NL9002308A - METHOD FOR CODING AND DECODING A SAMPLED ANALOGUE SIGNAL WITH A REPEATING CHARACTER AND AN APPARATUS FOR CODING AND DECODING ACCORDING TO THIS METHOD - Google Patents

Abstract

Frequency components are calculated from the STP-filtered speech signal. The amplitudes of these are combined in a manner such that the resultant values are associated with frequencies which are situated equidistantly on a linear Bark scale. Said components are quantised, possibly after scaling. In the decoder, the components are again distributed over the frequency spectrum. In the coder, the fundamental regularity D is determined with an LTP technique, after which it is transmitted. In the decoder the phases of the reconstructed signal at the spacing D in the past are determined. These phases are combined with the amplitudes already present in the frequency spectrum, after which transformation back to the time domain takes place. Inverse STP filtering is then carried out.

Description

Title: Method for encoding and decoding a sampled analog signal of a repetitive nature and an apparatus for encoding and decoding according to this method.

The invention relates to a method of encoding a sampled analog signal of a repetitive nature, wherein the sampled signal is split into successive segments each having a predetermined number of samples; in which a short-term prediction analysis is performed on these segments and in which the coefficients determined in this analysis are transferred and are also fed to a short-term prediction filter, in which a long-term prediction analysis is carried out on the residual signal available at the output of this filter and the information determined in this analysis is also transmitted, and the information contained in the residual signal is encoded and transmitted.

The invention also relates to a method for decoding a signal encoded in the manner described above, wherein the received information from the long-term prediction analysis and the received other information from the residual signal are combined and the combined signal together with the received coefficients of the short term prediction analysis is fed to an inverse short term prediction filter from which a series of samples is output that reconstructs the sampled analog signal.

The invention also relates to an apparatus for encoding and decoding according to the above-described methods.

It is known that analog signals of a highly coherent nature, such as, for example, speech signals, can be efficiently encoded after sampling by successively performing a number of different transformations on successive segments of the signal, each having a specific duration. One of the known transformations is linear predictive coding (LPC), for an explanation of which reference can be made to the book "Digital Processing of Speech Signals" by L.R. Rabiner and R.W. Schafer; Prentice Hall, New Jersey; Chapter 8.

As mentioned, LPC is always used for a signal segment with a specific duration, for example 20 ms for speech signals, and is counted as short-term prediction coding (STP). It is also known to use long-term prediction (LTP) in addition to a short-term prediction, whereby highly efficient coding is obtained by combining these two techniques. The principle of LTP is described in Frequenz, volume 42, no. 2-3, 1988; pp. 85-93; P. Vary et al .: "Sprachcodec für dass Europaische Funkfernsprechnetz" while an improved version of the LTP principle is described in Dutch patent application 9001985.

The object of the invention is to provide a method for transmitting the information relevant to the human ear in the residual signal remaining after the application of the STP principle in a very efficient manner, ie with a small number of bits / sec, without compromising the quality of the speech reconstructed by the decoder on the receiving side as perceived by the listener.

The encoding method according to the invention is to this end characterized in that the residual signal is transformed into the frequency domain, in that the amplitudes of at least some of the frequency components obtained during the transformation into the frequency domain are combined in such a way that the frequencies associated with the combined amplitudes are equidistant on a linear Bark scale and in that a signal representative of these combined amplitudes is transmitted.

The method for decoding according to the invention is characterized in that the original amplitudes in the frequency domain are reconstructed from the received amplitude values, that the information transmitted as a result of the long-term prediction analysis is used to calculate the phase values associated with these amplitudes and that the calculated phase values together with the associated amplitudes are transformed into the time domain.

In accordance with the present invention, the residual signal is encoded in one sense, which means that only that information is relevant to differences perceived by the human ear in the decoded received signal.

Firstly, use is made of the known fact that the human ear is not sensitive to absolute phase values, but only to phase relations, so that in principle it is not necessary to transfer the phase information from the residual signal to be encoded, as long as it is possible to reconstruct the original phase relationships on the receiving side.

