EP1130577B1 - Method for the reconstruction of low speech frequencies from mid-range frequencies - Google Patents

Abstract

The method involves determining at least two adjacent frequency components in the speech signal with increased amplitude above a frequency threshold (w0). The fundamental frequency (wg) of the speech signal is determined as the difference between the two or more adjacent frequency components and the low frequency range is reconstructed below the threshold frequency using the determined fundamental frequency. Independent claims are also included for the following: (1) the use of the method in a moving vehicle (2) an arrangement for reconstruction of low frequency speech components.

Description

The invention relates to a method and an apparatus for the reconstruction of low-frequency speech components from medium-high frequency components.

In the prior art of the digital processing of speech signals with a high noise level in the low-frequency range, the signal is improved in that either noise components are filtered out or very strongly disturbed frequency range are completely filtered out of the signal.

US Pat. No. 5,842,160 A discloses a method for improving the quality of a digital voice transmission in which different data volumes are assigned to different frequency bands depending on the energy content. The nature of the coding and transmission results in low-energy signal areas, which lead to gaps in the received signal spectrum. These gaps are filled by signals synthesized from the existing data so that a more natural sounding speech signal is achieved.

From US 4,091,237 A a method for determining the basic voice frequency of a digital voice signal in real time is known. Especially for signals with a limited frequency range, such as telephone signals, and with a high level of noise, the speech signal is improved by filtering out noise. The signal is split by a plurality of band-pass filters and a corresponding histogram is formed, from which the basic voice frequency is extracted. If the fundamental frequency is known, noise can be detected by the fact that they are not in harmonic relation to the fundamental frequency. The method described above serves to determine the fundamental frequency characteristic of a voice.

Furthermore, from DE 37 33 983 a method for attenuating noise in a hearing aid is known in which the signal is digitized and divided into individual frequency ranges. Frequency ranges with certain characteristics, such as fast or very slow spectral distribution changes are attenuated and / or the cutoff frequencies are shifted. The thus cleaned signal is converted into synthetic speech signals.

Furthermore, from 4,700,390 A a method for the reconstruction of low-frequency audio components from an audio signal is known.

The method described above and the associated devices is based on the disadvantage that the speech signal is not reconstructed at all or only in an inadequate form in order to produce the most natural possible source speech signal.

The methods described above can be used inter alia in digital voice enhancement (DVE). For example, two microphones are mounted above each row of seats in a motor vehicle, so that it is, for example. All vehicle occupants is allowed to participate in a telephone conversation. The system transmits the voice recorded at the front of the microphone to the rear standard loudspeakers and vice versa. The system is thus fully connected to the handsfree telephone and the radio / CD / navigation device. It significantly improves the communication within the vehicle, especially when driving fast.

The level of the vehicle interior noise increases very strongly to low frequencies, so that the language is covered there by the noise. In order to transmit as little ambient noise as possible through the DVE system, since this would unnecessarily increase the noise level inside the interior, in a part of the above-described methods all frequencies are cut off below, for example, 200 to 500 Hz, depending on the speed. The result is that the speech fundamental frequency and the first multiples (harmonics) in the transmitted signal are missing. The language thus sounds like a telephone, as typically a telephone network allows a sound transmission only above 350 Hz.

In addition to the use of a handsfree telephone can be carried out with the procedures and the speech communication within the vehicle. However, optimal sound quality is required in order to gain acceptance among the buyers.

In particular, in the methods that rid the speech of noise, z. As spectral subtraction or coherence filters, it comes that the variance of the frequency component of noise comes in the order of magnitude of the power of the speech signal. Thus, an effective noise reduction is no longer possible and the applied methods are no longer effective.

The invention is therefore based on the technical problem of further developing the known from the prior art method and the associated apparatus for the reconstruction of low-frequency speech components of medium frequency components and to design that for a reproduction of the disturbed speech signal as natural as possible reproduction possible.

