JPH10260699A - Voice coding method and apparatus - Google Patents

Abstract

(57) [Summary] (with correction) [PROBLEMS] To provide a speech encoding method and apparatus capable of obtaining a high compression gain. A method for encoding a speech electrical signal using backward adaptive prediction. A first time frame of the audio electrical signal to be encoded is received and transformed to the frequency domain using a modified discrete cosine transform (MDCT). The resulting frequency spectrum has 1024 spectral components. Subsequent time frames of the audio electrical signal are then received and a discrete cosine transform is applied to each of them in turn to generate a stream of spectral data values for each spectral component. For each stream, a set of prediction coefficients is calculated for each spectral value using a predetermined number of previously received consecutive spectral values of the stream. Using a set of linear prediction coefficients, a predicted spectral value is generated, and
Calculate the error between the predicted spectral value and the corresponding actual spectral value.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to a method for encoding and decoding an electronic signal and an apparatus for performing such a method.

[0002]

BACKGROUND OF THE INVENTION It is well known that the transmission of data in digital form increases the signal-to-noise ratio and increases the information capacity along the transmission channel. However,
There remains a need to further increase channel capacity by significantly compressing digital signals. For audio signals, two basic compression principles are customarily applied. The first of these involves removing statistical or deterministic redundancy in the source signal, while the second suppresses or removes components from the source signal that are redundant with respect to human perception. Accompanies you. Recently, the latter principle has become dominant in high quality audio applications,
Typically, this involves separating the audio signal into multiple frequency components (often referred to as 'subbands'), each of which is analyzed to identify inadequacies (for the listener). The quantization is performed with the quantization precision determined to be removed. ISO (Internationa
l Standards Organization)
MPEG (Moving Pictures Expe
The rt Group speech coding standard and other speech coding standards have further defined using this principle. However, MPEG (and other standards) also uses a technique known as 'adaptive prediction' to further reduce data rates.

[0003] A particular form of adaptive prediction is known as 'backward adaptive lattice prediction'. 'I by Fuchs et al.
mproving MPEG Audio Codin
gby Backward Adaptive Lin
ear Stereo Prediction ', AE
S Convention, New York, Pre
print 4086 Oct. 1995 describes one such backward adaptive grid prediction algorithm. For each spectral value ('current' value) of each frequency component, backward adaptive lattice prediction encodes the previously calculated spectral value of that component (by intermediate computation of the quantized spectral values). Generate a set of prediction coefficients in the vessel. These coefficients are then used to predict the value of the current spectral value. An error between the current spectral value and the predicted spectral value is determined, and the error value is transmitted (after quantization) to a receiver. At any given time, the current prediction factor is
It will be seen that it can be effectively obtained from all previously received sample values. At the receiver, the coefficients are similarly calculated and the reconstructed spectral values are obtained by combining the predicted spectral values with the received error values.

[0004] In some algorithms using backward adaptive prediction, the measurement of the achieved compression value is often determined during the compression process, and the error value is increased if a positive compression gain is achieved. Only sent. Otherwise, the actual quantized frequency component signal is transmitted instead.

[0005]

SUMMARY OF THE INVENTION New MPEG-2A
The AC standard uses psychoacoustic modeling with 1024 frequency components and backward adaptive linear prediction. New M
It is anticipated that the PEG-4VM standard will have similar requirements. However, such a large number of frequency components introduces a large computational overhead due to the complexity of the prediction algorithm, and a large area of memory can be used to store the calculated coefficients. is necessary. In addition to this, in backward adaptive lattice prediction, when the predictor is turned 'off' (eg, when it is not possible to gain the benefit of compression by sending error values), the decoder will have any When required, without any temporary degradation of performance,
The coefficients must be determined so that the predictor can be turned 'on' again. This causes additional computational overhead.

[0006] It is an object of the present invention to overcome or at least reduce one or more of the above disadvantages.

This object operates on a relatively large number of frequency components of the audio signal to be encoded and calculates a prediction coefficient for that component from a predetermined number of previously received sample values of the component. This is achieved by utilizing a backward adaptive prediction algorithm.

