JP2009069309A - Linear prediction model order determination device and method, and program, and recording medium thereof - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To determine an optimal model order of a linear prediction model without actually calculating a coding amount with each model order. <P>SOLUTION: A PARCOR coefficient of each linear prediction model up to a model order m, and prediction error energy at the time of the model order m, are calculated from an input signal. Then, the coding amount of a prediction error waveform which is an error between a prediction value of the linear prediction model of the model order m and the input signal, is estimated from the prediction error energy at the time of the model order m. Moreover, an individual coding amount corresponding to each PARCOR coefficient up to the model order m, or a whole coding amount corresponding to a whole PARCOR efficient is calculated. A sum coding amount of either the whole coding amount or a sum of each individual coding amount up to the model order j, and the estimated coding amount of the prediction error waveform, with each model order j, is calculated for a plurality of model orders, to determine one of the model orders from the sum coding amount. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a linear prediction analysis technique used for signal analysis and signal coding.

The linear prediction model includes a sample value x _t of a time-series discrete signal at a certain time point {t}, p time points {t-1} past the time point {t}, a time point {t-2},. - the sample value _{x t-1} of the time-series discrete signal at the time {t-p}, x t -2, ···, respectively coefficients φ _{i (i} = 1 to _{x t-p,} · · ·, p and those weighted by) may be represented by formula (a) as the assumed model and linear combination holds between the linear prediction error epsilon _t. For example, when any of the coefficients φ _i is 0, strictly speaking, the linear prediction model cannot be said to be based on the sample values of the past p time-series discrete signals, but including such a case, It is considered to be based on the past p time series discrete signal sample values. Here, p represents the number of samples required for the linear prediction model, and this is referred to as a model order.

  In all-pole linear prediction analysis, determination of the model order of the linear prediction model may be required. For example, taking compression coding as an example, it is necessary to determine the model order in order to minimize the sum of the code amounts of the auxiliary information related to the linear prediction coefficient and the prediction error waveform. The model order to be determined in this way is referred to as the optimal model order.

  In the prior art, the optimal model order is determined using FPE (Final Prediction Error: Final Prediction Error), AIC (Akaike Information Criterion: Akaike Information Criterion), etc. [see Non-Patent Document 1]. There is also a method for determining the optimal model order based on the MDL principle (Minimum Description Length Principle) [see Non-Patent Document 2].

  In the two-stage encoding based on the MDL principle, (code word length) = (description length of model) + (description length of data based on the model). If this is replaced with, for example, a conventional expression in lossless coding (Lossless Coding), (code amount necessary for lossless decoding) = (code amount necessary for PARCOR coefficient) + (code necessary for prediction error waveform) Amount)... (B). As shown in FIG. 8, the code amount (symbol 503) required for the prediction error waveform can be decreased as the model order is increased, but the code amount (symbol 502) required for the PARCOR coefficient is increased. By setting the model order large, it is not possible to reduce the amount of code (symbol 501) required for lossless decoding. Therefore, according to the equation (B), a model order that minimizes the amount of codes necessary for lossless decoding is searched, and this is determined as the optimum model order.

This will be described with reference to FIG. 9 as an example of calculating the PARCOR coefficient.
Each frame obtained by dividing a time discrete signal sampled at a predetermined time interval for each frame is set as an input signal y. Here, lossless encoding is taken as an example, and the input signal y is assumed to be an integer value obtained by integer conversion. The optimal model order is an n-order search range in which the model order is from N _min to N _max (where n is an integer that satisfies n ≧ 1). N _min and N _max are positive integers that satisfy N _max > N _min . ] Will be decided from those investigated. In the following, the model order m is a positive integer.

PARCOR coefficient calculation unit (901) is input with an input signal y, the model order is 1, 2, · · _·, PARCOR coefficient of each linear prediction model of _{N max} gamma _m [m = 1, 2, · · ·, N _max ] is calculated [refer to Non-Patent Document 3 or “Calculation of PARCOR coefficient / prediction error energy” described later for the PARCOR coefficient calculation method]. ]. Model order is 1, _2, each PARCOR coefficient when the _{N max} gamma _m [m = 1,2, _{..., N max]} becomes an input of PARCOR coefficient quantizer (902).

The PARCOR coefficient quantization unit (902) quantizes each PARCOR coefficient γ _m [m = 1, 2,..., N _max ], and the model order is 1, 2 _,. Quantized PARCOR coefficients qγ _m [m = 1, 2,..., N _max ] are calculated.
Further, the PARCOR coefficient quantizing unit (902) uses the quantization table for each quantized PARCOR coefficient qγ _m [m = 1, 2,..., N _max ] as an index. These are converted into numerical values and encoded to calculate coefficient codes Cp _m [m = 1, 2,..., N _max ].

Since the quantization method can use a known and well-known method, a detailed description thereof is omitted [for example, see Non-Patent Document 4]. ]. As described above, each of the quantized PARCOR coefficients qγ _m [m = 1, 2,..., N _max ] may be quantized, but some or all of them may be vector quantized together. Good. However, in the case of vector quantization, codes up to each model order included in the search range [that is, codes from model order 1 to N _min , codes from model order 1 to (N _min + n), models The orders are codes from 1 to (N _min + 2n),. ] Is set so that it can be selected by the optimal model order search / code selection unit (906), or each code is output to the optimal model order search / code selection unit (906). It takes a thing.
Each quantized PARCOR coefficient qγ _m [m = 1, 2,..., N _max ] is input to the linear prediction coefficient conversion unit (903).
Each coefficient code Cp _m [m = 1, 2,..., N _max ] is input to the optimum model order search / code selection unit (906).

The linear prediction coefficient conversion unit (903) calculates each model order for each n of the model orders from N _min to N _max from each quantized PARCOR coefficient qγ _m [m = 1, 2,..., N _max]. [Quantized] linear prediction coefficients a _k ^(m) _{m, k} [search target model order m = N _min , N _min + n, N _min + 2n,..., N _min + n × i,..., N _min + n × [(N _max −N _min ) / n], where [·] is a Gaussian symbol. As i _max = [(N _max −N _min ) / n], it is expressed as “i = 0, 1, 2,..., i _max ”. K = 1, 2,..., M. [Refer to Non-Patent Document 3 or << Calculation of PARCOR coefficient / prediction error energy >> described later for the calculation method of the linear prediction coefficient]. ]. That is, assuming that i = 0, 1, 2,..., I _max , the linear prediction coefficient a _k ^(g) of the linear prediction model of model order (N _min + n × i ⁾ [k = 1, 2,. , G] is calculated [N _min + n × i = g, and the symbols are replaced. ]. Note that the linear prediction coefficient a _g ^(g) is already given as a PARCOR coefficient.
The linear prediction coefficient a _k ^(g) [k = 1, 2,..., G] of the linear prediction model of model order (N _min + n × i) is a prediction error filter unit (904-i) [where i = 0, 1, 2,..., I _max ].

