JP2007034262A - Device, method, and program for deciding signal, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To decide a target signal which varies nonlinearly in the amplitude, fundamental frequency, etc., more accurately, in a signal with non-target signals such as noise being mixed therein. <P>SOLUTION: On the basis of a discrete signal x(t), autoregression coefficients of a nonlinear function autoregression model representing the discrete signal are found; and on the basis of the autoregression coefficient, it is determined whether the discrete signal x(t) is the target signal. The signal decision is made by deciding which of the found autoregression coefficient and a threshold is larger or performing pattern recognition, based on the learning contents previously obtained by learning with a known signal. Further, the found autoregression coefficient is normalized and smoothed, as necessary, and the signal determination is made, on the basis of the properly normalized and smoothed autoregression coefficients. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a signal determination device, a signal determination method, a signal determination program, and a recording medium for determining a signal having a property in which amplitude, fundamental frequency, and the like fluctuate nonlinearly.

  In signal processing such as analysis of signals (signals such as human voice, music, noise / noise, biological signals, wireless signals, etc.) and processing of the signals, a section where a signal to be processed exists is estimated. In other words, it is necessary to determine whether a signal in a certain section is a signal to be processed, and this determination accuracy greatly affects the effect of subsequent signal processing.

For example, when the signal is an acoustic signal, many of the acoustic signal processing techniques such as encoding of audio signals and music signals, suppression of noise signals, dereverberation, and automatic speech recognition include multiple types of input signals. Therefore, it is necessary to determine whether the acoustic signal in a certain section is an acoustic signal to be processed, and this determination accuracy greatly affects the effect of subsequent acoustic signal processing. As an acoustic feature (signal feature) for determining this acoustic signal, for example, in an audio signal determination method used in, for example, a mobile phone, each of the frequency spectrum of the audio signal, the energy of the entire band of the audio signal, and each after band division Band energy, the number of zero crossings of a speech signal waveform, and their time derivatives have been used (see Non-Patent Document 1, page 66). In the sound signal determination method using these acoustic features, the input acoustic signal is divided into a certain fixed length of, for example, about 25 ms, the above-described acoustic features are calculated in each divided section, and the values are separately obtained in advance. When the predetermined threshold value is exceeded, the signal in the section is determined as an audio signal, and otherwise, the signal in the section is determined as a non-speech signal.
Benyassine, A., Shlomot, E., and Su, HY. "ITU-T recommendation G.729 Annex B: A silence compression scheme for use with G.729 optimized for V.70 digital simultaneous voice and data applications", IEEE Communications Magazine, pp.64-73, September, 1997.

  However, each of the above signal features shows a similar value between a miscellaneous signal (for example, noise) to be excluded and a target signal (for example, a signal having a certain attribute (described later) such as an audio signal or a music signal). In some cases, there is a problem that sufficient signal determination accuracy cannot be obtained in an environment with a miscellaneous signal.

  Accordingly, the present invention provides a signal determination device, a signal determination method, and a signal determination method for more accurately determining a target signal having a property that the amplitude, fundamental frequency, and the like fluctuate nonlinearly in a signal mixed with a non-target signal such as a miscellaneous signal, An object is to provide a signal determination program and a recording medium.

  In order to solve the above problem, the present invention obtains an autoregressive coefficient of a nonlinear function autoregressive model expressing the discrete signal based on the discrete signal stored in the storage means, and based on the autoregressive coefficient, It is determined whether or not the discrete signal is the target signal.

  The signal determination may be performed by determining the magnitude of the obtained autoregressive coefficient and the threshold value, or may be performed by pattern recognition based on learning content obtained by learning with a known signal in advance.

  Further, the obtained autoregressive coefficient may be normalized / smoothed as necessary, and signal determination may be performed based on the autoregressive coefficient that has been appropriately normalized / smoothed.

  Further, the computer can be operated as a signal determination device by a signal determination program that causes the signal determination device of the present invention to function on the computer. Then, it becomes possible to cause another computer to function as a signal determination device, distribute the signal determination program, and the like by using a computer-readable program recording medium that records the signal determination program.

  According to the present invention, by performing signal determination based on the autoregressive coefficient of the non-linear function autoregressive model that is a signal model, the amplitude, the fundamental frequency, and the like are nonlinear in a signal mixed with a non-target signal such as a miscellaneous signal. It is possible to more accurately determine a signal having a varying property.

<Description of matters common to each embodiment>
The signal determination apparatus / method of the present invention calculates an autoregressive coefficient of a nonlinear function autoregressive model to be applied to an input signal, and performs signal determination based on the calculated autoregressive coefficient. Hereinafter, [signal] and [nonlinear function autoregressive model] in each embodiment will be described.

[signal]
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, a radio signal or a biological signal. It is assumed that the target signal is a signal having the property that the amplitude, the fundamental frequency, and the like fluctuate nonlinearly. The above-mentioned “certain attribute” is “a property that the amplitude, the fundamental frequency and the like fluctuate nonlinearly”.

  A sound collection signal obtained by sound collection means (for example, a microphone) (not shown) (this sound collection signal may include a noise signal which is a non-target signal in addition to a voice signal and a music signal which are target signals. .) Is sampled at a sampling rate of 8,000 Hz, for example, and converted into a quantized discrete signal as appropriate. Hereinafter, this discrete signal is simply referred to as an acoustic signal. Since the components (means) necessary for executing A / D conversion from the collected sound signal to the acoustic signal and the like are all achieved by conventional means in the known art, explanation and illustration are omitted.

  Then, an acoustic signal cutout unit (not shown) cuts out an acoustic signal having a predetermined time length from the acoustic signal as a frame while moving the start point with a certain time width in the time axis direction. For example, an acoustic signal having a length of 200 sample points (8,000 Hz × 25 ms) is cut out while moving the start point by 80 sample points (8,000 Hz × 10 ms). Each frame may be cut out by applying a known window function (for example, a Hamming window or a Gaussian window) to the acoustic signal. Since the cutout of the frame by applying the window function is achieved by known conventional means, the description and illustration are omitted.

  In addition, the frame width to be cut out may be the same for all the frames. For example, signal determination is performed on the acoustic signal of each section appropriately determined by a conventional method in advance, and sufficient accuracy is obtained. Therefore, a longer frame width is used for a section that is definitely the target signal or is determined not to be the target signal, and conversely, a frame width is shortened for a section that has no certainty in signal determination using the conventional method. It may be taken. Here, the certainty determination in the signal determination by the method of the prior art is, for example, energy that exceeds a sufficiently large threshold value by setting a threshold value relating to energy (power) of the acoustic signal sufficiently large or sufficiently small. It may be determined that the acoustic signal in the section having the signal is the target signal, and the acoustic signal in the section having energy not exceeding the sufficiently small threshold is not the target signal. In each of the following embodiments, description will be made assuming that the signal determination by the method according to the prior art is not performed.

  In this specification, “determination of whether or not the signal is a target signal (signal determination)” means that a discrete signal (in this case, an acoustic signal) to be determined is an “amplitude, fundamental frequency, etc.” It is assumed that it is determined whether or not the signal has a characteristic that fluctuates nonlinearly.

The acoustic signal x _i (t ^[i] ) [i ^{] of} each frame cut out as described above is a frame number. t ^[i] represents each sample point in the discrete signal of the frame of number i. Is output by the acoustic signal cutting means. Each frame output by the acoustic signal cutout means is input to an autoregressive coefficient calculation unit (described later) in the signal determination device of the present invention. Whether or not the signal is a target signal is determined for each frame, as will be apparent from the following description.

  The process from the sound collection signal obtained by the microphone to the output of the frame by the sound signal extraction means is to perform discrete signal conversion and extraction processes every moment while collecting sound by the microphone (so-called It is also possible to perform the same processing as the stream processing.) The collected sound signal obtained by the microphone is converted into a discrete signal and stored in an external storage device to be described later, and the stored and stored discrete signal (acoustic signal) ) May be performed (a process similar to a so-called buffer process), and the setting can be changed as appropriate.

  By the way, the signal determination apparatus / method of the present invention determines whether or not an acoustic signal of a certain frame is the target signal. After this determination, the signal determination apparatus / method determines the acoustic signal determined to be the target signal. It is normal that some signal processing is performed (of course, for the purpose of signal determination itself, the processing may be terminated without performing other signal processing). Here, for example, in the case of an acoustic signal, the signal processing is processing for noise suppression, howling prevention, echo suppression, and the like. For this reason, the above stream processing or buffer processing is not processing from the sound collection signal obtained by the microphone until the frame is output by the acoustic signal extraction means, but a predetermined signal from the sound collection signal obtained by the microphone. The processing may be performed until the processing is completed, and the setting can be appropriately changed depending on the content of the signal processing.

In each embodiment, sound signals of a plurality of frames obtained by converting a collected sound signal obtained by a microphone in advance into a discrete signal and performing a cutting process on the discrete signal (acoustic signal) by an acoustic signal cutting unit. It is assumed that the signal x _i (t ^[i] ) is stored in an external storage device described later.

[Nonlinear function autoregressive model]
The nonlinear function autoregressive model handled in each embodiment is, for example, an Exponential autoregressive model (hereinafter referred to as “exponential function autoregressive model”). An exponential autoregressive model has been proposed as a model for predicting time-series data having the property of nonlinear and irregular fluctuation (see Reference 1).
(Reference 1) Haggan, V. and Ozaki, T., "Modeling nonlinear random vibrations using an amplitude-dependent autoregressive time series model", Biometrika, Vol.68, pp.189-196, 1981

In this exponential function autoregressive model, the value x (t) of the acoustic signal at a certain time point {t} is changed to each time point {t−k} [k = 1, 2,.・ P and 1 ≦ p. Here, since p determines the number of terms of the exponential function autoregressive model, this is called the model order. ] Of the sound signal values x (t−k) in FIG. 4] multiplied by the autoregressive coefficient φ _k and the value x of the sound signal at each time point {t−k} before the time point {t}. For each of (t−k), an autoregressive coefficient π _k and a scale parameter γ (where 0 ≦ γ are autoregressive coefficients). ] Multiplied by the power of the base e of the natural logarithm with the product of the value x (t-1) of the acoustic signal at the time point {t-1} as the exponent, and the observation error ε ( t) and is a model predicted by the sum of (t), and can be expressed as in equation (1) (for example, the autoregressive coefficient of the term of the value x (t-2) of the acoustic signal at time {t-2}) When φ ₂ and the autoregressive coefficient π ₂ are 0, strictly speaking, the exponential function autoregressive model cannot be said to be based on the values of the past p acoustic signals. Suppose that it is based on the values of individual acoustic signals.)