The above-indicated technical problem is solved by a method having the features of claim 1. First, at least two adjacently arranged frequency components with an increased amplitude in the voice signal are determined above a cutoff frequency. Thereafter, the fundamental frequency of the speech signal is determined as a frequency difference between the at least two adjacent frequency components. Finally, the low-frequency frequency range below the cut-off frequency is reconstructed with the aid of the determined fundamental frequency and the speech signal. The thus generated synthetic speech signal can then be output directly via a playback device or stored for later transmission.

In other words, low-frequency signal components of the speech signal are generated synthetically, that is to say reconstructed, and admixed with the remaining recorded speech signal. The reconstruction of the low-frequency speech components is done on the basis of the non-filtered speech signals. This is exploited that the low-frequency speech components are accompanied by higher-frequency components of the harmonics, so that can be estimated from the remaining signal, the missing portions.

In addition to the fundamental frequency, the frequencies of the harmonics of the fundamental frequency arranged below the limit frequency are preferably determined and used in addition to the fundamental frequency for a reconstruction of the low-frequency frequency range. Thus, the maximum information regarding the undisturbed speech signal is utilized from the spectrally evaluated section of the speech signal. The frequencies used for the reconstruction are combined with a respective spectral distribution and a predetermined amplitude to form a synthetic spectrum which corresponds to the frequency range below the cutoff frequency in the speech signal. From this frequency section and the speech signal above the cutoff frequency, the reconstructed speech signal is then composed. The low-frequency speech component thus no longer has a noise signal, since it is composed exclusively of frequency components of the speech signal.

In a further embodiment of the invention, the low-frequency speech component can also be determined directly from the speech signal. For this purpose, a comb filter consisting of several band filters is set up on the basis of the fundamental frequency and the frequencies of the harmonics arranged below the cutoff frequency, the frequency positions of the individual bandpass filters corresponding to the cutoff frequencies and the harmonics. With the aid of the comb filter, the speech signal is then filtered in the range below the cutoff frequency, whereby the signal components which belong to the actual speech signal are transmitted. Also in this way, a reconstruction of a largely undisturbed speech signal in the low-frequency range of the speech signal is possible.

Decisive for the quality of the reconstruction of the low-frequency speech component is the accuracy of the determined fundamental frequency of the speech signal. Since the fundamental frequency changes continuously during speech due to the sentence melody, a further improvement of the method is achieved by determining the fundamental frequency from the speech signal at the beginning of a speech section and subsequently adaptively tracking it. Thus, in each case the current fundamental frequency is determined in the time course of the speech signal, so that the reconstruction of the speech signal can be adapted as accurately as possible to the voice response. An embodiment of such an adaptive tracking will be explained in detail below.

In a further preferred manner, the amplitude of the at least one frequency signal generated below the cutoff frequency is determined as a function of the amplitudes of the frequency signals analyzed above the cutoff frequency. In a further preferred manner, typical amplitude profiles of speech signals can be used in order to achieve as exact as possible adaptation to a natural speech signal not only in the frequency components but also in the amplitude distribution of the frequency components.

It is further preferred that the cutoff frequency is determined as a function of the noise level, that is to say, in particular, on the size of the interfering signal. Thus, at low noise levels, for example, it is only necessary to reconstruct the speech signal component below 200 Hz, while at high noise levels it is necessary to reconstruct the speech signal in the frequency range below 500 Hz. When using the method in a moving motor vehicle, the cut-off frequency can also be determined as a function of the driving speed.

Furthermore, a development consists in that the speech signal is subjected to a noise suppression before conversion. In this case, the conventional methods known from the prior art can be used to perform a pretreatment of the speech signal. The speech components then emerge more clearly in the spectrum and can be recognized more clearly and therefore more accurately and reconstructed.

One application of the method described above is to reproduce voice signals recorded in a moving motor vehicle in order to reproduce the most natural possible language impression.