The prediction error filter unit (904-i) [i = 0, 1, 2,..., I _max ] receives the input signal y as an input, and linear prediction coefficients a _k ^(g) [k = 1, 2, .., G] are passed through a prediction error filter with a prediction coefficient, for example, the input signal y is passed forward, and the input signal y and the prediction value of the linear prediction model of the model order g consisting of the linear prediction coefficient a _k ^(g) and it outputs a prediction error signal [prediction error waveform] e _i is the error.
The linear prediction error signal e _i is input to an error signal encoding unit (905-i) [i = 0, 1, 2,..., I _max ].

Each of the error signal encoding units (905-i) [i = 0, 1, 2,..., I _max ] encodes the prediction error signal e _i to generate an error of the model order (N _min + n × i). The signal code C _g [g = N _min + n × i] is output.
The error signal code C _g [g = N _min + n × i] is input to the optimum model order search / code selection unit (906).

The optimal model order search / code selection unit (906) performs error signal code C _g [g = N for model order (N _min + n × i) for i = 0, 1, 2,..., I _{max. min} + n × the amount of code i], the model order is _{(n min} + n × i) coefficients up code _{Cp m (m = 1,2, ···} , n min + n × i) the sum of the code amount of A certain total code amount is obtained, and a minimum total code amount is searched for among the obtained total code amounts. For example, when the code having the minimum total code amount is i = I, that is, when the model order is (N _min + n × I), the code amount and the model order of the error signal code C _g [g = N _min + n × I] are _{(n min} + n × I) to the coefficient code _{Cp m (m = 1,2, ···} , n min + n × I) the total code amount of the code amount of is assumed to be minimal, optimal model order is ( N _min + n × I).
Then, the communication path, coefficients corresponding to the PARCOR coefficient to the optimal model order _{(N min} + n × I) code _{Cp m (m = 1,2, ···} , N min + n × I) and optimal model order _{(n min} + n × I) error signal code _{C g} in the case of _{[g = n min} + n × I] is sent out. If the optimum model order cannot be determined from the coefficient code Cp _m (m = 1, 2,..., N _min + n × I), information representing the optimum model order [order information] is encoded as appropriate. The encoded order information is also sent to the communication path.

As in the above example, conventionally, the total code amount in each model order is obtained by brute force, and the smallest one is searched.
Osamu Ozaki, Genshiro Kitagawa, “Method of Time Series Analysis”, Asakura Shoten, pp.82-92, 1998 Information Theory and Its Applied Society, “Information Theory and Its Applied Series 1-I Information Source Coding-Undistorted Data Compression”, Baifukan, 1998 By Hino Mikio, “Spectral Analysis”, Asakura Shoten, 1979 Takehiro Moriya, “Speech coding”, The Institute of Electronics, Information and Communication Engineers, 1998

  For each model order, the auxiliary information related to the prediction coefficient and the prediction error waveform are encoded to obtain the total code amount, and the model order when the total code amount is minimized can be determined as the optimal model order. However, there is a problem that a lot of processing time is required.

  Therefore, in view of the above problems, the present invention provides a linear prediction model order determination apparatus, method, program, and recording medium that determine the optimal model order of a linear prediction model without actually obtaining the total code amount at each model order. The purpose is to provide.

In order to solve the above problems, the present invention is configured as follows. That is, from the input signal, the model order is 1,..., M [where m is an integer that satisfies 1 ≦ m ≦ N _sup for an integer N _{sup of} 2 or more. ] The PARCOR coefficient of each linear prediction model and the prediction error energy at the model order m are calculated. Then, from the prediction error energy when the model order is m, the prediction value of the linear prediction model of the model order m and the input signal composed of the linear prediction coefficients given by the respective PARCOR coefficients when the model order is 1,. The code amount [prediction error code amount] of the prediction error waveform, which is an error from the above, is estimated. Also, the code amount [individual code amount] corresponding to each PARCOR coefficient when the model order is 1,..., M, or the code amount [total code amount] corresponding to all PARCOR coefficients is calculated. For a plurality of model orders included in model orders from 0 to N _sup , the total code amount or the total of individual code amounts when the model order is 1,. A total [total code amount] with the code amount is obtained, and one model order is determined from the total code amount. The determined model order is the optimum model order.
In this way, the code amount at each model order is not actually obtained, but the prediction error code amount is estimated by the prediction error energy obtained in the PARCOR coefficient calculation process.

  According to the present invention, since the prediction error code amount is estimated based on the prediction error energy obtained in the PARCOR coefficient calculation process, the code amount at each model order [the coefficient code amount and the code amount of the prediction error waveform is actually calculated. The optimum model order of the linear prediction model can be determined without obtaining the total code amount].

[Technical explanation]
In the present invention, the optimal model order is determined by effectively using the prediction error energy obtained in the PARCOR coefficient calculation process.

<< Calculation of PARCOR coefficient and prediction error energy >>
A method for directly obtaining a PARCOR coefficient is known as one of all-pole linear prediction analysis. This calculation method is described in Non-Patent Document 4 and the like. Here, the Levinson-Durbin method will be described as an example.

Time series sampled values of the given input signals _{x 1, x 2, ···,} against _{x N,} consider a linear predictive model of equation (1). In general, p <N. Expression (2) is obtained by multiplying both sides of the p-th order autoregressive process represented by Expression (1) by x _t−τ to obtain an expected value. Note that {ε _t } is a white process. Also, v _m−n = v _n−m = E [x _t−m x _t−n ] = Σ _{t = 1} ^N x _t−m x _t−n represents an autocorrelation function.

_v-tau = v in mind that the _tau, tau = 1, 2 in the formula (2), ..., substituting p, to obtain a formula (3). Equation (3) is a so-called Yule-Walker equation. σ ² is a p-order prediction error variance (prediction error energy).

Here, for convenience of explanation, p is replaced with m, i is replaced with k, and symbols are replaced. In addition, the k-th linear prediction coefficient of the ^m- th model is denoted as a _k ^(m) .
At this time, Expression (4) and Expression (5) are defined. Equation (4) is the mth-order prediction error variance (prediction error energy), and Equation (5) is the mth-order partial correlation value.