The difference from the general autoregressive model (AR model) is that the coefficient of the acoustic signal value x (tk) depends on the variation of the acoustic signal value (this coefficient is not an autoregressive coefficient). Thus, an exponential term (π _k exp (−γx (t−1) ² )) whose value changes is introduced. The exponential autoregressive model can effectively model non-stationary acoustic signals with irregular fluctuations by the effect of this exponential term.

  Signals that exist in nature often have asymmetric vibration characteristics, and in order to model such signals, it is better to use the above-described nonlinear model than to use the AR model. Also, in many parts of audio signals and music signals, almost periodic signals have the property that their period and amplitude continue to vibrate finely. Since an exponential function autoregressive model is considered suitable, an example of a nonlinear function autoregressive model in each embodiment is an exponential function autoregressive model.

Note that the exponent term of the exponential function autoregressive model expressed by Equation (1) uses the power of the base e of the natural logarithm, but is not limited to the base e of the natural logarithm and may be a positive real number. When this positive real numbers ^{a, e -γ = a -μ,} μ = γ / log e since _a the relationship is, it is possible to change the representation of the exponential autoregressive model appropriately.

In addition, among the exponential terms of the exponential function autoregressive model expressed by the equation (1), exp (−γx ² ) [However, in the equation (1), x (t−1) is substituted for the variable x. It is preferable that the value of the function g represented by In other words, the function in the exponent term of the exponential function autoregressive model only needs to have a value that can be taken between 0 and 1. For example, the Boltzmann function [for example, g = exp (x) / (1 + exp (x))], the Gompertz function [for example, g = exp (−exp (−x))], Sigmoid function [for example, g = 1 / (1 + exp (−x))], Lorentz peak function (for example, g = α / (x ² + α); α is a positive constant), forked ( Various functions such as a peak function obtained by normalizing a Voigt function, a pseudo Vogt function, and a Pearson distribution family type 7 function are applicable. In addition, the value that the function can take should be within the range of 0 to 1, but as is apparent from the exponent term of the exponential function autoregressive model expressed by Equation (1), the function has an autoregressive coefficient π. _{Since k} is multiplied, it is not necessary to limit the value that can be taken by the function to be within the range of 0 to 1 if the value multiplied by the autoregressive coefficient π _k is included.

  Among the autoregressive coefficients of the exponential function autoregressive model of Equation (1), the autoregressive coefficient having a relationship with the nonlinearity of the signal is a scale parameter γ that is an exponent of the exponent term. Therefore, in each embodiment, description will be made on the assumption that the scale parameter γ is calculated (a specific value of the scale parameter γ is obtained from a discrete signal that is data). Of course, the purpose is not limited to the scale parameter γ, and other autoregressive coefficients φ and π can also be calculated and used for signal determination. Note that calculating the scale parameter γ (including other autoregressive coefficients) from the discrete signal is the exponential function that best represents this discrete signal (ie, most likely or appropriate) based on the discrete signal. Note that it is none other than finding the autoregressive coefficient of the autoregressive model.

  For reference, the scale parameter γ calculated based on a discrete signal (acoustic signal) is typified by a silent section, a section in which environmental noise is mixed, and an audio signal or music signal that is a substantially periodic signal. FIG. 1 shows that the range of values that can be taken is different in a section including a signal having a characteristic that continues while the period and amplitude vary nonlinearly. FIG. 1 is a diagram illustrating a relationship between an acoustic signal and a scale parameter γ.

  FIG. 1B shows the value of the scale parameter γ calculated by the signal determination device / method of the present invention for an acoustic signal having only a silent section and a voice section as shown in FIG. Show. As shown in the figure, it can be seen that the scale parameter γ has a high value in the silent section and a low value in the voice section. The acoustic signal shown in FIG. 1 (c) is obtained by adding station noise to the audio signal shown in FIG. 1 (a) with a signal-to-noise ratio of 10 dB. The value of the scale parameter γ calculated by the determination device / method is shown in FIG. Again, it can be seen that the value of the scale parameter γ is relatively high in the section where only station noise exists, and is relatively low in the speech section. This means that when the scale parameter γ shows a small value in the exponential function autoregressive model, the nonlinearity represented by the scale parameter γ of the acoustic signal is high, and when the scale parameter γ shows a large value. This is considered to mean that the nonlinearity represented by the parameter γ of the acoustic signal is low. In many cases, an acoustic signal has a high non-linearity represented by the scale parameter γ in a target signal such as a voice signal or a music signal, and a non-linearity represented by the scale parameter γ in a non-target signal such as an environmental noise. . FIG. 1 illustrates this trend.

  The present inventors have obtained knowledge that the autoregressive coefficient in the scale parameter γ and other nonlinear function autoregressive models is useful for signal determination against the background of such experimental facts. The present invention has been created based on such knowledge. Conventionally, there is no signal determination based on the autoregressive coefficient of the nonlinear function autoregressive model.

<First Embodiment>
1st Embodiment of the signal determination apparatus and method of this invention is described.
FIG. 2 is a configuration block diagram illustrating the hardware configuration of the signal determination apparatus (1) according to the first embodiment.
FIG. 3 is a functional configuration diagram of the signal determination device (1) according to the first embodiment.
FIG. 4 is a flowchart showing a process flow of a conventional autoregressive coefficient calculation method.
FIG. 5 is a functional block diagram showing processing functions in the autoregressive coefficient calculation method created by the present inventors.
FIG. 6 is a flowchart showing a process flow of the autoregressive coefficient calculation method created by the present inventors.
FIG. 7 is a functional block diagram showing processing functions in the autoregressive coefficient calculation method (Modification 1) created by the present inventors.
FIG. 8 is a flowchart showing the flow of processing of the autoregressive coefficient calculation method (Modification 1) created by the present inventors.
FIG. 9 is a functional block diagram showing processing functions in the autoregressive coefficient calculation method (Modification 2) created by the present inventors.
FIG. 10 is a flowchart showing the flow of processing of the autoregressive coefficient calculation method (Modification 2) created by the present inventors.
FIG. 11 is a functional block diagram showing processing functions in the autoregressive coefficient calculation method (Modification 3) created by the present inventors.
FIG. 12 is a flowchart showing the flow of processing of the autoregressive coefficient calculation method (Modification 3) created by the present inventors.

  As illustrated in FIG. 2, the signal determination device (1) includes an input unit (11) to which a keyboard or the like can be connected, an output unit (12) to which a liquid crystal display or the like can be connected, and the signal determination device (1) to communicate with the outside. A communication unit (13) to which connectable communication devices (for example, communication cables) can be connected, DSP (Digital Signal Processor) (14) [CPU (Central Processing Unit) may be used. A cache memory or the like may be provided. ], A RAM (15) as a memory, a ROM (16), an external storage device (17) as a hard disk, and an input unit (11), an output unit (12), a communication unit (13), a DSP (14), The bus (18) is connected so that data can be exchanged between the RAM (15), the ROM (16), and the external storage device (17). If necessary, the signal determination device (1) may be provided with a device (drive) that can read and write a storage medium such as a CD-ROM.

Furthermore, the signal determination device (1) can be connected to a microphone for collecting sounds such as voice, music, and noise, and is provided with a signal input unit that receives an input of a collected sound signal obtained by the microphone. A microphone is connected to the signal input unit. The collected sound signal is converted into an acoustic signal which is a discrete signal, the acoustic signal is output as a frame having a predetermined time width by the acoustic signal cutting means, and the obtained acoustic signals x of a plurality of frames are obtained. _{As described above, i} (t ^[i] ) is stored in the external storage device (17).

  The external storage device (17) of the signal determination device (1) stores and stores a program for signal determination, data necessary for processing of this program, and the like [if there is no external storage device, for example, The program may be stored in a ROM that is a read-only storage device. ]. Further, data obtained by the processing of these programs is appropriately stored and stored in a RAM or an external storage device.

  More specifically, a program for obtaining a scale parameter γ of an exponential function autoregressive model representing an acoustic signal, a scale parameter γ, is stored in the external storage device (17) [or ROM, etc.] of the signal determination device (1). And a program for determining whether or not the acoustic signal is the target signal, and data (such as an acoustic signal) necessary for the processing of these programs are stored and stored (the above-described stream processing is performed) In this case, a program for realizing acoustic signal extraction means for extracting a frame by applying a window function to the acoustic signal is also stored and stored. In addition, a control program for controlling processing based on these programs is also stored as appropriate.

  In the signal determination device (1) according to the first embodiment, each program stored in the external storage device (17) [or ROM, etc.] and data necessary for processing each program are stored in the RAM (15) as necessary. Are interpreted and executed and processed by the DSP (14). As a result, DSP (14) implement | achieves a predetermined function (an autoregressive coefficient calculation part, a signal determination part, and also an acoustic signal extraction part as needed), and signal determination is implement | achieved.

Then, next, with reference to FIGS. 4-12, the flow of the signal determination process in a signal determination apparatus (1) is demonstrated sequentially.
The signal determination device (1) of the first embodiment includes an autoregressive coefficient calculation unit (20) and a signal determination unit (21) (see FIG. 3).

In the first embodiment, signal determination is performed for each frame acoustic signal x _i (t ^[i] ). That is, for example, if the total number of frames is F and i = 1, 2,..., F, signal determination is performed on the acoustic signal x ₁ (t ^[1] ) of the _first frame, and then the second The signal determination is performed for the acoustic signal x ₂ (t ^[2] ) of the frame, and this process is repeated until the acoustic signal x _F (t ^[F] ) of the F th frame, thereby determining the signal for the acoustic signals of all frames. Is done.

It should be noted here that each of the acoustic signals x _i (t ^[i] ) of each frame has no correlation in the processing described below. Accordingly, the following description will be given focusing on the acoustic signal x _R (t ^[R] ) [i = R] of a certain frame. For convenience of explanation, x _R (t ^[R] ) is simply abbreviated as x (t). Further, the sample point t is set to t = 1, 2,.

