DE19860133C2 - Method and device for speech compression - Google Patents

Description

The invention relates to a method and a direction for compression of a digitized Speech signal using reference indices variable entries from sender and recipient existing reference memories.

For efficient voice transmission over one News channel is a compression of the speech absolutely indispensable. This applies in particular if the bandwidth of the message channel through physi cal wheel conditions is limited, such as. B. in the mobile telephone area.

Standard compression methods can be used which are compressed any time signal can distinguish from such procedures that finally for the compressed transmission of Language are suitable. With such special procedures ren a much higher compression rate achieve.

So-called codebook procedures for compressed transfer carrying language, such as that from the DE 35 13 243 A1, DE 40 33 350 A1 or DE 35 21 413 A1 are known can be divided into three different phases  subdivide, namely a transmitter-side detection fixed, process-specific language elements, a transfer of the sequence of those indices Speech elements corresponding to the spoken signal and a receiver synthesis of the language the language elements, corresponding to the transmitted ones Indices.

The procedural language elements may differ are at different levels of language. So according to DE 35 21 413 A1, the speech signal from gan zen words put together. However, this has the Disadvantage that only a limited, previously agreed Vocabulary can be transferred. Speech synthesis the level of the phonemes, d. H. that of human Speech apparatus generate basic sounds, is described in DE 30 06 339 A1 and DE 31 05 518 A1. Other, More abstract language elements can e.g. B. with help of the Mel-Cepstrum process define how it works DE 40 33 350 A1 is to be used.

Still other methods use as elementary signals no language modules in the real sense, but artificial signals that are put together so that an acoustic speech signal is generated. From DE 30 28 000 A1 and DE 32 18 755 A1 it is known between between voiced and unvoiced sounds divorce. In the latter process, Speech synthesis sound and noise generator signals ver uses that are modulated by controllable filters.

All of the above methods have in common that the Synthesis used procedural language elements are fixed and not on the respective transmitter side  Speakers can be adjusted. Thereby it is generally impossible to hear the sound of the voice of a speaker to faithfully reproduce what however, a telephone connection is desirable.

The present invention is based on the object de, such speech for the receiver-side speech synthesis to use elements that were previously ex have been married.

This object is achieved in that the reference memory is filled by an elementary signal s _E (t) at times t ₀ in a compression module on the transmitter side using the auto correlation function ϕ _ss (t, τ) of the speech signal s (t) ) and / or elementary spectrum S _E (f) is generated, which in the synthesis module on the receiver side as well as in the transmitter-side compression module as a reference signal s _i (t) and / or reference spectrum S _i (f) in the reference memory under the index i is stored if this elementary signal s _E (t) and / or elementary spectrum S _E (f) has no similarity to any of the existing reference memory entries.

These measures make it possible to reproduce the typical voice characteristics of a speaker in the speech signal synthesized on the receiver side. The speech elements consist of the smallest meaningfully definable units of the speech signal. With voiced sounds, an elementary signal s _E (t) corresponds to z. B. a period of the fundamental; in the case of unvoiced sounds, an elementary spectrum S _E (f) is used, which is obtained by Fourier analysis over a predetermined time interval. These Sprachele elements are stored on the transmitter and receiver side as reference signals s _i (t) or reference spectra S _i (f), so that they do not have to be retransmitted in the event of their repeated occurrence, which results in the compression.

Further advantageous measures are in the Unteran sayings described. The invention is in the lying drawing and is shown below described in more detail; it shows:

Figure 1 is a compression module that processes a digitized speech signal s (t) to a compressed code that consists of data blocks of variable length.

Fig. 2 shows a synthesis module that generates a speech signal s (t) from the compressed data, in which a received data block is to be processed, determines its identifier, addresses a reference signal si (t) or a reference spectrum Si (f) and reads, a signal generator adds this to the reference memory, in the case of a spectrum after previous inverse Fourier transformation, stores the speech signal s (t) to be synthesized, the volume and possibly frequency information E (t) contained in the data block are processed;

Fig. 3 is an averaging member for determining the instantaneous signal level E (t), the amount of the speech signal s (t) a time interval .DELTA.t is integrated;

Figure 4 is an autocorrelator for computing the autocorrelation function.

