Method of and arrangement for restoring samples of an equidistantly sampled signal which are considered invalid

0001The invention relates to a method for restoring a sequence of samples S_t+i, wherein i = 0, ..., m-1, which are considered invalid, of an equidistantly sampled signal on the basis of replacement values which are derived from sequences of valid samples.

0002Such method is known from EP-A-0 139 803. In this known method a restoration of samples of a speech signal is based on the fact that the speech signal or a derivation thereof is divided in sub-bands and that the sequence of samples for each sub-band is highly periodic. To obtain the replacement values from a period preceding the sequence of invalid samples, according to this known method, starting from the last valid sample, the first maximum value belonging to a sample n-p and the second maximum value belonging to a sample n-q are determined, and the invalid samples are replaced by corresponding samples in the period between the samples n-p and n-q, on the basis of the assumed periodicity of the signal. The present invention has for its object to improve the known method in such a way that a subdivision of the sampled signal in sub-bands and a high periodicity of these sub-bands is no longer a requirement in order to obtain a proper restoration of the invalid samples.

0003To that end, a method according to the invention is characterized in that    a sequence of values of samples present in a near pre-environment of the sample S_t in combination with values of p samples present in a near post-environment of the samples S_t+m-1 are compared with an equally large sequence of values of valid samples which occurred a predetermined period of time earlier and which are located in a near pre-environment of a sample s_q, wherein q represents a variable sample position in the range <maths id="math0001" num=""><math display="inline"><mrow><mtext>p ≦ q ≦ t-p</mtext></mrow></math><img file="EP0241980B1_D0001.tif" /></maths> and in a near post-environment of the sample s_q+m-1,    such a comparison is made for a plurality of different predetermined periods of time;    that environment which most closely approximates the environment of the sequence of samples to be restored is selected; and    the valid samples bounded by the environment chosen such are used for forming the replacement values.

0004More specifically, the method in accordance with the invention is characterized by the following steps: <ul id="ul0001" list-style="none"><li>a) values of samples in the near pre-environment of the sample s_t, in combination with values of the p samples in the near post-environment of the samples s_t+m-1 are correlated with values of samples which are present in the near pre-environment of the sample s_q, in combination with values of p samples present in the near post-environment of the sample s_q+m-1;</li><li>b) the correlation result of step a) is determined and recorded;</li><li>c) the steps a) and b) are repeated in that sequence for a plurality of differently chosen q-values;</li><li>d) from the set of correlation results determined after step c) the result having the lower value is determined;</li><li>e) the value q corresponding to the correlation result obtained after step b) is determined, more specifically q_opt; and</li><li>f) a sequence of value samples in the positions q_opt, ..., q_opt+m-1 is selected, the replacement values for the samples s_t, ..., s_t+m-1 being formed from the values of the samples thus selected.</li></ul>

0005A preferred embodiment of a method according to the invention, in which the correlation procedure is normalised, is characterized in that each environment determined by a value q opted for is scaled such by a factor a_q that the means square error relative to the environment of the samples to be restored is minimal and    each of the values of the samples in the indicated positions q_opt, ..., q_opt+m-1 is multiplied by the factor a_qopt.

0006It is illustrative of the invention that in the case of speech which is sampled at a rate of 8000 Hz, the time interval corresponding to said subsequent sequence can be reduced to approximately 25 samples, <u style="single">i.e.</u> to approximately 3 msec.

0007An arrangement for performing the method according to the invention is characterized by    comparator means for comparing values of samples which constitute an environment for the relevant sequence of samples to be restored, the value constituting a test environment of the relevant sequence of samples to be restored, which values or samples constitute a test environment of an equally large sequence of valid samples which appeared a predetermined period of time earlier;    indicating means for indicating several groups of samples shifted in time relative to each other, each group forming one of said test environments;    control means controlling the mode of operation of said comparator means in dependence on the indication result of the indicator means;    selection means which on the basis of the results provided by the comparator means select that test environment which approximates the environment of the sequence of samples to be restored most closely; and    arithmetic means deriving the replacement values from the values of the samples which are bounded by the selected test environment.