In addition, the present invention makes use of the insight that has existed for some time that human hearing actually functions as a sequence of a number of filters with contiguous frequency bands, but with different bandwidths, the so-called critical bands or Barks, where the bandwidth of such critical bands for low frequencies are much smaller than for high frequencies. A frequency scale formed according to this understanding is called a linear Bark scale. For a further explanation of the principle of the Bark scale, see B.Scharf and S.Bus "Stimulus, Physiology, Thresholds" in L.Kaufman, KRBoff and JPThomas, editors, Handbook of Perception and Human Performance, chapter 14 , biz. 1-43, Wiley, New York, 1986.

It should also be noted that the principle of first transforming a residual signal to be transmitted in speech coding to the frequency domain and then transmitting the information available after this transformation has already been proposed per se. For this, reference may be made, for example, to the article "Fourier Transform Vector Quantization for Speech Coding" by P. Chang et al in IEEE Transactions on Communications, Vol. COM 35, No. 10, biz.

1059 - 1068. According to this publication, however, vector quantization is used after the transformation, and there is no question of transferring only amplitude information.

The invention will be further elucidated hereinbelow on the basis of an exemplary embodiment with reference to the drawing, which shows:

Figure 1a: A block diagram of an exemplary embodiment of an encoding unit for the device according to the invention.

Figure 1b: A block diagram of an embodiment of a decoding unit for the device according to the invention.

An analog signal output from a microphone 1 is bandwidthed by a low-pass filter 2 and converted in an analog-to-digital converter 3 into a series of amplitude and time-discrete samples representative of the analog signal. The output of the converter 3 is applied to the input of a short-term analysis unit 4 and to the input of a short-term prediction filter 5. These two units provide the above-mentioned short-term prediction (STP) on segments of, for example, 160 samples and the analysis unit 4 provides an output in the form of short-term prediction filter coefficients, which are quantized, coded and transmitted to the decoder shown in Figure 1b. The structure and function of the filter 5 and the unit 4 are well known to those skilled in the art of speech coding and are of no further importance to the essence of the present invention, so that no further explanation can be given.

The STP filtered signal is fed to a long-term prediction (LTP) analysis unit 6. In this analysis unit, an LTP analysis is applied twice per segment of 160 samples, in a manner as is described, for example, in Dutch patent application 9001985 In such an LTP analysis, according to a certain search strategy, a signal segment to be encoded is always searched for the best possible segment in a signal period with a predetermined duration preceding that segment and a signal representative of the number of samples D becomes located between the start time of the found segment and the start time of the segment to be encoded, transmitted in encoded form.

The output signal of the STP filter unit 5 is called the residual signal and this residual signal is transmitted in coded form according to the invention in such a way that only the information which is perceptually relevant is transmitted. For this purpose, the segments of 160 samples of this residual signal in the chain 7 are divided into eight sub-segments of 30 samples. This is done by first dividing the offered segment into eight sub-segments of 20 samples and then supplementing them at the front with the last ten samples of the previous sub-segment. This means that the last ten samples of each segment must be saved in order to be able to supplement the first sub-segment of the next segment. Each sub-segment of 30 samples in a chain 8 is then multiplied by a window function, such as, for example, a cosine function. The window function is chosen such that for each sample in the overlapping parts of the sub-segments the sum of the squares of the two multiplication factors is one. The reason that this should apply to the squares is that the multiplication with the window function takes place both in the encoder and in the decoder shown in Figure 1b. A discrete Fourier Transform (DFT) is performed on the windowed sub-segments in a chain 9, whereby 16 different frequency components are obtained for each sub-segment. Of these 16 frequency components, numbered 0 to 15, the amplitudes A of components 1 to 13 are calculated in a chain 10. Components 0, 14 and 15 can be disregarded because they lie outside the frequency band of 300 - 3400 Hz chosen for voice communication. If a larger or a smaller frequency band is relevant, the number of amplitude components considered can be adapted accordingly. Starting from the said 13 components, four so-called Bark amplitude components are calculated in a chain 11. These are amplitudes associated with frequencies equidistant on a linear Bark scale. For example, the Bark amplitude components Bi through B4 can be calculated from the DFT amplitudes Αχ through A13 as follows:

(1)

If desired, an amplification factor G in scale 12 is calculated from the four Bark amplitude components as scale value according to:

The use of the scale value G has the advantage that the coding of the scaled amplitudes can be done in a more efficient manner. The value of G is quantized in a chain 13 and then transmitted to the decoder. If the scale factor G is calculated, each Bark component in a chain 14 is divided by the quantized gain factor G. The result of this division is quantized in a chain 15, coded and then also transferred to the decoder.

