonics to pass through to produce analog or digital output signals, and in which repeating signal areas are separated from the output signal and are drawn on, characterized in that
the output signal is examined for the repeating signal areas,
Loop start point (LP _S ) and loop end points (LP _E ) are formed for these signal areas, and
such loop start points (LP _S ) and loop end points (LP _E ) are selected that are separated from each other by an integer multiple of the interval of the fundamental frequency of the input signal. 2. The method according to claim 1, characterized in that to determine a repetitive interval Signal areas in an analog input signal this after conversion into a digital signal of a Fourier Transformation is subjected to multiple Fourier Components that generate the phases of the components can be adapted to one another so that a Fourier Transformation takes place, and that the periods of the peaks can be determined in the output data values.   3. The method according to claim 1, characterized in that to determine the location of repetitive signal areas specify a range within a signal ner length is shifted over the signal that under is searched whether corresponding areas in the Neighborhood of the start and end points of the shifted NEN area are similar to each other and that on dir ge location is decided if similarity is found becomes.