Another application of the method according to the invention is to reproduce a voice signal transmitted by means of a telephone connection. The underlying problem is that the voice signals for telephone connections in the frequency range below 350 Hz contain no information. Therefore, for a faithful reproduction of the speech signal, the low-frequency speech component must be reconstructed from the frequency range above 350 Hz. This can be carried out in a particularly advantageous manner by the method according to the invention.

Fig. 1

eine spektrale Innengeräuschverteilung in einem fahrenden Kraftfahrzeug für unterschiedliche Fahrgeschwindigkeiten,

Fig. 2

ein Spektrogramm eines im tieffrequenten Bereich von einem Störsignal überlagerten Sprachsignals,

Fig. 3

ein Spektrogramm des in Fig. 2 dargestellten Sprachsignals ohne Störsignal,

Fig. 4

ein Spektrogramm des in Fig. 3 dargestellten Sprachsignals ohne Frenquenzanteile unterhalb der Grenzfrequenz von ca. 400 Hz,

Fig. 5

ein Spektrogramm der im Spektralbereich unterhalb der Grenzfrequenz von ca. 400 Hz rekonstruierten Sprachanteile,

Fig. 6

das vollständige rekonstruierte Sprachsignal entsprechend dem in Fig. 3 dargestellten Sprachsignal ohne Störsignalanteil,

Fig. 7

ein Blockschaltbild eines Ausführungsbeispiels einer erfindungsgemäßen Vorrichtung zur Rekonstruktion tieffrequenter Sprachanteile aus mittelhohen Frequenzanteilen,

Fig. 8

eine Einrichtung zur adaptiven Nachführung der Grundfrequenz und

Fig. 9

die spektrale Verteilung der Kennlinien der Bandfilter des Regelelementes zum Feststellen der frequenzabhängigen Leistungsverteilung im Mischspektrum in Bezug auf die feststehende Mischungsfrequenz von 2000 Hz.

According to a further teaching of the present invention, the above technical problem is solved by a device having the features of claim 12, while in the claims 13 to 16 advantageous embodiments are given. The device and the method performed therewith will be explained in more detail below with reference to embodiments, reference being made to the accompanying drawings. In the drawing show

Fig. 1

a spectral internal noise distribution in a moving motor vehicle for different driving speeds,

Fig. 2

a spectrogram of a voice signal superimposed on a noise signal in the low-frequency range,

Fig. 3

a spectrogram of the speech signal shown in FIG. 2 without interfering signal,

Fig. 4

a spectrogram of the speech signal shown in FIG. 3 without frenquency components below the cut-off frequency of approximately 400 Hz,

Fig. 5

a spectrogram of the speech components reconstructed in the spectral range below the cutoff frequency of approximately 400 Hz,

Fig. 6

the complete reconstructed speech signal corresponding to the speech signal shown in FIG. 3 without interfering signal component,

Fig. 7

FIG. 2 shows a block diagram of an exemplary embodiment of an apparatus according to the invention for the reconstruction of low-frequency speech components from medium-high frequency components, FIG.

Fig. 8

a device for the adaptive tracking of the fundamental frequency and

Fig. 9

the spectral distribution of the characteristic curves of the band filter of the control element for determining the frequency-dependent power distribution in the mixed spectrum with respect to the fixed mixing frequency of 2000 Hz.

In Figs. 1 and 2, the starting point of the present invention is shown.

Fig. 1 shows a frequency-amplitude diagram of the interior noise level in a moving motor vehicle for different speeds between 60 Km / h and 160 Km / h. In this illustration, it is noticeable that, especially at low frequencies below about 500 Hz, the inner noise level rises sharply in comparison to the other frequencies of the inner noise signal. However, since with normal pitch the fundamental frequency and the first harmonics to the fundamental frequency in the frequency range below 1000 Hz and in particular below 500 Hz, a determination, so filtering out the speech signal from the interior noise signal is considerably more difficult.