If formula (3) is transformed and formula (4) is introduced, formula (6) is obtained.

When formula (5) is introduced into formula (6), formula (7) is obtained.

The equation (7) is regarded as a simultaneous equation, an arbitrary coefficient γ _{m + 1} is multiplied between m + 2 equations, subtraction is performed for each side of the equation, and the equation (7) is transformed to obtain the equation (8).

Since the coefficient γ _{m + 1} is arbitrary, it can be determined so that u _m −γ _{m + 1} σ _m ² = 0. At this time, the expression (8) can be understood as an expansion of m to m + 1 in the expression (6). Accordingly, Expression (9), Expression (10), and Expression (11) are obtained. γ _{m + 1} is an m + 1 order PARCOR coefficient, a _k ^{(m + 1)} is a kth linear prediction coefficient of the m + 1 order model, and σ _{m + 1} ² is an m + 1 order prediction error variance (prediction error energy).

The calculation of the PARCOR coefficient or the like based on the Levinson-Durbin method has been described above. However, the calculation is not limited to this method, and may be based on the Burg method, for example. For the Burg method, see Non-Patent Document 3 above.
This completes the description of “Calculation of PARCOR coefficient / prediction error energy”.

<< Information amount of PARCOR coefficient >>
Next, the information amount of the PARCOR coefficient will be described.
The calculation of the information amount of the PARCOR coefficient depends on the quantization method and the encoding method of the PARCOR coefficient. Therefore, here, a general method will be described as an example.
Generally, it is a data table in which a child element table [PARCOR coefficient and its quantized value correspond to each other. ] Is used to convert the PARCOR coefficient into an index value. Further, two examples are shown as encoding methods of the numerical values. The first method is to encode an index value with a predetermined number of bits including a sign bit (a bit representing positive / negative), and the coefficient code is a table of codes corresponding to the index value [index. Is a data table in which the numerical values of and the coefficient codes correspond. ] Can be obtained with reference to. The second method is a case where a variable length code is used depending on the numerical value of the index of the PARCOR coefficient. In this case, since the number of bits varies depending on the absolute value of the index value of the PARCOR coefficient, the code table corresponding to the absolute value of the index value [the data table in which the absolute value of the index value corresponds to the coefficient code. ] To obtain the coefficient code. The code amount of the coefficient code obtained here corresponds to the information amount of the PARCOR coefficient. The table used for encoding is an encoding table [data table. I will say.

In addition, the quantization table and the coding table, which are separate data tables, are not used, but the range of the PARCOR coefficient and the combination of the quantization value and the code corresponding to the range are all the ranges that the PARCOR coefficient can take. Quantization / coding table [Data table. ] May be used.
For example, in the case of fixed-length coding, a quantization / coding table that is simply exemplified in Table 1 can be used. In the case of variable length coding, a quantization / coding table that is simply exemplified in Table 2 can be used. The codes shown in Table 2 are codes that can be instantaneously decoded with 0 as an index.

Two methods are conceivable for calculating the code amount of the coefficient code. The first method is to calculate all code amounts [total code amount] of coefficient codes corresponding to all PARCOR coefficients obtained for each model order from model order 1 to model order m. The second method is to calculate the code amount [individual code amount] of the coefficient code corresponding to the PARCOR coefficient at each model order for each model order from model order 1 to model order m. . In the second method, the sum of the individual code amounts of the respective coefficient codes is handled as the information amount of the PARCOR coefficient. This will also be described in <Determination of Optimal Model Order> described later.
This is the end of the description of “PARCOR coefficient information amount”.

<< Estimation of code amount of prediction error waveform >>
Next, estimation of the code amount of the prediction error waveform that is an error between the prediction value of the linear prediction model and the input signal will be described. The estimation of the code amount of the prediction error waveform is performed using the prediction error energy.

When Expression (11) is used, Expression (12) is established for the m-th order normalized prediction error energy ξ _m ² . Here, the m-th order normalized prediction error energy ξ _m ² is given by, for example, dividing the prediction error energy σ _m ² by the number of samples N. Since σ ₀ ² = v ₀ = E [x _t ² ], ξ ₀ ² is the average energy of the sample series x ₁ , x ₂ ,..., x _N for one frame of the input signal. However, since normalization is the application of scaling to the prediction error energy σ _m ^2, it is not limited to dividing by the number of samples N, and may be divided by N−1 or divided by N−m. And can be set appropriately.

At this time, the mean absolute value sum D when the input signal follows a Gaussian distribution with mean 0 and variance ξ _m ² is given by Expression (13). However, when the average μ of the time series sample values of the input signal is not 0, the average 0 can be obtained by previously subtracting the time series sample values of the input signal by μ. Alternatively, the random variable x shown in the equation (13) may be rewritten to x−μ while keeping the time-series sample value of the input signal as it is. Even in this case, the same conclusion as in the equation (13) is obtained. This also applies to the application of the Laplace distribution described later.

That is, it can be approximated as D≈ξ _m × 0.798.

In the case of a Laplace distribution, the average absolute value sum D is given by Expression (14).

That is, it can be approximated as D≈ξ _m × 0.707.

In this way, it is possible to approximate D≈ξ _m × W using the weight W. Here, the weight W can be determined by the statistical properties of the signal of interest. Actually, D / ξ _m was determined for each model order using 15 audio files of 48 kHz, 16 bits and recording time of 30 seconds, and the frequency was as shown in FIG. 77). Obviously, it tends to be between 0.707 and 0.798. Therefore, when the target signal is an acoustic signal, the weight W may be set to a value of approximately 0.70 to 0.82.

By the way, for example, when a Golomb-Rice code as shown in Table 3 is used as a long code, an estimated value [estimated code amount] S required for a prediction error waveform in a frame is expressed by an average absolute value sum D. It is given by (15). In is a natural logarithm.

Also, when the input signal can assume a Laplace distribution like a speech signal, the entropy H of the Laplace distribution is given by Equation (16), so that the estimated amount of code required for the prediction error waveform in the frame [ Estimated code amount] S can be approximated by equation (17). E in the equation is the number of Napier. However, when the model order is increased, the Laplace distribution approaches the Gaussian distribution. Therefore, the estimation accuracy can be improved by correcting the estimated code amount S using the output value of the function having N and k as variables. The actual code amount in the processed frame (such as the previous frame) may be compared with the estimated code amount, and the weight of the corresponding frame may be changed using a weight that reduces the difference. In the case of multi-channel input, the weight coefficient may be changed in consideration of the results of other channels and the estimated amount. Further, weighting may be performed by considering not only the code amount minimization criterion but also the calculation amount (for example, the lower the model order, the smaller the number of filter taps, the smaller the calculation amount).