0008More specifically, an arrangement for performing the method according to the invention, is characterized by an arithmetic unit arranged for always calculating for a plurality of chosen values q, wherein q = p, ..., t-p the quantities a_q and ε(q, aq) in accordance with the formulae<maths id="math0002" num=""><img file="EP0241980B1_D0002.tif" /></maths>    by a search unit arranged for selecting from the pairs of quantities a_q and ε(q, a_q) as generated by the arithmetic unit, that pair for which ε(q, a_q) is minimal and to determine the associated values for q and a_q > o as q_opt and a_opt; and    a multiplying circuit arrangement for multiplying each of the sample values in the position q_opt+i wherein i = 0, ..., m-1 by a_opt and to substitute the product values as replacement values for the values of the invalid samples.

0009An arrangement illustrative of the invention is further characterized in that it is of such a structure that the flow charts of Figs. 4 and 5 can be performed.

0010The invention and how it can be put into effect will now be further described by way of example with reference to the accompanying drawing, wherein: <ul id="ul0002" list-style="none"><li>Fig. 1 shows a time diagram with reference to which the principle of the present invention will be described;</li><li>Fig. 2 is a general block circuit diagram to illustrate the basic function performed within the scope of the present invention;</li><li>Fig. 3 is a block circuit diagram of an embodiment of an arrangement according to the invention;</li><li>Fig. 4 is a flow chart to illustrate an input/output program performed within the scope of the invention; and</li><li>Fig. 5 is a flowchart for further illustration of the restoration technique according to the invention.</li></ul>

0011The invention can advantageously be used for a sampled speech signal of which a number of samples is mutilated such by interferences that at the receiver side of an associated transmission system the samples are considered as invalid. Such a situation can, for example occur in a telephony system of the type known as MAT (mobile automatic telephony). Due to what is commonly referred to as selective fading, the signal received and consequently the speech signal received fails regularly and for a short period of time. It has been found that a speech signal modulated on a carrier of, for example, approximately 300 MHz is received undisturbed and disturbed at the receiving side with given intervals which depend on the speed at which a relevant vehicle functioning as a mobile station of such a telephony system travels.

0012The restoration technique performed by the invention makes it possible to restore reliably invalid samples of a speech signal sampled at a frequency of 8000 Hz, if the interval containing unknown samples is less than approximately 12 msec, and if the relevant speech signal is available undisturbed on both sides of such an interval, for approximately 3 msec.

0013Hereinafter, the principle of the present invention will be further illustrated with reference to Fig. 1. The diagram shown in Fig. 1 is illustrative of a sequence of consecutive samples s_k, wherein k = 0, ..., N-1. Of the samples the samples s_t, ..., s_t+m-1, FIG. 1 interval B₁, form a sequence of samples which are considered invalid and must be restored. In other words, the replacement values for estimators ŝ_t+i, wherein i = 0, ..., m-1 must be found for these invalid samples. These replacement values must approach the original values of these invalid samples to the best possible extent, or in any case be such that in the ultimately restored signal substantially no observable errors occur.

0014The invention is based on the principle that such replacement values can be derived from the values of that sequence of valid samples s_q+i whose environment approaches the environment of the sequence s_t+i to the best possible extent, FIG. 1 interval B2. In the foregoing, q represents a sample position which can be chosen from a further specified range, while for i it holds that i = o, ..., m-1. As is shown in Fig. 1, the sequence s_t+i of invalid samples is bounded between the pre-environment defined by the samples s_t-p and s_t-1, and a post-environment of the sequence of the sequence of invalid samples. The sample position index q will now be chosen within the limitations of <maths id="math0003" num=""><math display="inline"><mrow><mtext>(1)   p ≦ q ≦ t-p; and 2 p+m &lt; t.</mtext></mrow></math><img file="EP0241980B1_D0003.tif" /></maths> It is now possible to define for each sequence s_q+i wherein i = 0, ..., m-1, defined by a sample index position q also a pre-environment, FIG 1 interval A2, and a post-environment FIG 1 interval C2, each comprising a number of p samples. The respective pre-environment and post-environment can now be defined by s_q-p and s_q-1 and s_q+m and s_q+m+p-1, respectively.