Fig. 2 shows a speech signal superimposed on a background signal in a time-frequency representation as a spectrogram. This spectrogram is obtained, for example, by a Fourier transform (FFT) from a microphone signal. In Fig. 2, different gray levels of the individual segments of the spectrogram indicate different intensities. On the one hand, one clearly recognizes the increasing intensity (lighter gray values) in the range of lower frequencies to the value zero and on the other hand, narrow-band frequency components that run largely parallel to each other over short periods of time. These latter narrow-band frequency components represent harmonics of the fundamental frequency of the corresponding voice signal, which are evaluated according to the invention as described below.

FIG. 3 shows a spectrogram of the speech signal shown in FIG. 2 without the background noise, so that the low-frequency speech components can also be recognized as narrow-band frequency components in the spectrogram below 500 Hz. These language parts need to be reconstructed.

FIG. 4 further shows the previously described speech signal, in which the speech components are cut off below a cutoff frequency of approximately 400 Hz. Such a signal is approximately the same as the voice signal transmitted on a telephone connection.

FIG. 5 shows an example of a reconstructed speech signal in the range below the cutoff frequency of approximately 400 Hz, and FIG. 6 shows the composite reconstructed speech signal from the reconstructed speech component shown in FIG. 5 and the frequency component shown in FIG. 4 above the cutoff frequency of the original one spectrum. How the reconstructed speech parts are obtained will be described in detail below with reference to FIGS. 7 to 9.

7 shows in a block diagram an apparatus for the reconstruction of low-frequency speech components from medium-high frequency components. The speech signal is _fed to a means 4 for determining frequency components ω _fa1 , ω _fa2 ,... Of maxima in the speech signal above a predetermined limit frequency ω ₀ . For this purpose, the speech signal is first passed through a bandpass filter 6, so that only the frequency components between the cutoff frequency ω ₀ and another frequency ω ₁ cut out and forwarded to further processing. In this case, ω ₀ is, for example, in the range of 200 to 500 Hz, in particular 350 Hz, while the frequency ω _{1 is,} for example, in the range of 800 Hz. The thus-filtered frequency portion of the speech signal is mixed in the mixing element 8, so that the sum and difference frequencies of the frequency components contained in the cut-out portion of the speech signal are formed. Of interest are the difference frequencies, so that the signal emerging from the mixing element 8 is processed by means of a low-pass filter, so that only frequency components below an adjustable frequency ω _{2 are} transmitted. Thus, the smallest difference frequency can be determined, the distance between two in the Speech signal adjacent to each other arranged spectral components corresponds. Since these are two harmonics of the fundamental frequency, the difference frequency represents the fundamental frequency ω _g . This fundamental frequency is then fed to means 12 for reconstructing the speech signal. Via a further input of the means 12, the speech signal via a delay stage 14 and a low-pass filter 16 is supplied. Thus, both the value of the fundamental frequency ωg and a predetermined frequency section of the speech signal are available to the means 12 for reconstruction of the signal containing the speech. The delay stage 14 serves to compensate for the time .DELTA.t, which is needed for the determination of the fundamental frequency ω _g and the low-pass filter 16 is a useful reduction of the amount of data, which is fed to the means 12 for the reconstruction of the speech signal.

The means 12 for the reconstruction of the speech signal below the cut-off frequency ω ₀ has two alternatives of procedures in terms of circuitry.

As a first alternative, the fundamental frequency ω _{g is} used to generate a signal in the reconstructed speech signal corresponding to the root of the speech. In addition, the frequencies of the harmonics to the fundamental frequency ω _g by simply multiplying with the numbers N = 2, 3, 4, ... are determined, so that for a reconstruction of the speech component below the cut-off frequency ω _{0 in} addition to the fundamental frequency ω _g also the frequencies ω _h1 , ω _h2 , ... of the first, second and further harmonics arranged below the cut-off frequency ω ₀ are used. The aim is to generate all harmonics in the frequency section of the speech signal to be reconstructed, that is to simulate them. For a spectral distribution around each of these frequencies, an approximate Gaussian distribution or other possible spectral distribution is assumed, which can be defined by a half-width and an amplitude. As a result, the spectral sections shown in FIG. 5 can be generated in the spectrogram, which can not be recognized or only slightly recognized in the noisy signal shown in FIG. 2.