The normalized prediction error energy ξ _m ² is not limited to the one obtained by dividing the prediction error energy σ _m ² obtained by Equation (11) by the number of samples, and the PARCOR coefficient γ _m of Equation (11) is it may be employed those obtained by dividing the prediction error energy sigma _m ² obtained by substituting the quantized PARCOR coefficient Qganma _m in samples.

In addition, here, the prediction error energy σ _m ² is described using the prediction error energy σ _m ² obtained by the Levinson-Durbin method. However, it is not intended to limit the prediction error energy σ _m ² obtained by the Levinson-Durbin method. _M- order average output P _m obtained by the Burg method [average output when data x ₁ , x ₂ ,..., X _N are passed forward and backward through a linear prediction model of model order m. ], The same relationship as in the equation (12) is obtained. Therefore, the forward prediction error energy and the backward prediction error energy based on this relationship can be obtained, and the arithmetic average thereof can be used as the prediction error energy σ _m ² .
This is the end of the description of “Estimation of Code Amount of Prediction Error Waveform”.

<Determining the optimal model order>
The method for determining the optimal model order is not particularly limited, and four methods are exemplified here.
In the first method (method A), each model order m in the search range set in advance [not all model orders in the search range, but a part thereof, for example, each model order for each n may be used. ], The information amount of the PARCOR coefficient in the model order [that is, the coefficient code amount, but when the overall code amount is calculated as the coefficient code amount, the entire code amount is used as it is, and the individual code amount is used as the coefficient code amount. Is calculated, the code amount obtained by summing up the individual code amounts corresponding to all model orders of model orders 1 to m is handled as the information amount of the PARCOR coefficient. The same applies hereinafter. And the estimated code amount of the prediction error waveform in the model order [total code amount] is determined, and the model order when the total code amount is minimized is determined as the optimal model order.

  In the second method (Method B), the model order is sequentially set so as to be larger than the model order set immediately before, and the total code amount in the set model order is set to the previous one. When the total code amount in the model order is set, the model order set immediately before is determined as the optimum model order.

  In the second method, when the total code amount in the set model order becomes larger than the total code amount in the previous model order, the model order set in the previous [provisional From the optimal order], a model order that is only a predetermined number (for example, a model order obtained by adding a positive integer F to the provisional optimal order). The total code amount [monitor total code amount] in each model order is obtained, and if there is a model order having a monitor total code amount smaller than the total code amount in the provisional optimum order, the provisional optimum order It is also possible to change to a method in which the model order when the minimum monitor total code amount is smaller than the total code amount in the above is the optimal model order. This modification method (method C) basically uses the model order set immediately before the total code amount larger than the total code amount in the model order set immediately before [provisional The optimal code order] is a promising candidate for the optimal model order, and as a precaution, the increase or decrease in the total code amount is checked up to the model order ahead.

In the fourth method (method D), when the total code amount in the set model order is larger than the total code amount in the model order set immediately before, the model set immediately before The total code amount in the order [provisional optimal order] is set as the provisional minimum total code amount, and the total code amount in the model order further from the provisional optimal order is sequentially obtained. This total code amount is calculated from the provisional minimum total code amount. When the threshold value δ is equal to or greater than a predetermined threshold value δ (or larger than the threshold value δ), the process is stopped and the provisional optimum order is determined as the optimum model order.
This completes the description of “Determining the optimal model order”.

[Embodiment]
Hereinafter, embodiments of the present invention will be described. In the present invention, the optimal model order of the linear prediction model is determined. The optimal model order may be determined as desired. For example, the optimal model order may be used for signal analysis or signal encoding. You may carry out as.
Here, from the viewpoint of clarifying the characteristics of the present invention from the comparison with the above-described prior art, the embodiment will be described by taking as an example the case where the present invention is used for signal coding.

<< First Embodiment >>
A first embodiment of the present invention will be described with reference to the drawings.
<Linear Prediction Model Order Determination Device of First Embodiment>
The linear prediction model order determining apparatus (1) may include an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, a CPU (Central Processing Unit) [cache memory, or the like. ] RAM (Random Access Memory), ROM (Read Only Memory), and external storage devices such as hard disks, and the exchange of data between these input units, output units, CPU, RAM, ROM, and external storage devices It has a bus that connects as much as possible. If necessary, the linear prediction model order determination device (1) may be provided with a device (drive) that can read and write a storage medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device (or ROM, etc.) of the linear prediction model order determination device (1) stores a program for determining the optimal model order of the linear prediction model, data necessary for processing of this program, and the like. These are read into memory as needed. Further, data obtained by the processing of these programs is appropriately stored in a memory such as a RAM, and further read out as appropriate for use in information processing.

Each frame obtained by dividing a time discrete signal sampled at a predetermined time interval for each frame is set as an input signal x. Here, lossless coding [Lossless Coding] is taken as an example, and the input signal x is composed of integer values obtained by integer conversion, and is expressed as x ₁ , x ₂ ,..., X _N as a data string of N samples. To do. It is assumed that the input signal x is stored as data in the external storage device. The optimal model order is an n-order search range in which the model order is from N _min to N _max (where n is an integer that satisfies n ≧ 1). N _min and N _max are positive integers that satisfy N _max > N _min . ] Will be decided from those investigated. In the following, the model order m is a positive integer. In general, if the upper limit of the model order is N _sup [the infinite model order is virtually impossible to investigate, the upper limit of the model order is determined. ], The search range can be set so that the model order is included in the range from 0 to _Nsup . That is, 0 ≦ N _min <N _max ≦ N _sup . However, in the case of N _max <N _sup , it is meaningless to obtain the PARCOR coefficient or the like for the model order from N _max +1 to N _sup, and therefore it is usually considered that N _max = N _sup . In this specification, N _max = N _sup . Also, the case where N _min = 0 is included because, for example, when the input signal is white noise, the zero-order linear prediction model has the smallest code amount (maximum compression rate).

  Signals handled in each embodiment are acoustic signals such as human voice, music, and noise / noise. Of course, the signal is not limited to an acoustic signal, and may be, for example, an earthquake wave or a biological signal (an electroencephalogram or a myoelectric waveform). In short, any signal that is preferably expressed as a linear prediction model may be used.