If no scale value is used, the chains 12, 13 and 14 can be omitted and the four calculated values for the Bark amplitude components after quantization in chain 15 can be directly transferred.

In the decoder, the four scaled Bark amplitude components after decoding in a chain 16 in a multiplier 18 are multiplied by the amplification decoded in a chain 17, ê, thereby obtaining the reconstructed Bark amplitude components B1 through B4. Obviously, this does not apply if no scaling factor is used in the encoder. In a chain 19, the amplitudes in the frequency domain Αχ to Ai3 (equidistant on the Hz scale) are calculated using the following formulas.

In order to be able to transform the 13 frequency components in the IDFT chain considered at the coder back into the time domain via an inverse DFT (IDFT), the amplitudes and the phases are required.

The phases are determined in the following manner using the LTP information decoded in a chain 23, which consists of the sample distance D.

The 120 most recent samples of the reconstructed STP residue, as present at the exit of the chain 22 to be discussed below, are always kept. In a chain 24, the sub-segment, which is situated at a distance from D samples in the past relative to the current sub-segment, is determined and in a chain 25, this sub-segment is multiplied by the same window function as that in the chain 8 in the encoder. applied. Subsequently, a DFT is applied to this sub-segment in a chain 26, after which the phases of the 13 components considered can be calculated in a chain 27. Using the phases determined in this way and the amplitudes already calculated, an IDFT is converted into the chain 20 where the amplitudes of Ao, & L4, Αχ5 and Αχς are set to zero.

A reconstruction of the sub-segment, 30 samples long, is now available at the output of circuit 20, but this has still been modified by the window function performed in the encoder. Therefore, the reconstructed sub-segment in a chain 21 is multiplied again with the window function. In the first ten samples of the sub-segment now multiplied twice by the window function, the last ten samples of the previous sub-segment multiplied twice by the window function are stored for this purpose in a chain 22. As a result, in the resulting ten samples, the sum of the multiplication factors equals one.

The last ten samples in this sub-segment are stored. The first twenty samples are part of the reconstruction of a segment of the STP residue. After reconstructing and pooling eight sub-segments, a fully reconstructed segment of the STP residue is obtained, which is ten samples in the past relative to the segment on which the STP analysis was performed in the encoder.

An inverse STP filtering is carried out on this segment in a filter chain 28 in a manner known per se using the received STP coefficients, the filter coefficients from the previous segment being used for the first ten samples.

The output signal of the filter 28 is converted in a digital-analog converter 29 into an analog signal, which is fed via a low-pass filter 30 to a loudspeaker 31, which gives a true-to-life reproduction of the speech signal presented to the microphone 1, which thanks to the measures according to the invention could be transmitted in a coded form with a small number of bits.

If desired, a circuit 23 'can be included between circuits 23 and 24 in order to first subject the value of D received by the decoder to a number of operations in order to obtain a value of D which is optimal for the reconstruction of the speech signal. are three consecutive operations.

1) If the row of received values D shows a trend, the current received D, if it falls outside this trend by a certain margin, is replaced by a value corresponding to this trend. Algorithms for determining a trend in a series of consecutive values and for determining a substitute value for a signal outside this trend are well known per se to those skilled in the art.

2) Three intermediate values (Iif I2 and I3) are calculated by means of interpolation between two consecutive values of D (Di and D2), possibly adjusted by means of such an algorithm. This is done, for example, in the following manner: I] _ = 0.75 * D] _ + 0.25 * D2 12 = 0.5 * D] _ + 0.5 * D2 13 = 0.25 * D] _ + 0.75 * D2

The interpolation occurs because the distance D in the encoder is determined twice per segment. Without interpolation, decoding of four consecutive sub-segments with the same value of D would be done. If there is no fundamental regularity in the signal in the encoder, this would incorrectly provide regularity in the decoder for four sub-segments. This problem is addressed by interpolation.