As a further alternative for a reconstruction of the low-frequency speech component, there is the possibility that the means 12 comprise a comb filter comprising a plurality of band filters whose spectral transmission functions are determined by the fundamental frequency ω _g and the frequencies ω _h1 , ω _h2 ,.. , The spectral transmission function of each bandpass filter is also defined over a predetermined width, so that corresponding spectral sections are filtered out of the speech signal in the range of low frequencies below the cutoff frequency ω ₀ . Since from the spectrogram only the proportions are filtered out, containing the speech signal, the speech signal is reconstructed from the spectrogram. If additionally a noise suppression is carried out, then the background noises are filtered out of the filtered-out signal components, so that an almost natural speech signal is generated.

As can further be seen in FIG. 7, the speech signal is delayed by a further delay stage 18 by a time difference Δt, in order to make it possible to adapt to the time span necessary for the reconstruction of the low-frequency speech component. After passing through a high-pass filter 20 in which the speech signal above the cut-off frequency ω _{0 is} filtered out, both this high-pass filtered signal and the reconstructed speech signal for frequencies ω <ω ₀ converge in the summation element 22, from which the reconfigured spectrogram shown in FIG becomes. This spectrogram therefore consists on the one hand of the frequency component reconstructed below the cutoff frequency ω ₀ and of the original frequency spectrum above the cutoff frequency ω ₀ . The spectrogram thus produced, after conversion to a loudspeaker signal, results in almost natural-sounding speech reproduction.

As already explained above, in general the fundamental frequency ω _g in a speech signal does not remain constant due to the speech melody. Therefore, it is necessary to constantly redetermine the fundamental frequency ω _g . This can on the one hand be done by constantly going through the process described above, which has been previously described with reference to the elements 4, 6, 8 and 10. On the other hand, however, a more accurate adaptive tracking of the fundamental frequency ω _g can be performed. This is possible with a device which is shown in Fig. 8.

The fundamental frequency ω _{g, 0} initially determined at the beginning of a speech signal is multiplied to N times the value by means of a multiplication element 24. Thus, the (N-1) th harmonic is calculated to the fundamental frequency. The frequency of these harmonics is hereinafter referred to as harmonic and the associated frequency denoted by ω _r .

The frequency ω _r is introduced via a Mehrtorschalter in a control loop. In an initialization phase at the beginning of a word, the output of the multiplication element 24 is transferred from the multi-port switch 26 to the mixing element 28. After a short time - as described below - an estimated value ω _r , new before and the Mehrtorschalter 26 is switched so that ω _r , new to the mixing element 28 is passed.

The aim of the control loop is to determine the difference between the (N-1) th harmonic and a fixed frequency of, for example, ω _m = 2000 Hz. Ideally, ω _{r is} exactly the frequency of the (N-1) th harmonic. The mixing element 28 forms the difference between ω _r and ω _m . A sine wave generator generates a sinusoidal signal with the frequency given by its input signal ω _d . This is fed to a mixing element 32 which mixes the speech signal and this sinusoidal signal. After mixing, the mixed signal is outputted from the mixing element 32, which is supplied to a control element 34 for detecting the frequency-dependent power distribution in the mixing signal with respect to the fixed frequency ω _m .

Assuming that the frequency ω _{r of} the control harmonics supplied to the mixing element 28 exactly matches a harmonic in the current speech signal, the sum of the difference frequency ω _d , which has been produced by the difference with the fixed mixing frequency ω _m and ω _r , and one of the rule harmonic corresponding frequency component of the speech signal exactly the mixing frequency ω _m . This is reflected in a power distribution (P distribution) in the service spectrum. The power distribution will be maximal at the mixing frequency ω _m .