0015The strategy in accordance with which the sequence of valid samples will ultimately be suitable to serve as a source for making available the replacement values of the samples which are considered invalid, implies that for a set of different q-values, the environment, more specifically the pre-environment and the post-environment of a sequence s_q+i, wherein i = 0, ..., m defined by the chosen q-value, is always compared with the environment, more specifically the pre-environment and the post-environment of the sequence s_t+i of unknown samples. That environment which, but for a scaling factor a_q, most closely approximates the environment of the sequence of invalid samples s_p+i then constitutes the boundary for the sequence s_q+i looked for. All this means in practice that the sampling values <maths id="math0004" num=""><math display="inline"><mrow><mtext>(2)   </mtext><msub><mrow><mover accent="true"><mrow><mtext>s</mtext></mrow><mo>ˆ</mo></mover></mrow><mrow><mtext>t+i</mtext></mrow></msub><msub><mrow><mtext> = a</mtext></mrow><mrow><mtext>q</mtext></mrow></msub><msub><mrow><mtext>s</mtext></mrow><mrow><mtext>q+i</mtext></mrow></msub></mrow></math><img file="EP0241980B1_D0004.tif" /></maths> wherein i = 0, ..., m-1, are the desired sources for the estimation values which can be filled up in the position of the invalid samples, if the environment, more particularly the pre-environment and post-environment, respectively, of these invalid samples is approached to the closest possible extent by the environment, more specifically the respective pre-environment and post-environment which can be represented by <maths id="math0005" num=""><math display="inline"><mrow><msub><mrow><mtext>a</mtext></mrow><mrow><mtext>q</mtext></mrow></msub><msub><mrow><mtext>s</mtext></mrow><mrow><mtext>q-p+i</mtext></mrow></msub></mrow></math><img file="EP0241980B1_D0005.tif" /></maths> , wherein i = 0, ..., p-1, and <maths id="math0006" num=""><math display="inline"><mrow><msub><mrow><mtext>a</mtext></mrow><mrow><mtext>q</mtext></mrow></msub><msub><mrow><mtext>s</mtext></mrow><mrow><mtext>q+m+i</mtext></mrow></msub></mrow></math><img file="EP0241980B1_D0006.tif" /></maths> , wherein i = 0, ..., p-1.

0016The scaling factor a_q depends on the value chosen for the sample position index q for a relevant equation. By introducing such a scaling factor a_q it can be accomplished that environments of equal shapes, which are compared with each other can also be made equal to each other as regards their magnitude.

0017The above-described strategy is based on the assumption that one set of samples is "similar" to another set of samples, if the environments of these two sets of samples may be considered as "similar". For the case of voiced speech, in other words in the case of a signal considered as periodical, this assumption is certainly correct. Because of this periodical character of voiced speech it is, in addition, substantially certain that in a finite interval preceding the sequence of samples to be restored, a sequence of valid samples is present whose values, after scaling, can serve as replacements for the samples considered as invalid. For the case of unvoiced speech, which has rather the character of noise, this is not always the case. Inspite thereof, it has been found that the restoration technique proposed leads to satisfactory results, also when a signal applied is unvoiced speech.

0018From the foregoing it will be obvious that looking for the sequence of valid samples s_q+i wherein i = 0, ..., m-1, which is suitable for supplying replacement values, which in practice means finding the correct values for the quantities p and q and a_q. As is also illustrated in Fig. 1, the samples suitable for use as the replacement values are looked for in a finite interval preceding the sequence of samples s_t+i, wherein i = 0, ..., m-1 to be restored. As will also be obvious from Fig. 1, the sequence s_k, wherein k = 0, ..., t+m+p-1, wherein k ≠ t, ..., t+m-1, is available for that purpose. Taking account of the limitations imposed by formula (1), it holds that <maths id="math0007" num=""><math display="inline"><mrow><mtext>t-p-m-p = t-2.p-m</mtext></mrow></math><img file="EP0241980B1_D0007.tif" /></maths> defines a stock of valid samples which are suitable for making replacement values available for the invalid samples. As will further be clear from formula (1), p must in any case be less than <maths id="math0008" num=""><math display="inline"><mrow><mtext>(t-m)/2</mtext></mrow></math><img file="EP0241980B1_D0008.tif" /></maths> . It was found experimentally that P ≧ 25 provides satisfactory results. With a customary sampling rate of 8000 Hz of a speech signal, this implies that the subsequent sequence produces a delay of 3.125 msec.