  The input signal x is not stored in the external storage device in advance. For example, the input signal x is read from a storage medium such as a CD-ROM and stored in the external storage device, or in the case of an audio signal. Suppose that the signal encoding device (2) including the linear prediction model order determining device (1) includes a microphone, and an input signal x is obtained by performing A / D conversion and integer conversion on the collected sound signal obtained by the microphone. It can be considered to be configured. In any case, the component (means) necessary for executing A / D conversion and the like and the data providing technology from the recording medium are achieved by conventional means of the publicly known technology, and therefore the description and illustration are omitted. .

  The external storage device stores the above-described quantization / coding table (data table).

  In the external storage device, a program for calculating the PARCOR coefficient / prediction error energy, a program for estimating the code amount of the prediction error waveform, a program for calculating the coefficient code amount, and an optimal model order are determined. A program for storing is stored. In addition, a control program for controlling processing based on these programs may be stored as appropriate.

  In the linear prediction model order determination device (1), each program stored in the external storage device and data necessary for processing each program are read into the RAM as necessary, and are interpreted and executed by the CPU. As a result, the CPU realizes predetermined functions (PARCOR coefficient calculation unit, prediction error code amount estimation unit, coefficient code amount calculation unit, optimal model order search unit, control unit), thereby determining the optimal model order of the linear prediction model. Realized.

<Linear Prediction Model Order Determination Process of First Embodiment>
Next, with reference to FIG. 1 and FIG. 2, the flow of the optimal model order determination process in the linear prediction model order determination apparatus (1) will be described descriptively. In the first embodiment, the optimal model order is determined by the above method A.

First, the PARCOR coefficient calculation unit (141) receives the input signal x and inputs the PARCOR coefficient γ _m [m = 1, 2,... At each model order for each model order from 1 to N _max. ., N _max ] and prediction error energy σ _m ² [m = 1, 2,..., N _max ] are calculated (step S11). The calculation method of the PARCOR coefficient / prediction error energy is as described in << Calculation of PARCOR coefficient / prediction error energy >>, the PARCOR coefficient γ _m is obtained according to the equation (9), and the prediction error energy σ _{m is} calculated according to the equation (11). ² can be obtained. Since the prediction error energy σ _m ² is obtained in the process of calculating the PARCOR coefficient according to the equation (9), the prediction error energy calculation unit (141a) outputs the prediction error energy σ _m ² stored in the PARCOR coefficient calculation process. do it.
Each prediction error energy σ _m ² [m = 1, 2,..., N _max ] is input to the prediction error code amount estimation unit (142). Each PARCOR coefficient γ _m [m = 1, 2,..., N _max ] is input to the coefficient code amount calculation unit (143).

Next, the prediction error code amount estimation unit (142) uses each prediction error energy σ _m ² [m = 1, 2,..., N _max ] to estimate the code error estimated value [estimated code]. Amount] Ce _m [m = 1, 2,..., N _max ] is calculated (step S12). The calculation of each estimated code amount is as described in << Estimation of code amount of prediction error waveform >>, and for each model order of m = 1, 2,..., N _max using any one method. To do. For example, the estimated code amount S is calculated by Expression (15) using the calculation result of D≈ξ _m ² × W. The weight W and the number N of samples used for normalization are stored in advance in an external storage device, for example. The estimated code amount S is the estimated code amount Ce _m [m = 1, 2,..., N _max ].
Each estimated code amount Ce _m [m = 1, 2,..., N _max ] is input to the optimum model order search unit (144).

In addition, the coefficient code amount calculation unit (143) uses each PARCOR coefficient γ _m [m = 1, 2,..., N _max ] to use the coefficient code amount Cp _m [m = 1, 2,. , N _max ] is calculated (step S13). The calculation of the coefficient code amount is as described in << Information amount of PARCOR coefficient >>, and is performed using any one method. In this embodiment, the individual code amount is obtained as an example.
Each coefficient code amount Cp _m [m = 1, 2,..., N _max ] is input to the optimum model order search unit (144).

  Either step S12 or step S13 may precede. That is, processing can be performed in the order of step S11, step S12, and step S13, and processing can also be performed in the order of step S11, step S13, and step S12.

Then, the optimal model order searching unit (144), each coefficient code amount Cp _m [m = _1,2, ···, _{N max]} and each estimated amount of code Ce _m [m = 1, 2, · · · , N _max ] determines the optimal model order of the linear prediction model. The process will be described.

First, the optimum model order search unit (144) calculates the total code amount in each model order for each model order from N _min to N _max that is the search range (step S14). That is, for each model order m [the search target model order] for each of n _min ≦ m ≦ N _max , the total of the individual code amounts up to the model order m, and the estimated code amount of the prediction error waveform at the model order m Find the sum of

Next, the optimum model order search unit (144) sets N _min to the parameter j for iterative processing, and sets an arbitrary value (initial value) to the parameter M for searching for the minimum total code amount (step S15). ). This initial value is set to a sufficiently large value.

Next, the optimum model order search unit (144) calculates the total code amount in the model order j (that is, the sum of the individual code amounts from the model orders 1 to j and the estimated code amount of the prediction error waveform in the model order j). Total. ] And M are determined (step S16). In the first embodiment, it is determined whether M> (total code amount in model order j) is satisfied.
If it is true that M> (total code amount in model order j) is true, optimal model order search unit (144) executes the process of step S17. If the establishment of M> (total code amount in model order j) is false, the optimal model order search unit (144) executes the process of step S18.

If it is true that M> (total code amount in model order j) is true, the optimum model order search unit (144) substitutes the value of parameter j for parameter N _opt for determining the optimum model order, and parameter M Is substituted with the total code amount in the model order j (step S17). Next, the process of step S18 is executed.

The optimum model order search unit (144) compares and determines the magnitude of the parameter j and the model order _Nmax (step S18). In the first embodiment, it is determined whether j ≧ N _max is satisfied.
If the establishment of j ≧ N _max is true, the optimum model order search unit (144) ends the optimum model order determination process. That is, the value of the parameter N _opt at this time is the optimal model order. The order information representing the optimum model order N _opt may be output from the signal encoding device (2). The optimal model order N _opt is an input to the PARCOR coefficient quantization unit (145).
If the establishment of j ≧ N _max is false, the optimum model order search unit (144) executes the process of step S19.

  The optimal model order search unit (144) sets n to the parameter j, that is, j + n as a new value of the parameter j (step S19), and performs the processing of steps S16 to S18.