However, if the frequency ω _{r of} the control harmonics does not correspond to the actual frequency of the corresponding harmonics in the speech signal, the power distribution will assume its maximum not at the frequency ω _m but at a position shifted by a difference value Δω. Thus, a correction value to Δω can be determined, which is added to the current value of the frequency ω _{r of} the control harmonics added. This results in the new value of the frequency ω _{r, new} , which is fed again via the multiport switch 26 of the control loop. Subsequently, a mixture is again in the mixing element 28 with subsequent control sequence, as has been previously described. Thus changes in the course of the speech signal, the fundamental frequency and thus the frequency of the corresponding harmonics in the speech signal, this is compensated by the control loop, so that constantly generates a current, with the fundamental frequency ω _r largely matching value ω _r .

9 shows the characteristics of a plurality of band filters which are provided for a determination of the power distribution in the control element 34. 9 results in a number of 7 band filters which are arranged around the fixed mixing frequency ω _m = 2000 Hz. If, for example, the maximum power drops into the passband of the middle bandpass filter, then the correction value Δω = 0 is set. If, on the other hand, the maximum lies in one of the adjacently arranged bandpass filters, a corresponding correction value Δω ≠ 0 is generated in order to shift the maximum of the spectral power distribution into the passband of the middle bandpass filter when the control is continued.

The value ω _r is diverted from the control loop via a multiplication element 38 and output, in which the current frequency ω _{r is applied} with the factor 1 / N to adapt the value of the fundamental frequency ω _{g, adapt} . Thus, the value of the fundamental frequency ω _{g is} constantly adaptively tracked, whereby the reconstruction of the low-frequency speech component from the medium-high frequency components is improved and brought closer to a natural speech signal.

Claims (16)

1. Method for reconstruction of low-frequency speech components from medium-level frequency components,

   - in which at least two frequency components (ω_fa1, w_fa2, ...) which are arranged adjacent and have an increased amplitude in the speech signal are determined above a cut-off frequency (ω₀), and

   - in which the fundamental frequency (ω_g) of the speech signal is determined as the frequency difference between the at least two adjacent frequency components (ω_fa1, ω_fa2, ...), and

   - in which the low-frequency frequency range below the cut-off frequency (ω₀) is reconstructed with the aid of the determined fundamental frequency (ω_g) and the speech signal.
2. Method according to Claim 1, in which the frequencies (ω_h1,ω_h2, ...) of the harmonics of the fundamental frequency (ω_g) which are arranged below the cut-off frequency (ω₀) are determined from the fundamental frequency (ω_g), and are used together with the fundamental frequency (ω_g) for reconstruction of the low-frequency frequency range.
3. Method according to Claim 1, in which the frequency positions of the band filters, with whose aid the speech signal is filtered in- the range below the cut-off frequency (ω₀), are set up with the aid of a comb filter, which has a plurality of band filters, on the basis of the fundamental frequency (ω_g) and the frequencies of the harmonics which are arranged below the cut-off frequency (ω₀).
4. Method according to one of Claims 1 to 3, in which, at the start of a speech section which contains speech, the fundamental frequency (ω_g) is determined from the speech signal, and the fundamental frequency (ω_g) is then adaptively readjusted.
5. Method according to Claim 4,

   - in which the frequency (ω_r) of a control harmonic is calculated as the N-th harmonic from the instantaneous value of the fundamental frequency (ω_g) for adaptive readjustment of the fundamental frequency (ω_g),

   - in which the difference between the frequency (ω_r) of the control harmonic and a fixed mixing frequency (ω_m) are formed,

   - in which a sinusoidal signal (sin(ω_d)) is produced using the difference or sum frequency (ω_d) resulting from the subtraction process,

   - in which the sinusoidal signal (sin (ω_d)) is mixed with the speech signal and a mixed signal is produced,

   - in which the frequency-dependent power distribution in the mixed signal is fixed with respect to the fixed mixing frequency (ω_m),

   - in which a correction value (Δω) for the frequency (ω_r) of the control harmonic is calculated from the power distribution,