Fig. 5 is a cross-correlator for calculating the correlation coefficients φ for comparing an elementary signal s _E (t) with the reference signals s _i (t) stored in the reference memory _iE;

Fig. 6 is a spectrum analysis for calculating ei nes elementary vector W _E of a mentarsignalspektrum Ele S _E (f).

The invention shown in FIGS. 1 to 6 is explained in more detail below using an exemplary embodiment. The compression module 10 and the synthesis module 22 are considered separately. At the beginning of the signal processing in the compression module 10 , the momentary volume E (t) of the digitized speech signal s (t) is determined with an averaging element 11 , as shown in FIG. 3.

For this purpose, an averaging element 11 , as shown in FIG. 3, determines the current volume E (t). An autocorrelator 12 , as shown in FIG. 4, calculates the autocorrelation function ϕ _ss (t, τ) of the speech signal. A downstream maximum detector 13 determines the maximum ϕ _ss (t ₀ , τ).

The delay time τ _max corresponds to the period length of an elementary signal s _E (t). A normalization element 14 brings the elementary signal to a predetermined period length and energy. At the same time, a Fourier transformer 15 calculates the current spectrum S (t, f) of the speech signal.

By means of spectral analysis 18 , as shown in FIG. 6, an elementary vector W _E is obtained from an elementary spectrum S _E (f), which element is compared in a comparison unit 19 with the reference vectors W _{i of} the reference memory 16 . If the distance to the most similar reference vector is less than a threshold ϑ, instead of the entire elementary signal s _E (t) the corresponding reference index i _min as well as volume information E (t) and frequency information (τ _max ) are output in a data block 21 . Alternatively, the elementary signal is compared with the reference signals s _i (t) by a cross correlator 17 , as shown in FIG. 5. If there is no suitable reference, the elementary signal s _E (t) and possibly the spectrum S _E (f) and the elementary vector W _{E are} stored in the reference memory. In this case, the entire elementary signal s _E (t) or the elementary spectrum S _E (t) is added to the data block ( 21 ).

With the aid of E (t), an autocorrelator 12 , as shown in FIG. 4, calculates the normalized autocorrelation function ϕ _ss (t, τ) of the speech signal s (t), ϕ _ss (t, τ) calculated at a defined time interval Δt and changes as time t progresses.

A fixed point in time t = t _{0 is} considered below. A maximum detector 13 determines the maximum of the auto correlation function ϕ _ss (t ₀ , τ), which is located at the point τ _max <0. It is particularly advantageous if the length of the time interval Δt on which ϕ _ss (t ₀ , τ) is calculated corresponds to the value of τ _max . Using the value of Wer _ss (t ₀ , τ _max ), the compression module 10 decides as follows whether the speech signal s (t) is voiced or unvoiced at time t ₀ . If ϕ _ss (t ₀ , τ _max ) exceeds a threshold to be specified, it is a voiced sound.

An unvoiced sound essentially consists of noise, so its autocorrelation function has no pronounced maximum for τ _max <0. If there is a harmonious sound, the speech signal is processed by a normalization element 14 to form an elementary signal s _E (t) which has a predetermined length, energy and phase position. This normalized elementary signal s _E (t) consists of a single period of the speech signal s (t). The length of this period corresponds to the value of τ _max . The elementary signal s _E (t) is compared with the reference signals s _i (t) stored in the reference memory. With voiced sounds, this can be done in two ways.

In the exemplary embodiment described below, voiced and unvoiced sounds are compared by means of spectral analysis 18 in the frequency domain. For this purpose, a Fourier transformer 15 calculates the time-varying spectrum S (t, f) of the speech signal s (t). The normalized magnitude spectrum at a point in time t = t ₀ is called the elementary spectrum S _E (f) below. S _E (f) can also be obtained by Fourier transformation of the autocorrelation ϕ (t ₀ , τ). Using spectral analysis 18 , an elementary vector W _{E is} calculated from an elementary spectrum S _E (f) by using filter functions F _k (f), as shown in FIG. 6.

With the help of this elementary vector W _E , the comparison with the entries of the reference memory 16 takes place in the following manner. A reference vector W _i corresponds to each stored reference signal s _i (t) or reference spectrum S _i (f), which corresponds to the element vector W _{E of} the respective reference signal just explained. The comparison of the elementary vector W _E with the reference vectors W _i takes place in a comparison unit 19 .