Hereinafter, processing related to signal encoding after obtaining the optimal model order will be described.
The PARCOR coefficient quantization unit (145) quantizes each PARCOR coefficient γ _m [m = 1, 2,..., N _opt ] obtained by the PARCOR coefficient calculation unit (141). Quantized PARCOR coefficients qγ _m [m = 1, 2,..., N _opt ] are calculated for each model order up to N _opt . Each quantized PARCOR coefficient qγ _m [m = 1, 2,..., N _opt ] is an input to the linear prediction coefficient conversion unit (146).
Further, the PARCOR coefficient quantizing unit (145) indexes the quantized PARCOR coefficient using the quantization table for each quantized PARCOR coefficient qγ _m [m = 1, 2,..., N _opt ]. Are encoded, and coefficient codes Cp _m [m = 1, 2,..., N _opt ] are respectively calculated. Each coefficient code Cp _m [m = 1, 2,..., N _opt ] is an output of the signal encoding device (2).

Next, the linear prediction coefficient conversion unit (146) calculates [quantized] from each quantized PARCOR coefficient qγ _{m, m} [m = 1, 2,..., N _opt ] according to the equation (10). ] Linear prediction coefficient a _k ^(g) [g = N _opt , k = 1, 2,..., G] is calculated. The [quantized] linear prediction coefficients a _k ^(g) [g = N _opt , k = 1, 2,..., G] are input to the prediction error filter unit (147).

Next, the prediction error filter unit (147) receives the input signal x as an input and inputs it to the prediction error filter using the linear prediction coefficients a _k ^(g) [k = 1, 2,..., G] as prediction coefficients. the signal x through prospective example, and outputs a prediction error signal [prediction error waveform] e _g is the error between the predicted value of the linear prediction model of model order g consisting of the input signal x and the linear prediction coefficients a _{k ^(g)} .
Linear prediction error signal _{e g} becomes an input of the error signal encoding unit (148).

Next, the error signal encoding unit (148) is to encode the prediction error signal _{e g,} to output an error signal code _{C g [g = N opt].} This error signal code C _g [g = N _opt ] is the output of the signal encoding device (2).

  In the first embodiment, the coefficient code amount calculation unit (143) has been described as the case of calculating the individual code amount. When the coefficient code amount calculation unit (143) calculates the total code amount, in the process of step S16, the optimum model order search unit (144) determines the total code amount in the model order j [that is, the total code amount and , And the estimated code amount of the prediction error waveform at the model order j. ] And M may be determined.

<< Second Embodiment >>
A second embodiment of the present invention will be described with reference to the drawings.
Since the hardware configuration and program of the linear prediction model order determination device of the second embodiment are the same as those of the first embodiment, description thereof will be omitted.
<Linear Prediction Model Order Determination Process of Second Embodiment>
With reference to FIG. 1, FIG. 3, and FIG. 4, the flow of the optimal model order determination process in the linear prediction model order determination apparatus (1) of the second embodiment will be described descriptively. In the second embodiment, the optimal model order is determined by the above method C. Note that << calculation of PARCOR coefficient / prediction error energy >>, << estimation of code amount of prediction error waveform >>, and << amount of information of PARCOR coefficient >> are as described above, and thus description thereof is omitted [[technical description] and See << First Embodiment >>. ].

First, the control unit (190) sets N _min for the parameter j and 1 for the parameter f (step S21).

Next, the PARCOR coefficient calculation unit (141) receives the input signal x and inputs the PARCOR coefficient γ _m [m = 1, 2,... , J] and prediction error energy σ _m ² [m = 1, 2,..., J] are calculated (step S22).
Each prediction error energy σ _m ² [m = 1, 2,..., J] is input to the prediction error code amount estimation unit (142). Each PARCOR coefficient γ _m [m = 1, 2,..., J] is input to a coefficient code amount calculation unit (143).

Next, the prediction error code amount estimation unit (142) uses each prediction error energy σ _m ² [m = 1, 2,..., J] to estimate the code error estimated value [estimated code amount]. ] Ce _m [m = j] is calculated (step S23).
Each estimated code amount Ce _m [m = j] is input to the optimum model order search unit (144).

The coefficient code amount calculating section (143), each PARCOR coefficient gamma _m [m = 1,2, ···, j] using a coefficient as a separate code amount code amount Cp _m [m = 1, 2, ..., J] are calculated (step S24). Each coefficient code amount Cp _m [m = 1, 2,..., J] is input to the optimum model order search unit (144).

  Either step S23 or step S24 may precede. That is, it can also process in the order of step S21, step S22, step S23, and step S24, and can also process in the order of step S21, step S22, step S24, and step S23.

Next, the optimal model order search unit (144) sequentially sets the coefficient code amount Cp _m [m = 1, 2,..., J] and the estimated code amount Ce _m [m = j] of the model order m. ] Is used to determine the optimal model order of the linear prediction model. The process will be described.

  First, the optimal model order search unit (144) calculates the total code amount in the model order j (step S25). That is, the sum of the individual code amounts from model orders 1 to j and the estimated code amount of the prediction error waveform at model order j is obtained.

Next, the optimal model order search unit (144) determines whether or not the parameter j is equal to N _min (step S26).
If the parameter j is equal to N _min , the optimal model order search unit (144) executes the process of step S29. If the parameter j is not equal to N _min , the optimal model order search unit (144) executes the process of step S27.

Since j = N _min at this stage, the process of step S29 is performed.
The optimum model order search unit (144) sets the total code amount in the model order j as the parameter M (step S29). Next, the control unit (190) sets the parameter j obtained by adding n, that is, j + n to a new value of the parameter j (step S30), and performs control so as to perform the processing of steps S22 to S26.

After this stage, since it is not determined that the parameter j is equal to N _min in the process of step S26, the process of step S27 is performed.
The optimal model order search unit (144) is the total code amount in the model order j [that is, the sum of the individual code amounts from model orders 1 to j and the estimated code amount of the prediction error waveform in model order j. . ] And M are determined (step S27). In the second embodiment, whether or not M> (total code amount in model order j) is determined as in the first embodiment.
If it is true that M> (total code amount in model order j) is true, optimal model order search unit (144) executes the process of step S28. If the establishment of M> (total code amount in model order j) is false, the optimum model order search unit (144) executes the process of step S32a.

The optimal model order search unit (144) substitutes the value of j−n for the parameter j _temp (step S32a).

Next, the optimum model order search unit (144) has a parameter f and a constant F [the same as the positive integer F described above, and is stored in the memory in advance. ] Is compared and determined (step S32b). In the second embodiment, it is determined whether f ≧ F is satisfied.
At this stage, establishment of f ≧ F is false, and the process of step S33 is executed.