<Calculation of scale parameters>
The autoregressive coefficient calculation unit (20) applies an exponential function autoregressive model to the input acoustic signal x (t) and calculates its scale parameter γ. In general, the scale parameter γ is not calculated alone, but other autoregressive coefficients are also calculated. There are no limitations on the autoregressive coefficient calculation method of the exponential function autoregressive model, and various methods can be applied. Here, the conventional autoregressive coefficient calculation method (A) and the present inventors previously An autoregressive coefficient calculation method (B) for which a patent application has been filed is shown. See Reference 2 and Reference 3 for details of the autoregressive coefficient calculation method (B).
(Reference 2) K.Ishizuka, H.Kato, and T.Nakatani, "Speech signal analysis with exponential autoregressive model", Proc. IEEE Intl, Conf. Acoustics, Speech and Signal Processing (ICASSP), Vol.1, I -225, March 18-23, 2005.
(Reference 3) Japanese Patent Application No. 2005-63162

[Conventional autoregressive coefficient calculation method (A)]
Based on the acoustic signal x (t) given by n sample points, the model order p which is an autoregressive coefficient of the exponential function autoregressive model, the scale parameter γ, the autoregressive coefficient φ _k (hereinafter referred to as “coefficient φ _k ”). ), A conventional method for calculating the autoregressive coefficient π _k (hereinafter abbreviated as “coefficient π _k ”) will be described with reference to FIG.

First, the calculation of the value of the model order p and scale parameter gamma, using Akaike information criterion AIC on Nonlinear Model (p), the value p _a to satisfy the condition of equation _(2), selecting a gamma _a by. Here, m represents the maximum value (arbitrary value) that the model order p can take. Also, argmin (•) is a function that outputs the value of a parameter (p and γ in equation (2)) for which the function in parentheses takes the minimum value.

In order to select p _a and γ _a in equation (2), it is necessary to determine the coefficient φ _k and the coefficient π _k in advance. Therefore, the autoregressive coefficient that uniquely minimizes the value of AIC (p) It is difficult to obtain a set (model order p, scale parameter γ, coefficient φ _k , coefficient π _k ), and it is difficult to avoid local solutions even when numerical optimization is performed.

Specifically, the autoregressive coefficient calculation unit (20) executes the following processing procedures, and sets a set of autoregressive coefficients (model order p, scale parameter γ, coefficient φ _k , coefficient π _k ). Will be asked.

It is assumed that the range of values that the model order p and the scale parameter γ can take are set in advance (for example, in a program). Here, it should be noted that the “possible value” is a discrete value because it is assumed that the autoregressive coefficient is calculated by a computer. The model order p represents the number of terms in the exponential function autoregressive model as described above, and thus an integer value is taken, and p = 1, 2,..., M. The scale parameter γ is a positive real value (AR model when γ = 0), and _{suppose that} [fin] values are taken as γ = γ ₁ , γ _{2 ,.} As for “range”, the range of the model order p is 1 ≦ p ≦ m here, but may be c ≦ p ≦ m for an integer c (≦ m) of 2 or more. The range of the scale parameter γ may be set arbitrarily. Here, γ ₁ ≦ γ ≦ γ _fin is set.

First, the autoregressive coefficient calculation unit (20) fixes the model order p and the scale parameter γ to arbitrary p ₀ and γ ₀ . Here, “arbitrary” is set because the coefficient φ _k and the coefficient π _k for the combination of the fixed model order p and the scale parameter γ are obtained in the later-described processing, and therefore the fixed model order p and the scale parameter γ are determined. This is because there is no necessity of ascending order and descending order. Here, for convenience of explanation, it is assumed that the model order p is sequentially fixed in ascending order (p = 1, p = 2,..., P = m order), and the scale parameter γ is in descending order (γ = γ _fin , γ = γ _fin−1 ,..., γ = γ ₁ (in order). That is, first, the autoregressive coefficient calculation unit (20) fixes the model order to p = 1 (step S1), and fixes the scale parameter to γ = γ _fin (step S2).

After that, the autoregressive coefficient calculation unit (20) uses the least square error estimation method described later to obtain a coefficient φ ^ _{k (fin)} and a coefficient π ^ _{k (fin)} (symbol ^ is the maximum likelihood estimator). (Step S3).

Next, the autoregressive coefficient calculation unit (20) determines whether the scale parameter γ is γ ₁ while the model order is fixed at p = 1 (step S6). If γ = γ ₁ , step S6 is performed. processing of, the process of gamma = gamma ₁ unless step S5. Here, since γ = γ _fin , the autoregressive coefficient calculation unit (20) executes the process of step S5 to set the scale parameter to γ = γ _fin−1 , and the minimum square error estimated value at this time is A certain coefficient φ ^ _{k (fin−1)} and coefficient π ^ _{k (fin−1)} are calculated by the least square error estimation method (step S3). By repeating this process until the scale parameter becomes γ = γ ₁ , a plurality of sets (γ _fin , φ ^) of (scale parameter γ, coefficient φ _k , coefficient π _k ) when the model order is p = 1. _{k (fin)} , π ^ _{k (fin)} ) _{p = 1} , ( _γfin-1 , φ ^ _{k (fin-1)} , π ^ _{k (fin-1)} ) _{p = 1} , ..., (γ ₁ , φ ^ _{k (1)} , π ^ _{k (1)} ) _{p = 1} .

Next, since γ = γ ₁ is established at this stage (step S4), the autoregressive coefficient calculation unit (20) determines whether the model order p is p = m (step S6), and p = m If so, the process of step S8 is performed. If p = m is not satisfied, the process of step S7 is performed. Since p = 1 at this stage, the autoregressive coefficient calculation unit (20) executes the process of step S7, sets the model order to p = 2, repeats the processes of step S2 to step S5, and performs a plurality of steps. (Scale parameter γ, coefficient φ _k , coefficient π _k ) (γ _fin , φ _{k (fin)} , π ^ _{k (fin)} ) _{p = 2} , (γ _{fin −1} , φ ^ _{k (fin −1) )} , Π ^ _{k (fin-1)} ) _{p = 2} ,..., (Γ ₁ , φ ^ _{k (1)} , π ^ _{k (1)} ) _{p = 2} .

The autoregressive coefficient calculation unit (20) repeats these processes until the model order p reaches its maximum value m. As a result, when it is determined in step S6 that p = m is satisfied, γ _fin , φ ^ _{k (fin)} , π ^ _{k (fin)} ) _{p = 1} , (γ _fin−1 , φ ^ _{k (fin− 1)} , π ^ _{k (fin-1)} ) _{p = 1} ,..., (Γ ₁ , φ ^ _{k (1)} , π ^ _{k (1)} ) _{p = 1} , (γ _fin , φ ^ _{k ( fin), π ^ k (fin} )) p = 2, (γ fin-1, φ ^ k (fin-1), π ^ k (fin-1)) p = 2, ···, (γ 1, φ ^ _{k (1)} , π ^ _{k (1)} ) _{p = 2} ,..., (γ _fin , φ ^ _{k (fin)} , π ^ _{k (fin)} ) _{p = m} , (γ _fin−1 , φ ^ _{k (fin-1)} , π ^ _{k (fin-1)} ) _{p = m} , ..., (γ ₁ , φ ^ _{k (1)} , π ^ _{k (1)} ) _{p = m} is obtained. ing. The autoregressive coefficient calculation unit (20) obtains the value of the Akaike information criterion AIC (p) for each of these obtained (model order p, scale parameter γ, coefficient φ _k , coefficient π _k ), A combination of the model order p, the scale parameter γ, the coefficient φ _k , and the coefficient π _k when AIC (p) is minimized is output as an autoregressive coefficient of the exponential function autoregressive model (step S8).

<Least square error estimation method>
As for the estimated least square error value of the coefficient φ _k and the coefficient π _k described above, the acoustic signal x (t) [t = 1, 2,..., N], the value of the model order p, and the value of the scale parameter γ are known. Then, the matrices α, X, and Y are configured as shown in Equation (3) and obtained as a solution of a normal equation using the same. Note that the letter T on the right side of the matrix indicates a transposed matrix.

When the matrices α, X, and Y are configured as in Expression (3), the matrix α ^ including φ ^ _k and π ^ _k , which are the least square error estimated values of the coefficient φ _k and the coefficient π _k , is expressed by Expression (4). ) Matrix calculation.

[Autoregressive coefficient calculation method (B)]
Next, the processing procedure of the autoregressive coefficient calculation method (B) will be described with reference to FIGS.
The model order p is assumed to be the same as that described in the conventional autoregressive coefficient calculation method (A).
In the autoregressive coefficient calculation method (B), the autoregressive coefficient calculation unit (20) includes a model order setting unit (141), a scale parameter initial value setting unit (142), a coefficient least square error estimation calculation unit (143), a scale. A parameter search unit (144), an extreme periodic coefficient search unit (145), a scale parameter determination unit (146), a maximum likelihood estimation amount output unit (147), a total maximum likelihood estimation amount output unit (149), a model order determination unit, The control unit is configured to execute each processing procedure as described below to obtain a set of autoregressive coefficients (model order p, scale parameter γ, coefficient φ _k , coefficient π _k ).

  First, the model order setting unit (142) sets the model order p to p = 1 under the control of the control unit (step S201).

Next, the scale parameter initial value setting unit (142), based on the equation (5), sets an initial value gamma ₀ of the scale parameter gamma (step S202). In Expression (5), Max (·) is a function for obtaining the maximum value among the values in parentheses, and ε is an initial value setting factor. It is assumed that the initial value setting factor ε is set in advance (for example, in a program for realizing the scale parameter initial value setting unit (142)). In the case of a signal whose nonlinearity is small at a location where the amplitude of the signal is large, such as an acoustic signal, the power of the base e of the natural logarithm of the exponential term of the exponential autoregressive model at the location where the signal amplitude is maximum The initial value setting factor ε is preferably a small value so that the value is minimized, but the value is preferably 1.0 × 10 ⁻³ or less. In particular, by setting the value to 1.0 × 10 ⁻⁷ or less, a more suitable autoregressive coefficient can be obtained.

The scale parameter initial value setting unit (142) is read from the external storage device (17) and stored in the acoustic signal x (t) storage area (1503) of the RAM (15) under the control of the control unit. The acoustic signal x (t) is read, and the maximum value Max (x (t) ² ) (where 1 ≦ t ≦ n) when the square value is maximum is obtained, and based on the equation (5). Then, the initial value γ ₀ of the scale parameter γ is obtained and stored in the scale parameter initial value γ ₀ storage area (1505) of the RAM (15).

Note that the initial value γ ₀ of the scale parameter γ is not determined only by the equation (5). The initial value γ ₀ of the scale parameter γ may be set based on the value of the acoustic signal having the maximum absolute value among the acoustic signals x (t). More specifically, the acoustic signal x (t) Among them, the value of the scale parameter γ is minimized with respect to the value of the acoustic signal having the maximum absolute value, and the minimum value is set as the initial value γ ₀ . The initial value γ ₀ obtained by Expression (5) is an example.