0019For each chosen value of the sample position index q wherein q = p, ..., t-p a value of a = a_q can be determined, as a result of which the expression<maths id="math0009" num=""><img file="EP0241980B1_D0009.tif" /></maths> minimizes. Thus, each environment, more specifically the preenvironment and post-environment, respectively, as described by <maths id="math0010" num=""><math display="inline"><mrow><msub><mrow><mtext>(4)   [(s</mtext></mrow><mrow><mtext>q-p+-i</mtext></mrow></msub><msub><mrow><mtext>)</mtext></mrow><mrow><mtext>i=0,....,p-1</mtext></mrow></msub><msub><mrow><mtext>, + (s</mtext></mrow><mrow><mtext>q+m+i</mtext></mrow></msub><msub><mrow><mtext>)</mtext></mrow><mrow><mtext>i=0, ...,p-1</mtext></mrow></msub><mtext>]</mtext></mrow></math><img file="EP0241980B1_D0010.tif" /></maths> q=p,...,t-p can be scaled such that the means square error as described by formula (3) relative to the environment, more specifically the pre-environment and post-environment, respectively, as defined by <maths id="math0011" num=""><math display="inline"><mrow><msub><mrow><mtext>(5)   [(s</mtext></mrow><mrow><mtext>t-p+i</mtext></mrow></msub><msub><mrow><mtext>)</mtext></mrow><mrow><mtext>i=0,...,p-1</mtext></mrow></msub><msub><mrow><mtext>, (s</mtext></mrow><mrow><mtext>t+m+i</mtext></mrow></msub><msub><mrow><mtext>)</mtext></mrow><mrow><mtext>i=0,...,.p-1</mtext></mrow></msub><mtext>]</mtext></mrow></math><img file="EP0241980B1_D0011.tif" /></maths> is minimal. The scaling factor a_q suitable for use as a selected sample position index q is that value for a for which the first derivative of the expression defined by (3) becomes equal to zero. From the equation obtained thus a_q can be dissolved in accordance with<maths id="math0012" num=""><img file="EP0241980B1_D0012.tif" /></maths> The expression ε(q, a_q) can be written as<maths id="math0013" num=""><img file="EP0241980B1_D0013.tif" /></maths> The correlation procedure proposed by this equation (7) is schematically illustrated in Fig. 1 in that the environments (test environments) in this Figure, defined by a given sample position index q are shown below the environment bounding the sequence of samples to be restored and with which the said first environments are compared.

0020The optimum value for the sample position index q, q = q_opt and the optimum value for the scaling factor a_q, a_q = a_opt is now that value of q and a_q for which the result of the equation (correlation) represented by equation (7) is minimal.

0021A method which is illustrative of the invention comprises the following steps:    For all the available values for the sample position index q, wherein q = p, ..., t-p, the value a_q as defined in equation (6) and the correlation result in accordance with equation (7) are each time determined. From all the pairs [a_q, ε(q, a_q)] _q=p,...,t-p thus determined that pair is determined for which ε(q, a_q) is minimal. The associated values for q and a_q are the values q_opt and a_opt. Thereafter, on the basis of the q_opt and a_opt thus determined, replacement values can be determined in accordance with<maths id="math0014" num=""><img file="EP0241980B1_D0014.tif" /></maths> It has been ascertained that a_opt must be positive. This implies that the pairs [a_q, ε(q, a_q)], wherein a_q &lt; 0 must be left out of consideration when looking for the minimum for ε(q, a_q).

0022Fig. 2 shows a general circuit diagram of a configuration of components dissolved in performing basic functions of the method according to the invention.