  The optimum model order search unit (144) sets a value obtained by adding 1 to the parameter f, that is, f + 1 to a new value of the parameter f (step S33).

  Subsequently, the optimum model order search unit (144) sets the parameter j plus 1, that is, j + 1 as a new value of the parameter j (step S34).

  And a control part (190) is controlled to perform the process of step S35-S39. The process of step S35 is the process of step S22, the process of step S36 is the process of step S23, the process of step S37 is the process of step S24, the process of step S38 is the process of step S25, and the process of step S39 is the step. Since it is the same as the process of S27, description is abbreviate | omitted.

If M> (total code amount in model order j) is true in the process of step S39, the optimum model order search unit (144) executes the process of step S40. Since the process of step S40 is the same as the process of step S29, description is abbreviate | omitted. Following the process of step S40, the process of step S41 is performed. That is, the optimal model order search unit (144) substitutes the value of the parameter j for the parameter j _temp (step S41).
If it is determined in step S39 that M> (total code amount in model order j) is false, the optimum model order search unit (144) executes step S32b.

In the process of step S32b, the processes of steps S33 to S41 are performed until f ≧ F is true.
If f ≧ F is true in the process of step S32b, the optimum model order search unit (144) substitutes the value of the parameter j _{temp for} the optimum model order N _opt (step S42). Then, the optimum model order search unit (144) ends the optimum model order determination process. The order information representing the optimum model order N _opt may be output from the signal encoding device (2). Further, the optimal model order N _opt can be input to the PARCOR coefficient quantization unit (145).

If M> (total code amount in model order j) is true in the process of step S27, the optimum model order search unit (144) compares and determines the magnitude of parameter j and model order N _max (step S27). S28). In the second embodiment, as in the first embodiment, it is determined whether j ≧ N _max is satisfied.
If the establishment of j ≧ N _max is true, the optimum model order search unit (144) substitutes the value of the parameter j for the optimum model order N _opt (step S31). Then, the optimum model order search unit (144) ends the optimum model order determination process. The order information representing the optimum model order N _opt may be output from the signal encoding device (2). Further, the optimal model order N _opt can be input to the PARCOR coefficient quantization unit (145).
If it is determined in step S28 that j ≧ N _max is false, the optimal model order search unit (144) executes step S29. The processing after step S29 is as described above.

  In the second embodiment, the case where the coefficient code amount calculation unit (143) calculates the individual code amount has been described. When the coefficient code amount calculation unit (143) calculates the total code amount, in the process of step S27, the optimum model order search unit (144) determines the total code amount and the estimated code of the prediction error waveform at the model order j. What is necessary is just to calculate | require the total with quantity. In the process of step S27, the optimal model order search unit (144) calculates the total code amount in the model order j [that is, the sum of the total code amount and the estimated code amount of the prediction error waveform in the model order j. . ] And M may be determined.

  For example, when processing related to signal encoding after obtaining the optimal model order is performed, the description is omitted because it is as described in the first embodiment.

<< Third Embodiment >>
A third embodiment of the present invention will be described with reference to the drawings.
Since the hardware configuration and program of the linear prediction model order determination device of the third embodiment are the same as those of the first embodiment and the second embodiment, description thereof will be omitted.
<Linear Prediction Model Order Determination Processing of Third Embodiment>
With reference to FIG. 1, FIG. 5, and FIG. 6, the flow of the optimum model order determination process in the linear prediction model order determination apparatus (1) of the third embodiment will be described narratively. In the third embodiment, the optimum model order is determined by the above method D.
Since the outline of the processing is the same as that of the second embodiment, only parts different from those of the second embodiment will be described here.

First, the control unit (190) sets N _min to the parameter j (step S21a). Subsequent to step S21a, the processing after step S22 described in the second embodiment is performed.

  In the third embodiment, if M> (total code amount in model order j) is false in the process of step S27, the processes after step S32 are executed.

The optimal model order search unit (144) substitutes the value of j−n for the parameter j _temp (step S32).

Next, the optimal model order search unit (144) performs the process of step S33. The process of step S33 is the same as the process of step S28.
If the establishment of j ≧ N _max is false, the process of step S34 is performed. If the establishment of j ≧ N _max is true, the process of step S42 is performed.

  The optimal model order search unit (144) sets a value obtained by adding 1 to the parameter j, that is, j + 1 as a new value of the parameter j (step S34). In this example, 1 is added, but two or more values may be added.

  And a control part (190) is controlled to perform the process of step S35-S39. The process of step S35 is the process of step S22, the process of step S36 is the process of step S23, the process of step S37 is the process of step S24, the process of step S38 is the process of step S25, and the process of step S39 is the step. Since it is the same as the process of S27, description is abbreviate | omitted.

If M> (total code amount in model order j) is true in the process of step S39, the optimum model order search unit (144) executes the process of step S40. Since the process of step S40 is the same as the process of step S29, description thereof is omitted. Following the process of step S40, the process of step S41 is performed. That is, the optimal model order search unit (144) substitutes the value of the parameter j for the parameter j _temp (step S41). After the process of step S41, the process of step S33 is performed.
If it is determined in step S39 that M> (total code amount in model order j) is false, the optimum model order search unit (144) executes step S39a.

  The optimal model order search unit (144) determines whether (total code amount in model order j) ≧ (M + δ). If the establishment of (total code amount in model order j) ≧ (M + δ) is false, the process of step S33 is performed. If the establishment of (total code amount in model order j) ≧ (M + δ) is true, the process of step S42 is performed.

In the process of step S33, the processes of steps S33 to S41 are performed until j ≧ N _max is true.
In the processing of step S33, if the establishment of j ≧ N _max is true, the optimal model order search unit (144) substitutes the value of the parameter j _{temp for} the optimal model order N _opt (step S42). Then, the optimum model order search unit (144) ends the optimum model order determination process. The order information representing the optimum model order N _opt may be output from the signal encoding device (2). Further, the optimal model order N _opt can be input to the PARCOR coefficient quantization unit (145).

  For example, when processing related to signal encoding after obtaining the optimal model order is performed, the description is omitted because it is as described in the first embodiment.

  In addition to the above embodiments, the linear prediction model order determination apparatus and method according to the present invention are not limited to the above-described embodiments, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the linear prediction model order determination device / method is not only executed in time series according to the order of description, but also in parallel or individually according to the processing capability of the device that executes the processing or as necessary. It may be executed.

  When the processing function in the linear prediction model order determination device is realized by a computer, the processing contents of the function that the linear prediction model order determination device should have are described by a program. Then, by executing this program on a computer, the processing function in the linear prediction model order determination device is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the linear prediction model order determination device is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware. Good.