Next, the coefficient least square error estimated value calculation unit (143) constructs the matrices α, X, and Y according to Equation (6), and each component of the matrix α (coefficient φ _k , coefficient π _k according to Equation (7)). ) To obtain a matrix α ^ having the least square error estimated value (coefficient φ ^ _k , coefficient π ^ _k ) as components (step S203). The coefficient least square error estimated value calculation unit (143) controls the value of the model order p stored in the model order p storage area (1502) under the control of the control unit (p = 1 at this stage), acoustic signal x stored in the acoustic signal x (t) storage area (1503) (t), reads the initial value gamma ₀ of the scale parameter gamma stored in the scale parameter initial value gamma ₀ storage area (1505), Based on these values, a known least square error estimation method (matrix α, X, Y using the matrix α, X, Y) represented by equation (7) is obtained from the matrix α, X, Y configured as shown in equation (6). The matrix α ^ is obtained by (solution of the equation), and the values of the coefficients φ ^ _k and coefficients π ^ _k which are components of the matrix α ^ are stored in the φ ^ _k , π ^ _k storage area (1506) of the RAM (15). .

Next, the scale parameter search unit (144) obtains the value γ ^ _A of the scale parameter γ based on the equation (8) (step S204). In equation (8), argmin (·) is a function that outputs a parameter (here, a scale parameter γ) that takes the minimum value of the function in parentheses. Further, AIC (p) in Expression (8) is the model order p at this stage and the least square error estimate (coefficient φ ^ _k , coefficient π ^) of (coefficient φ _k , coefficient π _k ) obtained in step S203. _This is the Akaike information criterion given by equation (2) with _k ) fixed. The scale parameter search unit (144) controls the value of the model order p stored in the model order p storage area (1502) under the control of the control unit (p = 1 at this stage), the acoustic signal x ( t) The acoustic signal x (t) stored in the storage area (1503) and the initial value γ ₀ , φ ^ _k , π ^ of the scale parameter γ stored in the scale parameter initial value γ ₀ storage area (1505) The value of coefficient φ ^ _k and coefficient π ^ _k stored in the _k storage area (1506) is read, and within the range of 0 ≦ γ ≦ γ ₀ , for example, gradient method (steepest descent method, Newton method, etc.) _A value γ ^ _A of the scale parameter γ that minimizes the value of the Akaike information criterion AIC (p) is obtained by numerical optimization. This is stored in the γ ^ _A storage area (1507) of the RAM (15).

  The information criterion is not limited to the Akaike information criterion AIC (p), and other information criterion (for example, BIC (Schwarz's Bayesian Information Criterion)) may be used.

Next, the coefficient least square error estimated value calculation unit (143) sets the value γ ^ _A of the scale parameter γ obtained in step S204 as γ ^ in the equation (9) (the model order is also fixed, In the stage, p = 1.), The matrix α, X, Y is constructed according to the equation (9), and the least square error estimation of each component (coefficient φ _k , coefficient π _k ) of the matrix α according to the equation (10). A matrix α ^ having values (coefficients φ ^ _k , coefficients π ^ _k ) as components is _obtained (step S205). The coefficient least square error estimated value calculation unit (143) controls the value of the model order p stored in the model order p storage area (1502) under the control of the control unit (p = 1 at this stage), The acoustic signal x (t) stored in the acoustic signal x (t) storage area (1503), the value γ ^ _A of the scale parameter γ stored in the γ ^ _A storage area (1507) are read, and these values are read. From the matrices α, X, and Y configured as shown in Equation (9) based on the above, the known least square error estimation method expressed in Equation (10) (the solution of the normal equation using the matrices α, X, and Y) ) To obtain the matrix α ^ and store the values of the coefficients φ ^ _k and the coefficients π ^ _k , which are elements of the matrix α ^, in the φ ^ _k , π ^ _k storage area (1506) of the RAM (15).

の固有解λ_０、λ⁻ _０（但しλ⁻ _０はλ_０の共役解を表し、λ_０およびλ⁻ _０をもって式（１１）の上段の固有方程式の全ての固有解を表す。つまり、例えばｒ個の固有解をλ_０１、λ_０２、・・・、λ_０ｒとすると、このｒ個の固有解をλ_０によって表すのである。その他も同様である。）、λ_∞、λ⁻ _∞（但しλ⁻ _∞はλ_∞の共役解を表し、λ_∞およびλ⁻ _∞をもって式（１１）の下段の固有方程式の全ての固有解を表す。）のそれぞれが、式（１２）を満たす範囲で（例えば｜λ_０｜^２＞１は、上記の例でいえば、ｒ個の固有解λ_０１、λ_０２、・・・、λ_０ｒのそれぞれが、｜λ_０１｜^２＞１、｜λ_０２｜^２＞１、・・・、｜λ_０ｒ｜^２＞１であることを意味する。その他も同様である。）、

赤池情報量規準ＡＩＣ（ｐ）を最小にする係数φ_ｋ、係数π_ｋの値φ＾＾_ｋおよびπ＾＾_ｋ（例えばφ＾＾は、φの上に＾が２つ付くものとする。）の組を、例えば勾配法（最急降下法やニュートン法など）などの数値的最適化により求める。極限周期係数探索部（１４５）は、この係数φ＾＾_ｋ、係数π＾＾_ｋの値をＲＡＭ（１５）のφ＾＾_ｋ、π＾＾_ｋ格納領域（１５０８）に格納する。

Then, the limit cycle coefficient searching unit (145), based on equation (13), the coefficient phi _k, determine the value phi ^^ _k and [pi ^^ _k coefficient [pi _k (step S206). The limit period coefficient search unit (145) controls the value of the model order p stored in the model order p storage area (1502) under the control of the control unit (p = 1 at this stage), the acoustic signal x. (T) The acoustic signal x (t) stored in the storage area (1503), the value of the scale parameter γ stored in the γ ^ _A storage area (1507), γ ^ _A , φ ^ _k , π ^ _k stored The values of the coefficients φ ^ _k and the coefficient π ^ _k obtained in step S205 stored in the area (1506) are read, and the two values of the equation (11) using the coefficient φ ^ _k and the coefficient π ^ _k as the coefficients are read. Eigen equation

Eigensolutions lambda _{0 of,} lambda ^- _{0 (where} lambda ^- ₀ represents a conjugate solution of lambda _0, lambda ₀ and lambda ^-. Represents all the specific solutions of ₀ with upper characteristic equation of Equation (11) that is, for example, When r _{eigensolutions} are λ ₀₁ , λ ₀₂ ,..., λ _0r , the r eigen solutions are represented by λ _{0. The} same applies to the others.), λ _∞ , λ ⁻ _∞ ( However lambda ^- _∞ represents a conjugate solution of lambda _∞, lambda _∞ and lambda ^-. each is) representing all unique solutions of lower specific equations have _∞ equation (11), in a range satisfying equation (12) (For example, | λ ₀ | ² > 1 means that, in the above example, each of r _{eigensolutions} λ ₀₁ , λ ₀₂ ,..., Λ _0r is | λ ₀₁ | ² > 1, | λ _02. | ² > 1,..., | Λ _0r | ² > 1 (the same applies to others).

Coefficient phi _k which minimizes the Akaike Information Criterion AIC (p), the coefficient [pi _k values phi ^^ _k and [pi ^^ _k (e.g. phi ^^ is on the phi ^ is a two stick things. ) Is obtained by numerical optimization such as a gradient method (steepest descent method, Newton method, etc.). The limit periodic coefficient search unit (145) stores the values of the coefficient φ ^ _k and coefficient π ^ _k in the φ ^ _k and π ^ _k storage area (1508) of the RAM (15).

In step S206, in order to better model a periodic signal in which the exponential function autoregressive model based on the coefficients φ ^ _k and π ^ _k obtained in step S205 has fluctuations like a speech signal, the expression (11) solutions lambda ₀ obtained by solving the characteristic equation ^{_{of, λ - 0, λ ∞,}} λ - for _∞, lambda _0, lambda ^- inside the _∞ the unit circle ^- outer, lambda _∞, lambda than ₀ is the unit circle This process is based on the fact that it is desirable to satisfy the condition (corresponding to the expression (12)) (see the above-mentioned Reference 1).

Next, the scale parameter search unit (144) obtains the value γ ^ _B of the scale parameter γ based on the equation (14) (step S207). The AIC (p) in equation (14) is the model order p at this stage and the least square error estimate (coefficient φ ^^ _k , coefficient π ^^) of (coefficient φ _k , coefficient π _k ) obtained in step S206. _This is the Akaike information criterion given by equation (2) with _k ) fixed. The scale parameter search unit (144) controls the value of the model order p stored in the model order p storage area (1502) under the control of the control unit (p = 1 at this stage), the acoustic signal x ( acoustic signal x stored in t) storage area (1503) (t), the initial value gamma ₀ of the scale parameter gamma stored in the scale parameter initial value gamma ₀ storage area _{(1505), φ ^^ k,} π ^^ coefficient _k stored is stored in the area (1508) phi ^^ _k, reads the value of the coefficient [pi ^^ _k, in the range of ₀ ≦ γ ≦ γ 0, for example, gradient method (steepest descent method or Newton's method The value γ ^ _B of the scale parameter γ that minimizes the value of the Akaike information criterion AIC (p) is obtained by numerical optimization such as. The scale parameter search unit (144) stores this in the γ ^ _B storage area (1509) of the RAM (15).

As described above, the scale parameter search unit (144) obtains γ ^ _A and γ ^ _B as the values of the scale parameter γ.

Next, the scale parameter determination unit (146) determines that the difference between the value γ ^ _A of the scale parameter γ calculated in step S204 and the value γ ^ _B of the scale parameter γ calculated in step S207 is a determination threshold ε ₂ ( For example, it is set to 1.0 × 10 ⁻¹⁰ ). It is assumed that the determination threshold value ε ₂ is set in advance (for example, in a program that implements the scale parameter determination unit (146)). Control unit reads the determination result stored in the determination result storage area (1511), in the case of the determination result of the determination threshold ε is not ₂ or less, the value of the scale parameter gamma obtained in step S207 gamma ^ _B is stored as a value γ ^ _A of the scale parameter γ in the γ ^ _A storage area (1507) (step S209), and the processing of steps S205 to S207 is performed again based on the value γ ^ _A of the scale parameter γ. Such processing is repeated until it is determined that the difference between the value γ ^ _A of the scale parameter γ and the value γ ^ _B of the scale parameter γ is equal to or less than the determination threshold ε ₂ .