Let i _{min be} the index of the reference vector W _i with the smallest distance of all W _{i from} the elementary vector W _E. The inverse of this distance is a measure of the similarity of the signals or spectra. If the minimum distance is smaller than a threshold value to be specified, the elementary signal s _E (t) or the elementary spectrum S _E (f) can be replaced by the corresponding reference. In this case, the comparison unit 19 only enters the index i _min together with the correct volume E (t) in the data block 21 to be transmitted. If the threshold value is selected accordingly, any sound can be replaced by a reference.

In the case of a voiced sound, the data block 21 additionally contains the value of τ _max , which is required for the synthesis of the speech signal with the correct fundamental frequency. Each data block 21 begins with an identifier 24 , in which the type of the information transmitted is encoded.

In another embodiment of the invention, separate reference memories are used for voiced and unvoiced sounds. In yet another embodiment of the invention, comparison unit 19 and reference memory 16 are combined in a self-organizing neural network, which is distinguished by a particularly clever handling of reference vectors W _i .

In yet another embodiment, voiced sounds are compared in the time domain. Instead of comparing an elementary vector W _E with the reference vectors W _i , the elementary signal s _E (t), as shown in FIG. 5, is compared by cross-correlation with the reference signals s _i (t). For this purpose, a cross correlator 17 calculates the correlation coefficients ϕ _iE .

If the minimum correlation coefficient ϕ _{iE is} smaller than a predefined threshold ϑ, the elementary signal is regarded as known and - as explained above - treated. The processing of voiceless sounds continues as in the first example.

All of the above-mentioned exemplary embodiments have in common that a complete elementary signal s _E (t), or in the case of unvoiced sounds, an elementary spectrum S _E (f) is transmitted in the data block 21 to be transmitted if no suitable reference is found. This is particularly the case when the reference memory is empty, e.g. B. at the beginning of a transmission in which references to a previous transmission or predetermined star references are not used. Each elementary signal s _E (t) or elementary spectrum S _E (f) entered in a data block 21 to be sent is simultaneously stored in the reference memory 16 together with the associated reference vector W _i .

If the size of the reference memory is limited, it is necessary to overwrite old entries. To do this, you select the cheapest entries that are relatively old and yet only rarely selected as suitable references for the transfer. The index i of the reference memory position of the new element is also entered in the data block 21 to be transmitted in order to store it in the reference memory of the receiver at the same index position i as the reference signal s _i (t), or reference spectrum S _i (f).

At the receiver end, the transmitted data blocks are evaluated by a synthesis module 22 , as shown in FIG. 2. A signal generator 23 composes the synthesized speech signal from the received elementary signals s _E (t) and elementary spectra S _E (f) as well as from indexed reference signals s _i (t) and reference spectra S _i (f). For this purpose, the identifier 24 of a data block specifies how the received data are to be treated. Spectra S _E (f) and S _i (f) must first be converted into time signals by inverse Fourier transformation.

The volume E (t) contained in data block 21 is used to emulate the correct signal volume. If voiced sounds are composed of elementary signals s _E (t) or reference signals s _i (t), the value of the delay τ _max also contained in the data block 21 serves to restore the correct frequency of the basic oscillation of the speech signal. The storage of received elementary signals s _E (t) and elementary spectra S _E (f) at the predefined index position i of a reference memory 16 is essential for the functioning of the synthesis module.

The corresponding reference index i is specified in each data block 21 . This ensures that the reference memory 16 of the compression module 10 on the side of the transmitter and the synthesis module 22 on the side of the receiver always have the same entries.

The method described allows the transmission of spoken speech with a transmission rate of less than 1 kbit / s with very good speech quality. Even with egg nem reference memory 16 , which only offers space for one reference and thus only allows repetition of the last transmitted elementary signal s _E (t) or elementary spectrum S _E (f), a considerable compression of the speech signal can already be achieved to reach.

reference numeral

10th

Compression module that a speech signal s (t) too Data blocks (

21

) processed;

11

Averaging element that determines the current volume E (t) of the speech signal s (t);

12th

Autocorrelator, determines the autocorrelation function ϕ _ss

(t) of the speech signal s (t);

13

Maximum detector, which is the maximum of the autocorrelation function ϕ _ss

(τ) determines that at the point τ _max

= 0;

14

Normalization element that converts a period of the speech signal s (t) to a normalized elementary signal s _E