  The present invention is useful for signal analysis and signal encoding using an all-pole linear prediction model.

The block diagram which shows the function structural example of a linear prediction model order determination apparatus. The figure which shows the processing flow of the linear prediction model order determination process concerning 1st Embodiment. The figure which shows the processing flow of the linear prediction model order determination process concerning 2nd Embodiment. The figure which shows the processing flow of the linear prediction model order determination process concerning 2nd Embodiment (it is a figure attached to FIG. 3). The figure which shows the processing flow of the linear prediction model order determination process concerning 3rd Embodiment. The figure which shows the processing flow of the linear prediction model order determination process concerning 3rd Embodiment (it is a figure attached to FIG. 5). The figure which showed the frequency of appearance of ratio of the average absolute value sum and the square root of prediction error energy when using actual speech data. The figure which shows the relationship between the code amount required for decoding losslessly, the code amount required for a PARCOR coefficient, and the code amount required for a prediction error waveform. The block diagram which shows the function structural example of the conventional linear prediction model order determination apparatus.

Explanation of symbols

1 linear prediction model order determination device 141 prediction error code amount estimation unit 142 coefficient code amount calculation unit 143 optimal model order search unit 145 PARCOR coefficient quantization unit 146 linear prediction coefficient conversion unit 147 prediction error filter unit 148 error signal encoding unit 190 Control unit

Claims (9)

The PARCOR coefficient calculation means has a model order of 1,..., M from the input signal, where m is an integer that satisfies 1 ≦ m ≦ N _sup for an integer N _{sup of} 2 or more. A PARCOR coefficient calculating step for calculating a PARCOR coefficient of each linear prediction model of FIG.
The prediction error code amount estimation means uses the prediction error energy when the model order is m, and the linearity of the model order m composed of the linear prediction coefficients given by the PARCOR coefficients when the model order is 1,. A prediction error code amount estimation step for estimating a code amount [prediction error code amount] of a prediction error waveform that is an error between a prediction value of a prediction model and the input signal;
The coefficient code amount calculation means has a code amount corresponding to each PARCOR coefficient when the model order is 1,..., M [individual code amount], or a code amount corresponding to all PARCOR coefficients [total code amount]. A coefficient code amount calculating step for calculating
The optimum model order search means, for a plurality of model orders included in the model orders from 0 to N _sup , for each model order j, when the overall code amount or model order is 1,. A linear prediction model order determination method comprising: an optimal model order search step for obtaining a total [total code amount] of a total of individual code amounts and the prediction error code amount and determining one model order from the total code amount. The prediction error code amount estimation step includes:
The average absolute value sum of the input signal is obtained by multiplying the normalized square root of the prediction error energy by a predetermined weighting factor, and the prediction error code amount is estimated using the average absolute value sum. The linear prediction model order determination method according to claim 1. The prediction error code amount estimation step includes:
The linear prediction model order determination method according to claim 1, wherein the prediction error code amount is estimated using entropy of a Laplace distribution in which the normalized prediction error energy is distributed. The optimal model order search step includes
For a plurality of model orders included in the model orders from 0 to N _sup , for each model order j, the total code quantity or the sum of the individual code quantities when the model order is 1,. The total code amount with the prediction error code amount is obtained, and the model order when the total code amount is minimized is determined as the optimal model order. The linear prediction model order determination method described. Until the model order is determined in the optimal model order search step, the control means sequentially sets the model order so that the model order is larger than the previous model order, and for each set model order, For the set model order j, there is a control step for executing the PARCOR coefficient calculation step, the prediction error code amount estimation step, the coefficient code amount calculation step, and the optimum model order search step.
The optimal model order search step whose execution is controlled in the control step is:
The total code amount of the individual code amount when the overall code amount or the model order is 1,..., J and the prediction error code amount in the model order j set in the control step. The model order r (provisional optimum order) when the total code quantity is larger than the total code quantity in the model order r set immediately before the model order j is determined in advance. When there is a model order that has a monitor total code amount that is smaller than the total code amount in the provisional optimum order for the total code amount [monitor total code amount] in each model order up to the model order, the monitor total code amount The model order when the minimum monitor total code amount is determined as the optimal model order, and the model order is the monitor total code amount smaller than the total code amount at the provisional optimal order. If but not, linear prediction model order determination method according to any one of claims 1 to 3, characterized in that what determines the temporary optimal order as the optimal model order. Until the model order is determined in the optimal model order search step, the control means sequentially sets the model order so that the model order is larger than the previous model order, and for each set model order, For the set model order j, there is a control step for executing the PARCOR coefficient calculation step, the prediction error code amount estimation step, the coefficient code amount calculation step, and the optimum model order search step.
The optimal model order search step whose execution is controlled in the control step is:
The total code amount of the individual code amount when the overall code amount or the model order is 1,..., J and the prediction error code amount in the model order j set in the control step. The model order r [provisional optimal order] when the total code quantity is larger than the total code quantity [provisional minimum total code quantity] in the model order r set immediately before the model order j is obtained. When the total code amount [monitor total code amount] in the model order is equal to or greater than a predetermined threshold δ than the provisional minimum total code amount, the provisional optimum order is determined as the optimum model order. The linear prediction model order determination method according to any one of claims 1 to 3, wherein From the input signal, the model order is 1,..., M [where m is an integer that satisfies 1 ≦ m ≦ N _sup for an integer N _{sup of} 2 or more. PARCOR coefficient calculating means for calculating the PARCOR coefficient of each linear prediction model and the prediction error energy when the model order is m,
From the prediction error energy when the model order is m, the prediction value of the linear prediction model of the model order m and the input including the linear prediction coefficient given by each PARCOR coefficient when the model order is 1,. A prediction error code amount estimation means for estimating a code amount [prediction error code amount] of a prediction error waveform which is an error with a signal;
Coefficient code amount calculation for calculating a code amount [individual code amount] corresponding to each PARCOR coefficient when the model order is 1,..., M, or a code amount [overall code amount] corresponding to all PARCOR coefficients. Means,
For a plurality of model orders included in the model orders from 0 to N _sup , for each model order j, the total code quantity or the sum of the individual code quantities when the model order is 1,. A linear prediction model order determining device comprising: an optimal model order search means for obtaining a total [total code amount] with the prediction error code amount and determining one model order from the total code amount.       The program for making a computer perform each process of the linear prediction model order determination method in any one of Claims 1-6.       A computer-readable recording medium on which the program according to claim 8 is recorded.

Images