Control unit, the determination result is read storage areas (1511) determination result is stored, in the case of the determination result that is determined threshold epsilon ₂ or less step S208, to execute the steps S210 and subsequent steps To.

Maximum likelihood estimator output unit (147), in step S208, when the difference between the value gamma ^ _A scale parameter gamma, and the value gamma ^ _B of the scale parameter gamma is determined to be the determination threshold epsilon ₂ or less , The value γ ^ _B of the scale parameter γ is read from the γ ^ _B storage area (1509) and output, and in step S207, the coefficient φ _k and the coefficient π when the value γ ^ _B of the scale parameter γ is given _k values φ ^ _k , π ^^ _k are read from the φ ^^ _k , π ^^ _k storage area (1508) and output, thereby obtaining these values as maximum likelihood estimators (step S 210). The maximum likelihood estimation amount output unit (147) stores these maximum likelihood estimation amounts in the maximum likelihood estimation amount storage area (1512) of the RAM (15). Since this maximum likelihood estimator is one when the model order p is p = 1, it is expressed as (γ ^ _B , φ ^^ _k , π ^^ _k ) _{p = 1} for convenience.

In step S210, the maximum likelihood estimator output unit (147) determines that the difference between the value γ ^ _A of the scale parameter γ and the value γ ^ _B of the scale parameter γ is less than or equal to the determination threshold ε ₂ in step S208. The value γ ^ _A of the scale parameter γ when it is determined to be present is read from the γ ^ _B storage area (1509) and output, and further, step S204 (after step S209 has elapsed, step S207 is performed). ), The values of the coefficient φ _k and the coefficient π _k when the value γ ^ _A of the scale parameter γ is given are read from the φ ^^ _k and π ^^ _k storage areas (1508), and output. May be obtained as the maximum likelihood estimation amount.

In step S208, it is assumed to determine whether a determination threshold epsilon ₂ or less, may be intended to determine whether it is less than the determination threshold epsilon _2.

  Subsequently, the model order determination unit reads the model order (p = 1 at this stage) stored in the model order p storage area (1502), and further stores it in the model order range storage area (1501). The maximum value m is read, and it is determined whether these values are equal (S211). If it is determined that p = m does not hold, the model order setting (141) adds 1 to the value of the model order p and stores it in the model order p storage area (1502) (step S212). Since p = 1 at this stage, p = 2 is set.

Then, the processing of steps S203 to S212 is repeated until it is determined in step S211 that the model order p is p = m (note that the above equations (6), (7), (8), (9), the expression (10), the expression (11), the expression (13), and the value of the symbol p in the expression (14) are the values updated in step S212. As a result, when it is determined in step S211 that p = m is established, the maximum likelihood estimator (γ ^ _B , φ ^^ _k ) for each value (p = 1, 2,..., M) of the model order p. , Π ^^ _k ) _{p = 1} , (γ ^ _B , φ ^^ _k , π ^^ _k ) _{p = 2} ,..., (Γ ^ _B , φ ^^ _k , π ^^ _k ) _{p = m} is stored in the maximum likelihood estimation amount storage area (1512). The total maximum likelihood estimator output unit (20) reads a set of these obtained (model order p, scale parameter γ, coefficient φ _k , coefficient π _k ) from the maximum likelihood estimator storage area (1512). The value of the Akaike information criterion AIC (p) is obtained for each group, and the combination of the model order p, the scale parameter γ, the coefficient φ _k , and the coefficient π _k when the AIC (p) is minimized is an exponential autoregressive function. The model is output as an autoregressive coefficient and stored in the total maximum likelihood estimator storage area (1514) (step S213).

<Variation 1 of autoregressive coefficient calculation method (B)>
Below, the modification 1 of said autoregressive coefficient calculation method (B) is demonstrated with reference to FIG. 7, FIG.
In the autoregressive coefficient calculation method (B) described above, the process of step S206 is executed to obtain the coefficient φ _k and the values φ ^^ _k and π ^^ _k of the coefficient π _k . However, by omitting this process, The amount of calculation can be reduced and the calculation time can be shortened.
That is, in the first modification of the autoregressive coefficient calculation method (B), the process of step S206 is omitted.
In this case, the coefficients φ _k and π _k values φ ^ _k and π ^^ _k in step S207 of the first modification of the autoregressive coefficient calculation method (B) are the first modification of the autoregressive coefficient calculation method (B). This depends on the values of the coefficients φ ^ _k and π ^ _k obtained in step S205. Processing in other steps is the same as the processing in the autoregressive coefficient calculation method (B).

<Modification 2 of autoregressive coefficient calculation method (B)>
Below, the modification 2 of said autoregressive coefficient calculation method (B) is demonstrated with reference to FIG. 9, FIG.
In the autoregressive coefficient calculation method (B), the model order p is set to p = 1, 2,..., M, and the autoregressive coefficient is obtained for each model order. In the second modification, the model order p is appropriately fixed to one value q. That is, in the autoregressive coefficient calculation method (B), the above steps S211, S212, and S213 are not required. By doing so, the amount of calculation can be further reduced and the calculation time can be shortened. Processing in other steps is the same as the processing in the autoregressive coefficient calculation method (B).

<Modification 3 of autoregressive coefficient calculation method (B)>
Below, the modification 3 of said autoregressive coefficient calculation method (B) is demonstrated with reference to FIG. 11, FIG.
In the modified example 1 of the autoregressive coefficient calculation method (B), the model order p is set to p = 1, 2,..., M, and the autoregressive coefficient is obtained for each model order. In the second modification of (B), the model order p is appropriately fixed to one value q. That is, in the modified example 1 of the autoregressive coefficient calculation method (B), the above steps S211, S212, and S213 are not necessary. By doing so, the amount of calculation can be further reduced and the calculation time can be shortened. The processing in other steps is the same as the processing in Modification 1 of the autoregressive coefficient calculation method (B).

<Modification 4 of autoregressive coefficient calculation method (B)>
Below, the modification 4 of said autoregressive coefficient calculation method (B) is demonstrated.
Modification 4 of the autoregressive coefficient calculation method (B) includes the autoregressive coefficient calculation method (B), the first modification of the autoregressive coefficient calculation method (B), and the modification of the autoregressive coefficient calculation method (B). In any one of the methods of Example 2 and Modification 3 of the autoregressive coefficient calculation method (B), the processing after step S205 is not executed, and the scale parameter γ (γ ^ _A ) obtained in step S204 is used as the autoregressive coefficient. The output of the calculation unit (20).

<Signal judgment>
As described above, the autoregressive coefficient calculation unit (20), based on the acoustic signal x (t), the scale parameter γ (another autoregressive) of the exponential function autoregressive model expressing the acoustic signal x (t). Coefficient). The calculated scale parameter γ (other autoregressive coefficient) is appropriately stored in the RAM (15) or the like.
The scale parameter γ thus obtained becomes an input to the signal determination unit (21).

The signal determination unit (21) reads the scale parameter γ stored and stored in the RAM (15) and determines whether or not the acoustic signal x (t) is a target signal. A signal determination method in the signal determination unit (21) will be described below.
FIG. 13 is a flowchart illustrating processing of a signal determination method based on a threshold.
FIG. 14 is a flowchart showing processing of a signal determination method based on a support vector machine.

<Determination based on threshold>
First, a method for determining a signal by comparing with the signal determination threshold τ will be described.
It is assumed that the signal determination threshold τ is set in advance (for example, in a program that implements the signal determination unit (21)).
Regarding the relationship between the scale parameter γ and the acoustic signal, as described above, when the acoustic signal is a target signal such as an audio signal, the value of the scale parameter γ decreases, and the acoustic signal is a non-target signal such as a noise signal. The value of the scale parameter γ increases. This indicates that signal determination can be performed by determining the magnitude of the preset signal determination threshold τ and the scale parameter γ.

  Therefore, in the program that realizes the signal determination unit (21), when τ> γ is satisfied in the determination of τ and γ, it is output that “is the target signal”, and when τ> γ is not satisfied. It is described in association with each other so as to output “is a non-target signal” or “not a target signal”. Specifically, a logical value 1 is output when “is a target signal”, and a logical value 0 is output when “is a non-target signal” or “not a target signal”. The signal determination unit (21) realized by interpreting and executing this program by the DSP (14) determines the magnitude between the preset signal determination threshold τ and the scale parameter γ calculated by the autoregressive coefficient calculation unit (20). (Step S50), if τ> γ holds, the signal is determined to be a target signal and a logical value 1 is output (step S51). Otherwise, it is a non-target signal (or target signal). Is not) and a logical value 0 is output (step S52). This output becomes the output of the signal determination device (1).

Note that the present invention is not limited to the magnitude determination between the preset signal determination threshold τ and the scale parameter γ calculated by the autoregressive coefficient calculation unit (20). Autoregressive coefficient calculation unit (20) may be one signal determined by size determination of the autoregressive coefficients other than has been scaled parameter γ to calculate (for example the average value of the coefficient [pi _k) and the signal discrimination threshold tau ₁ by or, The magnitude of the scale parameter γ appropriately weighted with another autoregressive coefficient and the signal determination threshold may be determined, and can be changed as appropriate according to the type and nature of the signal.

<Signal judgment based on pattern recognition>
Next, a method for determining a signal by pattern recognition will be described.
The signal determination unit (21) functions as a so-called discriminator in pattern recognition. The scale parameter γ corresponds to a feature amount in pattern recognition. Since the autoregressive coefficients other than the scale parameter γ are also feature quantities of the acoustic signal, pattern recognition can be performed using these autoregressive coefficients as feature quantities. Here, for convenience of explanation, only the scale parameter γ is used as a feature amount in pattern recognition.

There are various pattern recognition methods, such as Support Vector Machine, K-Nearest Neighbor Method, Hidden Markov Model, Neural Network (Neural). Network). In the present invention, these pattern recognition methods can be applied without particular limitation.
Here, a case using a support vector machine will be described.

<Support vector machine>
The support vector machine is a kind of linear classifier (extendable to a non-linear classifier) and is expressed by a discrimination function f (x). For example, if the output f (X) is positive for an input of certain data X, the data X is identified as class A, and if the output f (X) is negative, the data X is identified as class B. The signal determination unit (21) corresponding to the support vector machine identifies whether or not the acoustic signal x (t) (corresponding to the input scale parameter γ) is the target signal with respect to the input of the scale parameter γ. .