(t) defined length and energy processed;

15

Fourier transformer, calculates the spectrum S (f) of the speech signal s (t);

16

Reference memory in which reference signals s _i

(t) and reference spectra S _i

(f) and the associated reference vectors W _i

get saved;

17th

Cross correlator for calculating the cross correlation coefficient ϕ _iE

;

18th

Spectral analysis based on a spectrum S _E

(f) an elementary vector W _E

generated;

19th

Comparison unit that decides whether in the Refe reference memory (

16

) is a suitable reference to the elementary signal s _E

(t) to replace;

20th

Transmission channel over which the data blocks (

21

) are sent;

21

Data block that contains the information on the speech synthesis on the receiver side;

22

Synthesis module that blocks the data (

21

) receives and a reference memory (

16

) and a signal generator for speech synthesis;

23

Signal generator consisting of received elementary signals s _E

(t) and stored reference signals s _i

(t) speech signal s (t) generated;

24th

Identifier of a data block (

21

), which characterizes its data;
E (t) instantaneous volume of the speech signal s (t);
f frequency
k Component indexing of elementary vector W _E

and the filter functions F _k

(f);
F _k

(f) Filter functions with which by spectral analysis (

18th

) from a spectrum S _E

(f) an elementary vector W _E

is calculated;
i Index with which references in the reference memory are addressed;
i _min

Index of the reference vector with the minimum distance from the elementary vector W _E

;
s (t) digitized speech signal;
S (t, f) normalized magnitude spectrum of the speech signal in a predetermined time interval;
s _E

(t) normalized elementary signal, corresponds to a period of the fundamental wave of s (t) at a time t = t ₀

, scaled to constant length and energy;
S _E

(f) normalized elementary spectrum, corresponds to the spectrum S (t, f) at time t = t ₀

;
s _i

(t) reference signal, in the reference memory (

16

) filed elementary signal s _E

(t);
S _i

(f) reference spectrum, in the reference memory (

16

) stored elementary spectrum S _E

(f);
t ₀

arbitrary point in time
W _E

Elementary vector, by spectral analysis (

18th

) from elementary spectrum S _E

(f) generated;
W _Ek

Components of the elementary vector W _E

;
W _i

Reference vector, one in the reference memory (

16

) stored elementary vector W _E

;
Δt time interval at which the volume E (t) and the autocorrelation function ϕ _ss

(t, τ) or the cross correlation coefficient ϕ _iE

be calculated;
ϕ _ss

(t ₀

, τ) autocorrelation function of the speech signal s (t) at time t ₀

;
ϕ _iE

Cross-correlation coefficient of elementary signal s _E

(t) and reference signal s _i

(t);
ϑ Threshold when comparing elementary signal s _E

(t) and reference signals s _i

(t);
τ delay time as argument of the autocorrelation function ϕ _ss

(t, τ)
τ _max

Delay time at which the maximum of the Autocor relation function ϕ _ss

(t ₀

, τ) occurs for τ <0, which speaks to the period length of the speech signal s (t) ent.

Claims (16)