[0008]

According to a first aspect of the present invention, there is provided an encoding method for encoding an audio electric signal using backward adaptive prediction, comprising: (a) an audio electric signal to be encoded; Receiving a first time frame of
(B) transforming the time frame into the frequency domain to generate a frequency spectrum having 512 or more spectral components; and (c) receiving the next time frame of the audio electrical signal and performing Repeating steps (b) for these frames in order to generate a stream of spectral data values; and (d) for each said stream, a predetermined number of previously determined reconstructed spectral values of a stream. Calculating a set of predictive coefficients for each spectral value using the covariance of the set, and using the set of predictive coefficients to generate a predicted spectral value; Calculating an error between the spectral values of the spectrum values. Providing an encoded representation of the over-time,
A method is provided in which the error can be recombined with the predicted spectral values to obtain a reconstructed spectral value.

In the case of a conventional backward adaptive prediction algorithm, the method of the present invention does not directly calculate a set of prediction coefficients from all previous spectral components. That is,
The prediction coefficients are recalculated for each spectral value and are not simply adapted from a previously calculated set. In this way, it is not necessary for the decoder to keep updating the coefficients while the predictor is off.

A backward adaptive prediction algorithm that calculates prediction coefficients from a predetermined number of covariances of previous spectral values encodes a speech signal subdivided into a relatively small number of frequency subbands (eg, 32). While such algorithms are not generally suitable for such applications, such algorithms require that the audio signal be composed of a relatively large number of frequency subbands (eg, MPEG
-4 has been found to be appropriate if it is subdivided into 1024) as specified in the draft standard. this is,
When a large number of subbands are defined, the order of the prediction algorithm (ie, the number of prediction coefficients) can be reduced, and the algorithm embodying the present invention provides high performance and computational efficiency for low orders Is good. Preferably, the prediction order is one or two. More preferably, the prediction order is two.

[0011] Preferably, said predetermined number of previously received consecutive spectral values is used to obtain a corresponding number of quantized spectral values. Therefore, it is the quantized value used to calculate the prediction coefficient.

Preferably, the time window viewed from the audio signal is:
overlapping. For example, each window may include 2048 sample points with adjacent windows having 50% overlap. However, the windows may be adjacent.

In one embodiment of the invention, a new set of prediction coefficients can be calculated for each and every spectral value. However, in other embodiments, it is possible to recalculate the prediction coefficients only for every second or third (or other multiple) spectral value, and use the same coefficient for several consecutive spectral values. , May be more computationally efficient. As soon as a transition of the audio signal is detected, a low coefficient update rate (eg, every second value) and a high coefficient update rate (eg, every spectral value)
It may be appropriate to switch between.

The lower bound for the predetermined number of previously received sample points used to calculate each set of prediction coefficients is determined by the required coding quality. However, preferably, the number is four or more. The upper limit for this number is determined by memory and computational constraints. Preferably, the number is 10 or less. More preferably, the predetermined number is six.

[0015] Any suitable method of evaluating the prediction coefficients,
For example, an autocorrelation method can be used. However, the least squares method has been found to be particularly advantageous.

Preferably, the prediction coefficients used to calculate the predicted spectral values are linear prediction coefficients.

It will be appreciated that the present invention is intended for use with psychoacoustic correction, and that the quantization of the error signal can be controlled accordingly.

According to a second aspect of the present invention, there is provided a method of decoding an audio electric signal encoded using the method of the first aspect, comprising the steps of: Receiving, as an input signal, an error value for each stream, and for each stream, using a set of prediction coefficients to determine a corresponding predicted spectral component value for each error value, wherein the prediction coefficient is Combining the error value and the predicted spectral value to provide a reconstructed spectral value calculated using the covariance of a predetermined number of previously determined continuous predicted spectral components And by combining the reconstructed spectral values of all streams and performing a frequency-to-time conversion,
Substantially reconstructing the audio signal.

It will be appreciated that the details of a particular embodiment of the encoding method determine most of the details (eg, the prediction order) of the embodiment of the decoding method.