The signal determination unit (21) executes a process expressed by a discriminant function obtained in advance by “learning”. Pre-learning is performed as follows.
First, a plurality of signals (preferably sufficiently large in order to improve learning accuracy) whose signal types (the target signal and the non-target signal, which correspond to “class”) are known. Each value of the scale parameter γ of the signal is obtained, and a value corresponding to the purpose type for each value of the scale parameter γ (this will be referred to as a class value. If it is an external signal, it is set to -1.). A set in which the scale parameter γ is associated with the class value in this way is referred to as learning data.

  A hyperplane (in this case) that separates two types of data (scale parameter γ corresponding to the target signal and scale parameter γ corresponding to the non-target signal) that are classified by the class value based on the learning data thus obtained Is a 0-dimensional hyperplane) and a separation hyperplane that maximizes the distance between two types of data sorted by class value (this is referred to as “margin”). Finding the separation hyperplane from the training data is a quadratic programming problem by formulating a function representing the margin by the Lagrange undetermined multiplier method under the constraint that the training data is discriminated into two types without mistakes. To return to. By solving this quadratic programming problem by numerical calculation, parameters of the discriminant function (coefficients of the discriminant function) are obtained.

  Here, since only the scale parameter γ is used as the (acoustic) signal feature, the discriminant function f (x) is expressed by f (x) = w · x + b. The coefficients w and b of the discriminant function f (x) are obtained by numerical calculation of the quadratic programming problem.

  The signal determination unit (21) executes processing based on the discriminant function f (x) expressed by the coefficients w and b obtained from the learning data as described above. In the above example, since the class value of the target signal is associated with 1 (positive) and the class value of the non-target signal is associated with −1 (negative) in the learning data, the signal determination unit (21) A discrimination function f (x = γ) is obtained with respect to the input of γ, and it is determined whether or not this value is 0 or more (step S53). A logical value 1 is output (step S54), and if the value of f (x = γ) is negative, a logical value 0 indicating that the signal is a non-target signal (or there is no target signal) is output (step S55). . This output becomes the output of the signal determination device (1).

  In the above, only the scale parameter γ is used as the (acoustic) signal feature, but other autoregressive coefficients calculated by the autoregressive coefficient calculating unit (20) can also be used, and the same method as described above The discriminant function is obtained. And the signal determination part (21) performs the process represented by this discriminant function by using the autoregressive coefficient used for obtaining the discriminant function as an input. Further, the autoregressive coefficient to be used is not limited to one type but can be a plurality of types.

In the present invention, obtaining the discriminant function of the support vector machine by learning is not a main component, but has significance in signal determination based on the discriminant function obtained by learning. Learning of the discriminant function of the support vector machine may be based on a known one, for example, see Reference 4.
(Reference 4) Koji Tsuda, “What is Support Vector Machine”, IEICE Journal, 2000, p.

As described above, it is determined whether the acoustic signal x (t) is a target signal or a non-target signal.
By the way, the acoustic signal x (t) here is an acoustic signal x _R (t ^[R] ) [i = R] of a certain frame as already described. The signal determination described above is performed on the acoustic signals x _i (t ^[i] ) [i = 1, 2,... Judgment can be made. That is, signal determination is performed for the acoustic signal x ₁ (t ^[1] ) of the ^first frame, then signal determination is performed for the acoustic signal x ₂ (t ^[2] ) of the _second frame, and this processing is performed as F By repeating up to the acoustic signal x _F (t ^[F] ) of the second frame, signal determination can be performed for the acoustic signals of all frames.

Second Embodiment
Next, a second embodiment of the signal determination apparatus / method of the present invention will be described with reference to FIG.
FIG. 15 is a functional configuration diagram of the signal determination device (1) in the second embodiment.
In the first embodiment, the signal determination is performed for each frame of the acoustic signal x _i (t ^[i] ) [i = 1, 2,..., F]. the acoustic signal ^{_{x i (t [i])}} , each of the scale parameter gamma _i [i = 1, 2, · · ·, F] is calculated, then the scale parameter gamma _i [i = 1, 2, · .., F] to perform signal determination of the acoustic signals x _i (t ^[i] ) of all frames. Since the hardware configuration and functional configuration of the signal determination apparatus in the second embodiment are the same as those in the first embodiment, only the differences from the first embodiment will be described.

First, the autoregressive coefficient calculation unit (20) performs the acoustic signal x ₁ (t ^[ ₁ ^] of the first frame on the acoustic signals x _i (t ^[i] ) of all frames input to the signal determination device (1) ^{. 1])} is calculated scale parameter gamma ₁ of calculating the scale parameter gamma ₂ of the audio signal of the second frame _x 2 ^{(t [2]),} the acoustic signal _x F (t of F-th frame of this process ^[F] ), the scale parameter γ _i [i = 1, 2,..., F] is calculated for the acoustic signals of all frames. The calculated scale parameters γ _i [i = 1, 2,..., F] are appropriately stored and stored in the RAM (15) or the like. Then, the signal determining unit (21), RAM (15) scale parameter stored stored in the gamma _i [i = 1, 2, · · ·, F] reads, each scale parameter gamma _i [i = 1, 2,..., F], the signal determination of the acoustic signal x _i (t ^[i] ) corresponding to each scale parameter γ _i is performed.

  The autoregressive coefficient calculation method by the autoregressive coefficient calculation unit (20) and the signal determination method by the signal determination unit (21) are the same as described in the first embodiment.

<Third Embodiment>
Next, a third embodiment of the signal determination apparatus / method of the present invention will be described with reference to FIGS.
FIG. 16 is a functional configuration diagram of the signal determination device (2) of the third embodiment.
FIG. 17 is a flowchart illustrating a flow of signal determination processing in the signal determination device (2) of the third embodiment.
In the second embodiment, the scale parameters γ _i [i = 1, 2,..., F] are calculated for the acoustic signals x _i (t ^[i] ) of all frames, and then the scale parameter γ The signal determination of the acoustic signals x _i (t ^[i] ) of all frames is performed from _i [i = 1, 2,..., F]. In the third embodiment, in consideration of the fact that the amplitude (magnitude) of the scale parameter γ differs depending on the acoustic signal of each frame, the autoregressive coefficient calculation unit (20) uses the influence of the influence on the signal determination. Normalization of the calculated scale parameter γ _i [i = 1, 2,..., F] is performed. Therefore, only the differences from the second embodiment will be described.

  The external storage device of the signal determination device (2) of the third embodiment stores and stores a signal determination program and data necessary for processing of this program [when there is no external storage device] For example, the program may be stored in a ROM that is a read-only storage device. ]. Further, data obtained by the processing of these programs is appropriately stored and stored in a RAM or an external storage device.

  More specifically, in the external storage device (or ROM, etc.) of the signal determination device (2), a program for obtaining the scale parameter γ of the exponential function autoregressive model expressing the acoustic signal, normalization of the scale parameter γ A program for determining whether or not the acoustic signal is the target signal from the scale parameter γ, and data (such as an acoustic signal) necessary for processing these programs are stored and stored. (Note that when performing the above-described stream processing, a program for realizing acoustic signal extraction means for extracting a frame by applying a window function to the acoustic signal is also stored.) In addition, a control program for controlling processing based on these programs is also stored as appropriate.

  In the signal determination device (2) according to the third embodiment, each program stored in an external storage device (or ROM, etc.) and data necessary for processing each program are read into the RAM (15) as necessary. Then, the interpretation is executed and processed by the DSP (14). As a result, the DSP (14) realizes predetermined functions (autoregressive coefficient calculation unit, normalization unit, signal determination unit, and, if necessary, an acoustic signal extraction unit), thereby realizing signal determination. .

As described in the second embodiment, first, the autoregressive coefficient calculation unit (20) performs the first operation on the acoustic signals x _i (t ^[i] ) of all frames input to the signal determination device (2). calculates the scale parameter gamma ₁ of the audio signal frame _x 1 in ^{(t [1]),} to calculate the scale parameter gamma ₂ of the second frame acoustic signal ^{_{x 2 (t [2])}} , the process F By repeating up to the acoustic signal x _F (t ^[F] ) of the first frame, the scale parameter γ _i [i = 1, 2,..., F] is calculated for the acoustic signals of all frames (step S56). . The calculated scale parameters γ _i [i = 1, 2,..., F] are appropriately stored and stored in the RAM (15) or the like. Next, the normalization unit (22) reads the scale parameter γ _i [i = 1, 2,..., F] stored and stored in the RAM (15), and according to the equation (15), the scale parameter γ _i [ i = 1, 2,..., F] is normalized (step S57). γ ^- _i represents a value obtained by normalization. max (·) and min (·) represent functions for obtaining a maximum value and a minimum value, respectively.

In the normalization expressed by the equation (15), the minimum value of γ ^- _i is 0 and the maximum for the acoustic signals x _i (t ^[i] ) of all frames that are input to the signal determination device (2). Indicates that normalization is performed so that the value becomes 1. Of course, such is not limited to the normalization, for example, gamma mean _i, dispersion, using other statistics such as standard deviation, gamma ^- the range of values of _i 0 ≦ γ ^- the _i ≦ 1 Without limitation, normalization may be performed so that the value of γ ^- _i falls within a certain range.

The normalized scale parameter γ ^- _i [i = 1, 2,..., F] calculated by the normalization unit (22) is appropriately stored and stored in the RAM (15) or the like. Next, the signal determination unit (21) reads the scale parameters γ ^- _i [i = 1, 2,..., F] stored and stored in the RAM (15), and each scale parameter γ ^- _i [i = Based on 1, 2,..., F], signal determination of the acoustic signal x _i (t ^[i] ) corresponding to each scale parameter γ ^- _i is performed (step S58).

  The autoregressive coefficient calculation method by the autoregressive coefficient calculation unit (20) and the signal determination method by the signal determination unit (21) are the same as described in the first embodiment.

<Fourth embodiment>
Next, a fourth embodiment of the signal determination apparatus / method of the present invention will be described with reference to FIGS.
FIG. 18 is a functional configuration diagram of the signal determination device (3) of the fourth embodiment.
FIG. 19 is a flowchart illustrating a flow of signal determination processing in the signal determination device (3) of the fourth embodiment.
In the second embodiment, the scale parameters γ _i [i = 1, 2,..., F] are calculated for the acoustic signals x _i (t ^[i] ) of all frames, and then the scale parameter γ The signal determination of the acoustic signals x _i (t ^[i] ) of all frames is performed from _i [i = 1, 2,..., F]. In the fourth embodiment, the auto-regression coefficient calculation unit (20) calculates in order to correct a rapid variation between the scale parameters γ with respect to the acoustic signals of temporally adjacent frames cut out from the acoustic signal having a predetermined time length. The scale parameter γ _i [i = 1, 2,..., F] is smoothed. Therefore, only the differences from the second embodiment will be described.