1. A method for compressing a digitized speech signal s (t) by using reference indices (i) of variable entries from the transmitter and receiver side existing reference memories ( 16 ), characterized in that the reference memories ( 16 ) are filled by in a compression module ( 10 ) on the side of the transmitter using the autocorrelation function ϕ _ss (t, τ) of the speech signal s (t) at times t ₀ an elementary signal s _E (t) and / or elementary spectrum S _E (f) is generated, which both in the synthesis module ( 22 ) on the side of the receiver and in the transmitter-side compression module ( 10 ) as reference signal s _i (t) and / or reference spectrum S _i (f) in the reference memory ( 16 ) under the index i, if this is stored Elemental signal s _E (t) and / or elementary spectrum S _E (f) has no overlap similarity to any of the existing reference memory entries. 2. The method according to claim 1, characterized in that when storing new reference signals s _i (t) or reference spectra S _i (f) in the reference memory ( 16 ), provided that there is no free memory, old entries that are rarely overwritten Show similarity with elementary signals s _E (t) or elementary spectra S _E (f). 3. The method according to claim 1 or 2, characterized in that the period length of an elementary signal at time t ₀ corresponds to the delay time τ _max at which a maximum of the autocorrelation function funktion _ss (t ₀ , τ) occurs for τ <0. 4. The method according to any one of claims 1 to 3, characterized in that from the speech signal s (t) at any time t ₀ an elementary vector W _E by applying filter functions F _k (f) on the spectrum S (t ₀ , f ) is calculated, a correspondingly defined reference vector W _{i is} stored for each reference signal s _i (t) and each reference spectrum S _i (f) of the reference memory ( 16 ), the comparison unit ( 19 ) of the compression module ( 10 ) storing the Similarity between elementary signal s _E (t) and reference signal s _i (t) and between elementary spectrum S _E (f) and reference spectrum S _i (f) as an inverse distance ∥W _E - W _i ∥ ^-1 between the elementary vector W _E and the corresponding reference vector W _{i is} defined. 5. The method according to one or more of claims 1 to 4, characterized in that separate reference memories ( 16 ) are used for voiced and unvoiced sounds. 6. The method according to one or more of claims 1 to 5, characterized in that the reference memory ( 16 ) and comparison unit ( 19 ) are combined in a self-organizing neural network. 7. The method according to one or more of claims 1 to 4, characterized in that all reference signals s _i (t) of the reference memory are scaled to the same energy and length and with the aid of the cross correlation coefficient ϕ _iE with an elementary tarsignal s _E (t ) are compared, while reference spectra S _i (f) and elementary spectra S _E (f) are compared as in claim 4. 8. The method according to one or more of claims 1 to 7, characterized in that data blocks ( 21 ) containing the compressed speech information leg are transmitted from the compression module ( 10 ) to the synthesis module ( 22 ), each data block ( 21 ) Has identifier ( 24 ) for characterizing the information contained and a reference index i for specifying a reference storage position. 9. The method according to one or more of claims 1 to 8, characterized in that a data block ( 21 ) additionally contains information about the correct Si signal volume E (t) and the correct fundamental frequency τ _{max of} the speech signal. 10. The method according to one or more of claims 1 to 9, characterized in that a data block ( 21 ) additionally contains an elementary signal S _E (t) or an elementary spectrum S _E (f). 11. Device for compressing a digitized speech signal s (t) by using reference indices (i) variable entries from the transmitter and receiver side existing reference memories, characterized by a compression unit ( 10 ) on the side of the transmitter, the data blocks ( 21 ) va riabler length via a transmission channel ( 20 ) to a synthesis module ( 22 ) on the receiver side. 12. The apparatus according to claim 11, characterized in that the compression module ( 10 ) has an averaging element ( 11 ) for determining the signal volume E (t), an autocorrelator ( 12 ) for determining the autocorrelation function ϕ _ss (t, τ) of the speech signal s (t), a maximum detector ( 13 ) for determining the maximum of the autocorrelation function ϕ _ss (t ₀ , τ _max ) at time t ₀ with the delay τ _max <0, a normalization element ( 14 ) for generating a normalized elementary signal s _E ( t) with period length τ _max , a Fourier transformer for calculating an elementary spectrum S _E (f) and a spectral analysis ( 18 ) for calculating elementary vectors W _E. 13. The apparatus of claim 11 or 12, characterized in that the compression module ( 10 ) in addition to a reference memory ( 16 ) for storing reference signals s _i (t), reference spectra S _i (f) and reference vectors W _i , and a Comparison unit for determining the reference signal s _i (t) and / or reference spectrum S _i (f), which is most similar to an elementary signal s _E (t) and / or elementary spectrum S _E (f). 14. The apparatus of claim 11 or 12, characterized in that the compression module ( 10 ) additionally has a self-organizing neural network that stores reference signals s _i (t), reference spectra S _i (f) and reference vectors W _i , and determines a reference signal s _i (t) and / or reference spectrum S _i (f) which is most similar to an elementary signal s _E (t) and / or elementary spectrum S _E (f). 15. Device according to one of claims 11 to 14, characterized in that the compression module additionally has a cross correlator ( 17 ) for calculating a cross correlation coefficient, with the aid of which an elementary signal s _E (t) is compared with a reference signal s _i (t) becomes. 16. The apparatus according to claim 11, characterized in that the synthesis module ( 22 ) has a signal generator ( 23 ) for synthesis of a speech signal s (t) and a reference memory ( 16 ) for storing at least one reference signal s _i (t) or Has reference spectrum S _i (f).