According to a third aspect of the present invention, there is provided an apparatus for encoding a speech electrical signal using backward adaptive prediction, comprising: an input for receiving a speech electrical signal to be encoded; A time-frequency domain converter for transforming a received time frame of the signal from the time domain to the frequency domain to provide a frequency spectrum having 512 or more spectral components, and receiving the associated spectral values as a stream. Calculating a set of prediction coefficients for each spectral value using a covariance of a predetermined number of previously reconstructed spectral values, and calculating the set of prediction coefficients to generate a predicted spectral value. And using signal processing means associated with each spectral component to calculate the error between the predicted value and the corresponding actual spectral value. An apparatus is provided wherein the error provided provides an encoded representation of a stream of received spectral values, said error being able to recombine with the predicted spectral values to obtain a reconstructed spectral value. .

According to a fourth aspect of the present invention, there is provided an apparatus for decoding an audio electric signal encoded using the apparatus according to the third aspect of the present invention, comprising: Separating the set of values into separate spectral component streams, and an input for receiving a set of error values corresponding to
Signal processing means for determining a corresponding predicted spectral component value using a set of prediction coefficients for each error value, wherein the signal processing means comprises a predetermined number of previously determined consecutively determined spectral component values. A signal processing means arranged to calculate the prediction coefficients using the covariance of the reconstructed spectral values, wherein the signal processing means combines each error value with a corresponding predicted spectral value and reconstructs the reconstructed spectral value. Are provided.

According to a fifth aspect of the present invention, there is provided a communication system comprising a combination of the devices of the third and fourth aspects of the present invention.

According to a sixth aspect of the invention, there is provided a mobile communication device comprising a combination of the devices according to the third and fourth aspects of the invention.

For a better understanding of the present invention,
To illustrate how the invention may be carried out, reference is made to the accompanying drawings as an example.

[0025]

DETAILED DESCRIPTION OF THE INVENTION Referring to FIG. 1, a pulse code modulated (PCM) speech input signal g to be encoded
(T) is given to the input of the first signal processing unit 1 of the coding device. This first unit 1 converts the input signal g (t) from the time domain to the frequency domain in frame units where each frame n consists of 2048 sample values and adjacent frames have 50% overlap. Arranged for. In particular, the unit for transforming the signal into the frequency domain, such that the output of unit 1 consists of 1024 separate streams of spectral values x _j (n), each stream j corresponding to a different spectral component, 1 uses a modified discrete cosine transform (MDCT). It should be noted that other transformation methods can be used, for example a Fourier transform.

Each stream of data values x _j (n) is provided to a corresponding input of a backward adaptive predictor 2 whose operation is described in detail below. In brief,
For each spectral value x _j (n) of each stream, predictor 2 uses in turn the reconstructed quantized spectral value obtained later and the previously received spectral value of that stream. And a set of prediction coefficients a _j
Calculate (n). The prediction coefficients are used in turn to calculate error values e _j (n) for the spectral values. The error value for each stream is provided to the input of a quantizer 3 arranged to generate a quantized error e _j (n) for the next digital transmission.
The quantized errors e _j- (n) are provided to a multiplexer 4 for generating a multiplexed error signal 9 for transmission, and are also fed back to a predictor 2.

Another signal processing unit 5 is also provided for controlling the operation of the signal processing unit 1 and the quantizer 3 depending on the psychoacoustic characteristics of the input audio signal g (t). The operation of this unit is conventional and will not be described in detail here.

For each spectral component j, x
(N), x ＾ (n), and x〜 (n) are the input signal to the predictor 2, the predictor output signal, and the reconstructed quantized signal, and e _j (n) And e _j ~ (n)
Are the prediction error signal and the quantized prediction error signal. One set of prediction coefficients is

(Equation 3) , Which is time-dependent, where the superscript T stands for Transpose.
The output signal x ＾ (n) of the predictor 2 is

(Equation 4) Where is calculated by

(Equation 5) And P is the prediction order or number of coefficients. The predictor error is

(Equation 6) And the reconstructed quantized signal is

(Equation 7) It is. The calculation of the predictor coefficients is based on minimizing the mean square prediction error. a (n) is

(Equation 8)

Once the autocorrelation function r (n) is obtained, it will be seen that a linear predictor can be obtained by solving a normal equation. However, here the least squares algorithm is represented for estimating the linear predictor coefficients for each sample. Least squares often yield better linear prediction coefficient estimates than autocorrelation methods, especially when the number of available data is small. It is shown below that when the order of the predictor is low, especially when it is only 2, the complexity of the least squares algorithm is comparable to or less than that of the prior art adaptive lattice algorithm.