  The external storage device of the signal determination device (3) of the fourth embodiment stores and stores a signal determination program and data necessary for the processing of this program [when there is no external storage device] For example, the program may be stored in a ROM that is a read-only storage device. ]. Further, data obtained by the processing of these programs is appropriately stored and stored in a RAM or an external storage device.

  More specifically, in the external storage device (or ROM, etc.) of the signal determination device (3), a program for obtaining the scale parameter γ of the exponential function autoregressive model expressing the acoustic signal, smoothing of the scale parameter γ A program for determining whether or not the acoustic signal is the target signal from the scale parameter γ, and data (such as an acoustic signal) necessary for processing these programs are stored and stored. (Note that when performing the above-described stream processing, a program for realizing acoustic signal extraction means for extracting a frame by applying a window function to the acoustic signal is also stored.) In addition, a control program for controlling processing based on these programs is also stored as appropriate.

  In the signal determination device (3) according to the fourth embodiment, each program stored in an external storage device (or ROM, etc.) and data necessary for processing each program are read into the RAM (15) as necessary. Then, the interpretation is executed and processed by the DSP (14). As a result, the DSP (14) realizes a predetermined function (autoregressive coefficient calculation unit, smoothing unit, signal determination unit, and, if necessary, an acoustic signal extraction unit), thereby realizing signal determination. .

As described in the second embodiment, first, the autoregressive coefficient calculation unit (20) performs the first operation on the acoustic signals x _i (t ^[i] ) of all frames input to the signal determination device (3). calculates the scale parameter gamma ₁ of the audio signal frame _x 1 in ^{(t [1]),} to calculate the scale parameter gamma ₂ of the second frame acoustic signal ^{_{x 2 (t [2])}} , the process F By repeating up to the acoustic signal x _F (t ^[F] ) of the th frame, the scale parameter γ _i [i = 1, 2,..., F] is calculated for the acoustic signals of all frames (step S59). . The calculated scale parameters γ _i [i = 1, 2,..., F] are appropriately stored and stored in the RAM (15) or the like. Next, the smoothing unit (23) reads the scale parameter γ _i [i = 1, 2,..., F] stored and stored in the RAM (15), and according to the equation (16), the scale parameter γ _i [ i = 1, 2,..., F] is smoothed (step S60). γ ^to _h represent values obtained by smoothing.

The smoothing represented by the equation (16) indicates that a moving average of the scale parameter γ is taken with respect to the acoustic signal x _i (t ^[i] ) of a predetermined range of frames that is input to the signal determination device (3). . ν is an integer value representing the number of frames in a predetermined range for which a moving average is taken. Of course, the present invention is not limited to such smoothing, and may be smoothed by an interpolation method, for example.

  Since it is rare in reality that an audio signal or music signal as a target signal suddenly occurs (exists) only in a short time of, for example, about 25 ms, the scale parameter γ is smoothed to obtain a scale parameter. It is possible to improve the signal determination accuracy by correcting the rapid fluctuation between γ.

The smoothed scale parameters γ ^to _h [h = 1, 2,..., F−ν] calculated by the smoothing unit (23) are appropriately stored and stored in the RAM (15) or the like. Next, the signal determination unit (21) reads the scale parameters γ ^to _h [h = 1, 2,..., F−ν] stored and stored in the RAM (15), and each scale parameter γ ^to _h [ Based on h = 1, 2,..., F−ν], the sound signal x _i (t ^[i] ) corresponding to each of the scale parameters γ ^to _h is determined (step S61).

  The autoregressive coefficient calculation method by the autoregressive coefficient calculation unit (20) and the signal determination method by the signal determination unit (21) are the same as described in the first embodiment.

<Fifth Embodiment>
Next, a fifth embodiment of the signal determination apparatus / method of the present invention will be described with reference to FIGS.
FIG. 20 is a functional configuration diagram of the signal determination device (4) of the fifth embodiment.
FIG. 21 is a flowchart showing the flow of signal determination processing in the signal determination apparatus (4) of the fifth embodiment.
In the third embodiment, the normalization unit (22) normalizes the scale parameter γ _i [i = 1, 2,..., F] calculated by the autoregressive coefficient calculation unit (20), and then The signal determination unit (21) performs signal determination based on the normalized scale parameter γ ^- _i [i = 1, 2,..., F]. In the fifth embodiment, the scale parameter γ ^- _i [i = 1, 2,..., F] normalized by the normalization unit (22) is smoothed by the smoothing unit (23), and the signal determination unit (21) performs signal determination based on the scale parameter γ ^- _i [i = 1, 2,..., F] smoothed by the smoothing unit (23). The hardware configuration and functional configuration of the signal determination device are the same as those of the third and fourth embodiments. The fifth embodiment is the same as that of the fourth embodiment after the normalization processing described in the third embodiment. Since the smoothing process described above is executed, the flow of the process of performing the smoothing after the normalization process will be described here.

First, the autoregressive coefficient calculation unit (20) in the signal determination device (4) of the fifth embodiment performs the first operation on the acoustic signals x _i (t ^[i] ) of all frames input to the signal determination device. calculates the scale parameter gamma ₁ of the audio signal frame ^{_{x 1 (t [1])}} , to calculate the scale parameter gamma ₂ of the audio signal of the second frame ^{_{x 2 (t [2])}} , F th this process The scale parameter γ _i [i = 1, 2,..., F] is calculated for the acoustic signals of all the frames by repeating until the acoustic signal x _F (t ^[F] ) of the current frame (step S62). The calculated scale parameter γ _i [i = 1, 2,..., F] is appropriately stored in a RAM or the like. Next, the normalization unit (22) reads the scale parameter γ _i [i = 1, 2,..., F] stored and stored in the RAM, and according to the equation (15), the scale parameter γ _i [i = 1]. , 2,..., F] are normalized (step S63). The normalized scale parameter γ ^- _i [i = 1, 2,..., F] is appropriately stored and stored in a RAM or the like. For the method of normalization, as described above, the average of the gamma _i, dispersion, using other statistics such as standard deviation, gamma ^- the range of values of _i 0 ≦ _γ ^- not limited to _i ≦ 1, Normalization may be performed so that the value of γ ^- _i falls within a certain range.

Subsequently, the smoothing unit (23) reads the normalized scale parameter γ ^- _i [i = 1, 2,..., F] stored and stored in the RAM, and according to the equation (17), the scale parameter Smoothing of γ ⁻ _i [i = 1, 2,..., F] is performed (step S64). γ ^to _h represent values obtained by smoothing. As described above, the smoothing method may be smoothed by an interpolation method, for example.

The smoothed scale parameters γ ^to _h [h = 1, 2,..., F−ν] calculated by the smoothing unit (23) are appropriately stored and stored in the RAM (15) or the like. Next, the signal determination unit (21) reads the scale parameters γ ^to _h [h = 1, 2,..., F−ν] stored and stored in the RAM (15), and each scale parameter γ ^to _h [ Based on h = 1, 2,..., F−ν], the sound signal x _i (t ^[i] ) corresponding to each of the scale parameters γ ^to _h is determined (step S65).

  The autoregressive coefficient calculation method by the autoregressive coefficient calculation unit (20) and the signal determination method by the signal determination unit (21) are the same as described in the first embodiment.

<Sixth Embodiment>
Next, a sixth embodiment of the signal determination apparatus / method of the present invention will be described with reference to FIGS.
FIG. 22 is a functional configuration diagram of the signal determination device (5) of the sixth embodiment.
FIG. 23 is a flowchart illustrating a flow of signal determination processing in the signal determination device (5) of the sixth embodiment.
In the fourth embodiment, the smoothing unit (23) smoothes the scale parameter γ _i [i = 1, 2,..., F] calculated by the autoregressive coefficient calculation unit (20), and then The signal determination unit (21) performs the signal determination based on the smoothed scale parameters γ ^to _h [h = 1, 2,..., F−ν]. In the fifth embodiment, the scale parameters γ ^to _h [h = 1, 2,..., F−ν] smoothed by the smoothing unit (23) are normalized by the normalization unit (22), The determination unit (21) performs signal determination based on the scale parameter γ ⁻ _h [h = 1, 2,..., F−ν] normalized by the normalization unit (22). The hardware configuration and functional configuration of the signal determination device are the same as those of the third and fourth embodiments, and the embodiment of the sixth embodiment is the same as that of the third embodiment after the smoothing process described in the fourth embodiment. Since the described normalization process is executed, the flow of the normalization process after the smoothing process will be described here.

First, the autoregressive coefficient calculation unit (20) in the signal determination device (5) of the sixth embodiment performs the first operation on the acoustic signals x _i (t ^[i] ) of all frames input to the signal determination device. calculates the scale parameter gamma ₁ of the audio signal frame ^{_{x 1 (t [1])}} , to calculate the scale parameter gamma ₂ of the audio signal of the second frame ^{_{x 2 (t [2])}} , F th this process The scale parameter γ _i [i = 1, 2,..., F] is calculated for the acoustic signals of all the frames by repeating until the acoustic signal x _F (t ^[F] ) of the current frame (step S66). The calculated scale parameter γ _i [i = 1, 2,..., F] is appropriately stored in a RAM or the like. Next, the smoothing unit (23) reads the scale parameter γ _i [i = 1, 2,..., F] stored and stored in the RAM, and according to the equation (16), the scale parameter γ _i [i = 1]. , 2,..., F] is smoothed (step S67). The smoothed scale parameters γ ^to _h [h = 1, 2,..., F−ν] are appropriately stored and stored in a RAM or the like. As described above, the smoothing method may be smoothed by an interpolation method, for example.

Subsequently, the normalization unit (22) reads the smoothed scale parameters γ ^to _h [h = 1, 2,..., F−ν] stored and stored in the RAM, according to the equation (18). The scale parameters γ ^to _h [h = 1, 2,..., F−ν] are normalized (step S68). γ ^- _h represents a value obtained by normalization. For the method of normalization, as described above, the average of the gamma ^~ _i, dispersion, using other statistics such as standard deviation, gamma ^- the range of values of _i 0 ≦ γ ^- not limited to _i ≦ 1 Normalization may be performed so that the value of γ ^- _i falls within a certain range.