0023An equidistantly sampled signal, for example a speech signal is applied in the shape of a sequence of samples s_k. A flag bit which indicates whether a relevant sample must be considered as valid or invalid, must be added to each of the applied samples. The samples are applied to a flat bit-recognizing means 2.1 which at an output generates a flag bit signal f_k which informs a control arrangement 2.2 of the fact whether an applied sample is be qualified as being valid or invalid. <u style="single">Via</u> a different output of this flag bit-recognizing means the incoming signal samples s_k are applied to a randomly accessible register 2.3. This register has such as storage capacity that all the incoming samples required to complete a restoration cycle relating to a sequence of samples qualified as invalid, are available. The control arrangement 2.2 is <u style="single">inter alia</u> arranged for determining in accordance with a predetermined routine, on the basis of flag bit signals applied and samples present in the register 2.3, whether a restoration procedure for restoring a number of samples which are considered invalid must be initiated and completed or not. As soon as this control arrangement has ascertained that of a sequence of samples stored in the register 2.3 a set must be restored, all the sampled positions q, wherein q = p, ..., t-p, illegible therefor are read from the register 2.3, all this in such a manner as defined with reference to Fig. 1, the sample values corresponding therewith being inputted in an arithmetic unit 2.4. Fundamentally, such an arithmetic unit performs the function whereby an environment of a relevant sequence of samples to be stored is compared to an environment of an equally large sequence of valid samples which appear in a chosen period of time earlier, such a comparison being made for a plurality of different periods of time chosen thus. More specifically, the arithmetic unit 2.4 is arranged to always calculate for a selected value q the quantities a_q and ε(q, a_q) in accordance with the formula (6) and (7), defined in the foregoing, from the sample values read from the register 2.3. The results of the operation performed by the arithmetic unit 2.4, more specifically the pairs a_q and ε(q, a_q) generated for the chosen q-values, are applied to a search unit 2.5. Fundamentally, this search unit performs the function whereby the environment which most closely approximates the environment of the sequence of samples to be restored is selected. More specifically, this search unit 2.5 is arranged for finding from the results applied from the arithmetic means 2.4 to this search unit, that combination for which the relevant quantity ε(q, a_q) wherein a_q &lt; 0 is minimal. The values associated therewith for q and a_q, namely q_opt and a_opt are supplied by this search unit 2.5 on the basis of the search result found, and applied to a multiplier 2.6. This multiplier is also arranged to read under the control of the control arrangement 2.2 the sample position q_opt, ... q_opt+m-1 determined by the quantity q_opt supplied, and consequently to insert automatically the corresponding sample values. This multiplier is then operative for multiplying each of the sample values thus inserted, by the quantity a_opt supplied by the search unit 2.5, so that the ultimate replacement values can be filled up in the sample position t, ..., t+m-1. Thus, this multiplier performs basically the function whereby the valid samples which are bounded by the environment chosen by the search unit, are used to form the replacement values of the samples to be restored.

0024After such a restoration cycle the restored sequence of samples is supplied from the output of the register 2.3, also under the control of the control unit 2.2.

0025Generally the invention is suitable for use in those situations in which digitized signals, such as audio signals and more particularly speech signals, can be received in a disturbed state. The positions of the disturbed sample must be known, as also the position of the samples which determine environments. The invention is also suitable for use for disturbed analog speech signals, after the signals have been converted to the discrete form.

0026The invention is also suitable for use in restoring speech signals which during transmission were protected by an error-correcting and detection code. In that case those samples can be restored which can indeed be detected as an error, but can not be corrected anymore.

0027Fig. 3 shows a block circuit diagram of an embodiment of an arrangement according to the invention. The block 3.1 is a receiver which at its output can develop from a received signal a sequence of samples s_k such as they were described with reference to Fig. 2. The block 3.2 corresponds to the flag bit-recognizing circuit shown in Fig. 2. The values of the incoming sample appear at the output 3.3 in the form of 16-bit words. When at the output 3.3 a sample appears of a value which can be qualified as invalid, a flag bit signal formed by a 1-bit word and indicating that the relevant sample is invalid appears at the output 3.4. The values of the samples and the flag bit signals are applied to the respective input arrangement 3.6 and 3.7 of a microprocessor configuration 3.5 arranged for implementating a restoration cycle for sample to be restored. In addition to an outputting member 3.8 <u style="single">via</u> which the sample values are outputted, this processor 3.5 comprises a central processing unit 3.9 of which a control member, an arithmetic member and a logic unit form part. Furthermore, the processor 3.5 includes three memories 3.10, 3.11 and 3.12. The memories 3.10 and 3.11 are random access memories, the memory 3.10 acting as a cyclic buffer and the memory 3.11 acting as a working store for storing the intermediate results. The memory 3.12 is a static memory in which the programs for restoring invalid samples are stored. The memories 3.10, 3.11 and 3.12 are coupled to the central processing unit 3.9 <u style="single">via</u> a data bus 3.13, <u style="single">via</u> which bus data can be inputted and outputted. For conveying addresses, the memories 3.10, 3.11 and 3.12 and also the inputting and outputting arrangements 3.6, 3.7 and 3.8 are coupled to the central processing unit 3.9 <u style="single">via</u> an address bus 3.14. Furthermore, the processor 3.5 includes a clock 3.15 under whose control the operation of the central processing unit can be interrupted (interrupt) in order to have the inputting and outputting of data take place. Namely, digital audio signals require a constant through-put in order to accomplish that the frequency with which the samples appear at the output is the same as the rate at which the relevant signal is sampled. Inputting and outputting of data is controlled <u style="single">via</u> a program stored in the program store 3.12. Fig. 4 illustrates a flow-chart of this program.