The reconstructed quantized signal has x〜
Assume again that it is indicated by (n). If the prediction order is 2 and the block length is L, the covariance of the reconstructed signal is

(Equation 9) Is calculated by An efficient algorithm is

(Equation 10) Becomes With these covariances, two linear predictor coefficients can be calculated as follows:

[Equation 11]

It will be appreciated that the linear prediction coefficients are obtained from a predetermined or fixed, relatively small number of previous spectral values. The calculation of the coefficients does not depend on all previously received spectral values.

To increase the robustness of backward adaptive prediction for channel errors and numerical roundoff errors, bandwidth extension can be performed after the linear prediction coefficients are obtained. The linear prediction coefficients calculated by the above equation are represented by a _i ,
It is assumed that i = 0, 1, and 2. Here, a ₀ = 1. The bandwidth extension operation replaces each a _i with γ ⁱ a _i . Here, γ is a constant slightly smaller than 1.

As can be seen from the previous section, the covariance function is updated on a sample basis. Therefore, by solving the normal equation, a linear prediction coefficient can be obtained for each sample. However, the linear prediction coefficients can be calculated less frequently to save computation. For example, the linear prediction coefficients may be calculated once every two samples. The loss of the average prediction gain may be ignored. However, the loss of prediction gain is clearly perceived when a transition occurs in the audio signal to be coded. Thus, a transition detector 10 is included that switches the predictor from a normal low update rate (eg, every second spectral value) to a high update rate (eg, every spectral value) when a transition is detected. The high update speed only needs to be maintained for a short time after detecting the transition.

Assume that G _l represents the expected gain in magnification band l. If G _l > 0, the predictors in this subband can switch depending on the overall prediction gain calculated as follows.

(Equation 12) Where N _s is the number of magnification band. If G corrects the extra bits needed for predictor side information, ie, G> T ₁ (dB), or if the prediction gain does not drop sharply, ie, G ^Present −
If G ^Previous <T ₂ (dB), the complete side information is transmitted and the predictor producing the positive gain is switched on, otherwise the predictor is not used, which is It also means that a transition occurs. After the transition frame is detected, the backward adaptive prediction coefficient is calculated on a sample basis. After a certain number of samples, the prediction coefficients are calculated every second sample.

FIG. 2 shows an apparatus for decoding an encoded signal using the method described in detail above.
The received multiplexed error signal 9 has the signal 1024
Are supplied to the input of a demultiplexer 6 which separates the spectral value streams into e _j (n). These streams are then sent to the signal processing unit 7. For each stream, this unit 7 calculates, for each error value, a predicted or estimated spectral value. A predetermined number of these predicted values are used in turn to calculate a linear prediction coefficient so that a predicted value can be calculated for the current sample. This process is consistent with that described for the encoding process. The reconstructed spectral values are obtained by combining the received error signal with the corresponding predicted values. The reconstructed stream of spectral values is provided to another processing unit 8 that performs inverse MDCT on the data to substantially regenerate the original audio signal.

FIG. 3 shows a mobile telephone 11 incorporating in its transmitter a device 12 (corresponding to the device of FIG. 1) for encoding a radiotelephone signal using the encoding method described above. Is shown. The telephone also has, in its receiver, an apparatus 1 for decoding the received encoded telephone signal.
3 (corresponding to the device of FIG. 2) is also incorporated.

[Brief description of the drawings]

FIG. 1 schematically shows an apparatus for encoding a speech signal using backward adaptive prediction according to an embodiment of the present invention.