If fundamental regularity is present in the speech signal, the repeat distance in the signal will generally vary slowly. Thanks to the interpolation, the variation in the value of D now also has a smooth character with the decoder.

3) After equalizing the values of D by calculating a replacement value, if necessary, and after interpolation, the calculated distance D corresponds as closely as possible to the repeat distance actually present in the signal. However, if this distance D is less than 30, D is multiplied by an integer, chosen such that the result is at least equal to 30. This is necessary because of a sub-segment at a distance less than 30 from the current segment not all samples have been reconstructed yet, so that they cannot be used for calculating the phases.

The reason that distances D less than 30 are nevertheless transmitted is that in this way it is prevented that, if the fundamental regularity in the signal covers a number of samples smaller than 30, the decoded distance D takes values which are unequal multiples of the actual repeat distance. This would make the equalization algorithm less likely to detect a tendency.

Claims (9)

A method of encoding a sampled analog signal of a repetitive nature, wherein the sampled signal is split into successive segments each having a predetermined number of samples; in which a short-term prediction analysis is performed on these segments and in which the coefficients determined in this analysis are transferred and are also fed to a short-term prediction filter, in which a long-term prediction analysis is carried out on the residual signal available at the output of this filter and the information determined in this analysis is also transmitted, and in which the information contained in the residual signal is encoded and transmitted, characterized in that the residual signal is transformed into the frequency domain, in that the amplitudes of at least some of the ones in the transformation into the frequency domain obtained frequency components are combined in such a way that the frequencies associated with the combined amplitudes are equidistant on a linear Bark scale and in that a signal representative of these combined amplitudes is transmitted. A method of decoding a signal encoded with the method of claim 1, wherein the received long-term prediction analysis information and the received other information from the residual signal are combined and the combined signal together with the received short-term coefficients prediction analysis is applied to an inverse short-term prediction filter from which a series of samples is output which is representative of the sampled analog signal, characterized in that the original amplitudes in the frequency domain are reconstructed from the received amplitude values, which the long-term prediction analysis transferred information is used to calculate the phase values associated with these amplitudes and that the calculated phase values together with the associated amplitudes are transformed into the time domain. Method according to claim 1, characterized in that the amplitudes of thirteen frequency components Αχ to A13 obtained during the transformation to the frequency domain are transformed into amplitudes Βχ to B4 of four equidistant frequency components located on a Bark scale according to:

and that these values for B are transferred after quantization. Method according to claim 3, characterized in that for the four equidistant frequency components Βχ to B4 located on a Bark scale, a scale factor G is calculated according to:

that this value for G is quantized and that the values from Βχ to B4, before being quantized, are divided by the quantized scale factor. Method according to claims 2 and 3 or 4, characterized in that combined amplitude values B1 to B4 are reconstructed from the received information, and amplitude values A1 to A13 in the frequency domain are obtained therefrom according to:

and that the information transferred as a result of the long-term prediction analysis is representative of the number of samples D which lie between the start time of a previously transferred group of samples found by means of the long-term prediction analysis and the start time of a group to be decoded. monsters. Method according to claim 5, characterized in that the previously transferred group of samples, which is located at a distance D with respect to a group of samples to be decoded, is transformed into the frequency domain, that of at least some of the ones with these frequency components calculated the phase value is determined, that these phase values are combined with the amplitude values Ai to Ai3 and that these combinations are transformed to the time domain. Method according to claim 5 or 6, characterized in that the drift in the received values of D is equalized according to a predetermined algorithm by calculating, if necessary, a substitute value for a received value of D and then between two successive values of D by means of interpolation three intermediate values for D are calculated. Method according to claim 7, characterized in that three intermediate values Iir I2 and I3 are calculated from the known values Di and D2 according to: 11 = 0.75 * Di + 0.25 * D2 12 = 0.5 * Di + 0.5 * D2 13 = 0.25 * Di + 0.75 * D2 Device for carrying out the method according to at least one of claims 1 to 8.