0028This program can be described as follows: (for captions see Table 1).

0029Block 4.1, caption: "wait for interrupt"; description: data can be inputted and outputted during an interrupt which is initiated when a clock pulse from data clock 3.5 appears.

0030Block 4.2: caption "output/input"; description: upon the appearance of a clock pulse the value of a new sample is inputted <u style="single">via</u> the input 3.6 and entered into the first free address in the memory 3.10, and the oldest present in memory 3.4 is read and outputted.

0031Block 4.3, caption: "input error?"; description: upon entering a new sample, the flag bit signal is at the same time entered <u style="single">via</u> input 3.7. When no invalid sample is detected, the subsequent clock pulse is waited for repeating the procedure.

0032Block 4.4, caption: "update error table"; description: if a sample is detected a being invalid, the address of this sample is entered into the memory 3.10, in a Table of invalid samples in the working memory 3.11.

0033Under the regime of the above-described program, the cyclic buffer store 3.10 operates as a delay line, whose delay is determined by the period of time necessary for a restoration cycle measured from the arrival of the first sample, which is qualified as invalid, increased by the period of time of the required post-environment. The program stored in the memory 3.12, with which the replacement values for the samples to be corrected are calculated, is started when the input/output program detects sample as being invalid.

0034This restoration program will now be described in greater detail with reference to Fig. 5, which shows a flow chart with this program. (For captions see Table II). <dl id="dl0001"><dt>Block 51:</dt><dd>"wait for an error pattern that can be restored", description: as long as no pattern if invalid samples is detected the restoration program is inoperative.</dd><dt>Block 52:</dt><dd>caption "compute q_opt and a_opt"; from a row of possible values of q = p, ..., t-p a value of a = a_q is determined for which the expression<maths id="math0015" num=""><img file="EP0241980B1_D0015.tif" /></maths> is minimised. The scaling factor a_q which is suitable for use for a selected sample position index q is that value of a for which the first derivative of the expression defined by (3) becomes equal to zero. From the equation of a_q obtained therefrom there are dissolved in accordance with<maths id="math0016" num=""><img file="EP0241980B1_D0016.tif" /></maths> The expression ε(q, a_q) can be written as<maths id="math0017" num=""><img file="EP0241980B1_D0017.tif" /></maths></dd></dl> The correlation procedure proposed by this equation (7) is schematical illustrated in Fig. 1 in that the environment defined by a given sample position index q (test environments) are shown in this Figure below the environments bounding the sequence of samples to be restored and with which said first environments are compared.

0035The optimum value for the sample position index q, q = q_opt and the optimum value for the scaling factor a_q, a_q = a_opt is now that value of q and a_q for which the result of the equation (correlation) represented by equation (7) is minimal. It was found that a_opt must be positive. This implies, that the pairs [a_q, ε(q, a_q)] wherein a_q &lt; 0 must be ignored in the search for the minimum for ε(q, a_q). <dl id="dl0002"><dt>Block 53:</dt><dd>caption "compute replacement samples ŝ_t+i, then the sample ŝ_t+i is determined in accordance with<maths id="math0018" num=""><img file="EP0241980B1_D0018.tif" /></maths></dd><dt>Block 54:</dt><dd>caption "replicate buffer", here the values of the invalid samples are replaced by the estimated values s_t+i, I = 0, ..., M-1. Thereafter the program is again cycled through.</dd></dl><tables id="tabl0001" num="0001"><img file="EP0241980B1_D0019.tif" /></tables>