FIG. 2 schematically shows a device for decoding a speech signal encoded by the device of FIG. 1;

FIG. 3 shows a mobile telephone incorporating the apparatus of FIGS. 1 and 2;

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 first signal processing unit 2 backward adaptive predictor 3 quantizer 4 multiplexer 5 signal processing unit 6 demultiplexer 7 signal processing unit 8 processing unit 9 error signal 10 transition detector 11 mobile telephone 12 encoding wireless telephone signal 13 Device for decoding telephone signals

Claims (15)

[Claims] 1. An encoding method for encoding an audio electrical signal using backward adaptive prediction, comprising: (a) receiving a first time frame of the audio electrical signal to be encoded; Transforming a time frame into the frequency domain to generate a frequency spectrum having 512 or more spectral components; and (c) receiving a next time frame of the audio electrical signal and obtaining a spectral data value for each spectral component. (B) repeating step (b) for these frames in turn to generate a stream of the previously determined reconstructed spectral values of a predetermined number of streams for each said stream. Calculate a set of prediction coefficients for each spectral value using to generate the predicted spectral values Using the set of prediction coefficients to calculate an error between a predicted spectral value and a corresponding actual spectral value, the calculated error being encoded in the spectral value stream. Providing a representation and recombining said error with a predicted spectral value to obtain a reconstructed spectral value. 2. The method according to claim 1, wherein the prediction order is two. 3. The method according to claim 1, wherein the prediction coefficients are recalculated only after receiving a multiple of the spectral values, and the same coefficients are used for several successive spectral values. 4. The method of claim 3, wherein said multiple is two. 5. The method according to claim 3, further comprising the step of switching between a low coefficient update rate and a high update rate as soon as a transition of the audio signal to be coded is detected. 6. The method according to claim 1, wherein the predetermined number of spectral values is four or more. 7. The method according to claim 1, wherein the predetermined number of spectral values is less than or equal to 10. 8. The method according to claim 1, wherein a least squares method is used to determine the prediction coefficients. 9. The method of claim 1, wherein the covariance is 9. The method according to claim 8, when dependent on claim 2, characterized in that: 10. The prediction coefficient is given by: 10. The method according to claim 9, wherein the method is determined according to: 11. A method for decoding an encoded audio electrical signal, comprising: receiving a series of error values corresponding to the encoded audio signal as an input signal; Using a prediction coefficient to determine a corresponding predicted spectral component value for each error value, wherein the prediction coefficient is a covariance of a predetermined number of previously determined continuous predicted spectral components for the stream. Combining the error values and the predicted spectral values to provide the reconstructed spectral values calculated using: and combining the reconstructed spectral values of all streams into a frequency-to-time transform. And thereby substantially reconstructing the audio signal. 12. An apparatus for encoding an audio electrical signal using backward adaptive prediction, comprising: an input for receiving an audio electrical signal to be encoded; and a time frame receiving the received signal. A time-frequency domain converter for transforming from the domain to the frequency domain to provide a frequency spectrum having 512 or more spectral components; receiving the associated spectral values as a stream and reconstructing a predetermined number of previously reconstructed Calculating a set of predictive coefficients for each spectral value using the covariance of the calculated spectral values, using the set of predictive coefficients to generate a predicted spectral value; and And each of the spectral components and associated signal processing means for calculating an error between the corresponding actual spectral value and the calculated error is received. Providing an encoded representation of spectral values stream, the error encoding apparatus characterized by capable of recombining with predicted spectral values to obtain the spectral values reconstructed. 13. An apparatus for decoding an encoded audio electrical signal, comprising: an input for receiving a series of error values corresponding to the encoded audio signal; Signal processing means for determining a corresponding predicted spectral component value using a set of prediction coefficients for each error value, wherein the signal processing means comprises: Arranged to calculate a prediction coefficient using the covariance of the determined successively reconstructed spectral values, wherein signal processing means combine each error value with a corresponding predicted spectral value. A decoding device for providing reconstructed spectral values. apparatus. 14. A communication system comprising a combination of the devices according to claim 12 and 13. 15. A mobile communication device comprising a combination of the devices according to claim 12 and claim 13.