   - in which the frequency (ω_r) of the control harmonic is changed by the correction value (Δω), and is supplied to a mixing process once again, with the fixed mixing frequency (ω_m), and

   - in which the fundamental frequency (ω_g) which corresponds to the corresponding fraction 1/N of the frequency (ω_r) is emitted.
6. Method according to Claim 5, in which, in order to determine the power distribution, the mixed signal is supplied to a plurality of band filters (BF_n), which cover adjacent frequency ranges, centred about the fixed mixing frequency.
7. Method according to one of Claims 1 to 6, in which the amplitude of the at least one frequency signal which is produced below the cut-off frequency is determined as a function of the amplitudes of the frequency signals which are analysed above the cut-off frequency.
8. Method according to one of Claims 1 to 7, in which the cut-off frequency is determined as a function of the noise level.
9. Method according to one of Claims 1 to 8, in which the speech signal is subjected to an interference-signal removal process before conversion to a spectrogram.
10. Use of a method according to one of Claims 1 to 9 for reproduction of a speech signal which has been recorded in a moving motor vehicle.
11. Use of a method according to one of Claims 1 to 9 for reproduction of a speech signal which is transmitted by means of a telephone link.
12. Apparatus for reconstruction of low-frequency speech components from medium-level frequency components, in particular for carrying out a method according to one of Claims 1 to 11, and

    - having means (4) for determination of frequency components (ω_fa1, ω_fa2, ...) of maxima in the speech signal above a predetermined cut-off frequency (ω₀),

    - having means (8) for mixing of the frequency components (ω_fa1, ω_fa2, ...) for determination of the fundamental frequency (ω_g) of the speech signal as the difference frequency between two respectively adjacent frequency components (ω_fa1, ω_fa2, ...), and

    - having means (12) for reconstruction of the speech signal below the cut-off frequency (ω₀) from the determined fundamental frequency (ω_g) and the speech signal.
13. Apparatus according to Claim 12, characterized in that the means (12) for reconstruction of the speech signal below the cut-off frequency (ω₀) determine the spectrogram from the fundamental frequency (ω_g) and the frequencies (ω_h1, ω_h2, ...) of those harmonics of the fundamental frequency (ω_g) which are arranged below the cut-off frequency (ω₀) with a predetermined spectral distribution and a predetermined amplitude distribution.
14. Apparatus according to Claim 12, characterized in that the means (12) have a comb filter with a plurality of band filters, with the frequencies of the band filters being variable on the basis of the fundamental frequency (ω_g) and, possibly, one or more harmonics of the fundamental frequency (ω_g) which are arranged below the cut-off frequency (ω₀).
15. Apparatus according to one of Claims 12 to 14, characterized in that the following items are provided for adaptive readjustment of the fundamental frequency (ω_g):

    - a multiply element (24) for production of the N-th harmonic of the fundamental frequency as the frequency (ω_r) of a control harmonic,

    - a mixing element (28) for mixing of the frequency (ω_r), of the control harmonic with a fixed mixing frequency (ω_m),

    - a sine-wave generator (30) for' mixing of the difference or sum frequency (ω_d) which results from the mixing process,

    - a mixing element (32) for mixing of the sinusoidal signal (sin (ω_d)) with the speech signal and for production of a mixed signal,

    - a control element (34) for fixing the frequency-dependent power distribution in the mixed signal with respect to the fixed mixing frequency (ω_m) and for calculation of a correction value (Δω) for the frequency (ω_r) of, the control harmonic from the power distribution,

    - a mixing element (36) for variation of the frequency (ω_r) of the control harmonic by the correction value (Δω), and

    - having a multiply element (38) for calculation of the fraction 1/N of the frequency (ω_r) as the fundamental frequency (ω_g).
16. Apparatus according to Claim 15, characterized in that the control element (34) has a plurality of band filters, which cover adjacent frequency ranges centrally with respect to the mixing frequency (ω_m).

Images