The normalized scale parameter γ ⁻ _h [h = 1, 2,..., F−ν] calculated by the normalization unit (22) is appropriately stored and stored in a RAM or the like. Next, the signal determination unit (21) reads scale parameters γ ⁻ _h [h = 1, 2,..., F−ν] stored and stored in the RAM, and each scale parameter γ ⁻ _h [h = 1]. , 2,..., F−ν], the signal determination of the acoustic signal x _i (t ^[i] ) corresponding to each scale parameter γ ⁻ _h is performed (step S69).

  The autoregressive coefficient calculation method by the autoregressive coefficient calculation unit (20) and the signal determination method by the signal determination unit (21) are the same as described in the first embodiment.

<Seventh embodiment>
Next, a seventh embodiment of the signal determination apparatus / method of the present invention will be described with reference to FIGS. 24, 25, 26, and 27. FIG.
FIG. 24 is a functional configuration diagram of the signal determination apparatus according to the seventh embodiment.
FIG. 25 is a diagram illustrating a configuration and a connection relationship between the target signal section recognition HMM and the non-target signal section recognition HMM.
FIG. 26 is a diagram illustrating a correspondence relationship between the learning data, the target signal section recognition HMM, and the non-target signal section recognition HMM in learning.
FIG. 27 is a diagram illustrating signal section determination of the time series pattern of the scale parameter γ by the HMM (signal determination model) obtained by learning.

In each of the embodiments described above, the signal determination of the acoustic signal of each frame has been described. This is because the acoustic signal of each frame is cut out from the acoustic signal having a predetermined time length. This is nothing but the detection (determination) of the signal section in the acoustic signal having the predetermined time length.
Then, the scale parameter γ _i [i = 1, 2,..., F] calculated for each acoustic signal of each frame cut out in time series from the acoustic signal having a predetermined time length corresponds to each frame. It is possible to detect (determine) a signal section in an acoustic signal having a predetermined time length from the scale parameter γ _i [i = 1, 2,... It will be possible. Such signal section determination based on the scale parameter γ _i as a time series pattern can be achieved by a Hidden Markov Model (hereinafter referred to as “HMM”).

The signal determination unit (21) uses a scale parameter γ _i [i = 1, 2,..., F] as a time series pattern (however, the scale parameter γ _i may be one that has not been smoothed). Alternatively, it may be smoothed or normalized, or may be normalized and smoothed.) And an HMM is used as a predetermined before each frame is cut out. Detection (determination) of a signal section in a time-length acoustic signal is performed. In each of the above embodiments, the frame, the scale parameter, and the signal determination have a one-to-one correspondence, but such a relationship does not necessarily hold in the signal determination using the HMM. That is, in the signal determination using the HMM, it is determined whether the signal width given by the plurality of frames corresponding to the plurality of scale parameters of a certain time series pattern is the target signal section or the non-target signal section. Keep in mind.

  Note that the autoregressive coefficient calculation method by the autoregressive coefficient calculation unit (20) is the same as that described in the first embodiment, and normalization and smoothing are the same as those described in the above embodiments. Therefore, here, the signal section determination using the HMM in the signal determination unit (21) will be described.

The signal determination unit (21) performs signal section determination using the HMM obtained by the previous “learning”. Pre-learning is performed as follows.
First, two HMMs (the model structure may be the same) are prepared for target signal section recognition and non-target signal section recognition (see FIG. 25). These are referred to as a target signal HMM and a non-target signal HMM. Here, since time series patterns are handled, for example, a so-called left-to-right model (LR model) may be prepared. The number of states and the normal distribution mixture number in each HMM may be arbitrary. Note that the target signal HMM and the non-target signal HMM are learned by the scale parameter of the acoustic signal that is known to be the target signal and the scale parameter of the acoustic signal that is known to be the non-target signal, respectively, and the sequential state division method ( It may be an HMM network formed by SSS (Successive State Splitting). For this learning, a known EM algorithm or Baum-Welch algorithm may be used. By this learning, the state transition probability between states in each HMM and the output that is the probability of outputting the value of the scale parameter (in the transition between states or states). Probability distribution is obtained.

  It is assumed that the target signal HMM and the non-target signal HMM are connected as shown in FIG. The target signal section and the non-target signal section are expressed by the state transition of the target signal HMM and the non-target signal HMM expressed by this connection relation. Further, the target signal HMM and the non-target signal HMM are not limited to one each, but a plurality of HMMs of the same type may be connected to each other in order to correspond to a long time series parameter pattern.

  Next, each HMM is used by using learning data obtained by labeling the target signal section and the non-target signal section on the time series pattern of the scale parameter obtained from the acoustic signal whose target signal section and the non-target signal section are already known. (See FIG. 26. In FIG. 26, arrows indicating transitions between the HMMs are omitted. In addition, the target signal HMM is “trust” and the non-target signal HMM is “out”.) Abbreviated.) In this learning, the transition probability is obtained so that the likelihood for the learning data is maximized.

  As described above, an HMM (signal determination model) for performing signal determination from the time series pattern of the scale parameter is obtained. The signal determination model obtained in this way is assumed to be stored in an external storage device such as a hard disk of the signal determination device.

  When the signal determination unit (21) reads the time series pattern of the scale parameter stored and stored in the RAM, the signal determination unit (21) uses the signal determination model stored and stored in the external storage device to display the time series pattern of the read scale parameter. The path in the signal determination model with the highest probability (likelihood) (the order of passing through the target signal HMM and the non-target signal HMM) is used as the output of the signal section determination (see FIG. 27. The path with the bold line in FIG. 27 selected. For example). This output becomes the output of the signal determination device. In addition, since searching for every route is redundant, a route with low possibility may be pruned to search for a route with high possibility.

In the present invention, obtaining a signal determination model by learning is not a main component, but has significance in signal determination based on the signal determination model obtained by learning. Learning of the signal determination model may be based on a known one. For example, see Reference 5.
(Reference 5) Kenji Kita, Satoshi Nakamura, Masaaki Nagata, “Spoken Language Processing”, Morikita Publishing Co., Ltd., 1996

  The signal determination apparatus / method according to the present invention is not limited to the above-described embodiment, and can be appropriately changed without departing from the gist of the present invention. Further, the processing described in the signal determination device / method is not only executed in time series in the order described, but also executed in parallel or individually as required by the processing capability of the device that executes the processing. It is good.

  When the processing function in the signal determination device is realized by a computer, the processing content of the function that the signal determination device should have is described by a program. Then, by executing this program on a computer, the processing function of the signal determination device is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. The computer-readable recording medium may be any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory. 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 the program stored in its own recording medium and executes the process according to the read program. As another execution form of the program, the computer may directly read the program from the 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 signal determination apparatus is configured by causing a computer to execute a predetermined program. However, at least a part of these processing contents may be realized by hardware.

<Effectiveness of exponential function regression model>
It shows that the exponential autoregressive (ExpAR) model is effective for acoustic signals—especially acoustic signals that can be modeled as nonlinear vibration systems.
For example, a characteristic that contributes to the naturalness and individuality of speech is said to be a vocal cord sound source. The conventional voice transmission system (speech generation model) is vocal fold vibration → vocal tract transmission system → voice radiation. Theoretically, the vocal fold vibration is caused by the generation of glottal volume flow and the temporal change of glottal impedance. Non-linear effects on the vocal tract transmission system. Vocal cord vibration is said to be related to the rapid decay of formant energy. Therefore, it is necessary to model the interaction between the self-excited vibration mechanism of the vocal cords and the transmission mechanism of the vocal tract.

  However, at present, the vocal cord vibration and the vocal tract transmission system are modeled separately. Mainly, the vocal fold vibration is a vibration model, and the vocal tract transmission system is a linear autoregressive model represented by LPC (Linear Predictive Coding). The above-mentioned AR model) is used. Thus, at present, voice data is modeled in two stages, but the ExpAR model can express the characteristics of both vocal fold vibrations and vocal tract transmission systems that are nonlinearly affected by the vocal fold vibrations. It is. The physical characteristics of the ExpAR model will be described below.

A continuous differential equation indicating self-excited vibration can be expressed by Equation (19). Here, x represents a signal, and a and b represent coefficients.

The root ρ of the characteristic equation of this differential equation is given by equation (20). Here, i is an imaginary number. Expression (19) represents a damped oscillation (attenuation fast when a is large, and slow attenuation when a is small) because -a / 2 <0 when a> 0, and-when a <0, Since a / 2> 0, it represents divergent vibration.

The differential equation (19) is characterized by a second-order AR model represented by the equation (21) (hereinafter referred to as “AR (2)”) [the differential equation (19) as a first-order simultaneous differential equation] Rewrite it. ]. Here, x (t) represents the signal x at time t, φ ₁ and φ ₂ represent AR coefficients, and ε (t) represents normal white noise.

When the characteristic root of AR (2) is λ, the characteristic root λ is given by Expression (22), and a signal satisfying AR (2) has a spectrum having a peak in the vicinity of f _{0 of} Expression (23). When the signal is an audio signal, the spectrum peak group represents a formant.

Assuming that the complex conjugate roots of equation (22) are λ ₁ and λ ₂ , when | λ _1,2 | ≦ 1, | λ _1,2 | . In addition, when | λ _1,2 |> 1, | λ _1,2 | Incidentally, lambda _{1, 2} becomes notation represents lambda ₁ or lambda ₂ (hereinafter the same).

The self-excited vibration of van der Po1 is expressed by the equation (24), and the first-order difference coefficient a (x) is a function of the signal x. The magnitude of the amplitude of the signal x determines whether the coefficient is positive or negative, and vibration attenuation / divergence is determined. In other words, the self-excited vibration of van der Po1 has an amplitude-dependent nonlinear model structure.

Model of the vocal cords vibration is represented by the equation of motion of the formula (25), a model structure of the amplitude-dependent which determines the state of the vibration in the magnitude of K _t. Equation (25) represents a damped vibration when K _t > 0, and a divergent vibration when K _t <0. K _t : energy loss time change within the pitch period, ω ₀ : resonance angular frequency at K _t = 0.

The ExpAR model is a non-linear model in which the amplitude-dependent term of the signal is represented by the signal x (t−1) one period before, as is clear from Expression (26). The coefficient part of the signal x (t−k) becomes an element of the transition matrix in the form of state space expression, and the information of the previous signal x (t−1) is sufficient to show the characteristics. Meaning. Equation (26) shows that the vibration characteristic changes depending on the magnitude of | x (t−1) |. When | x (t−1) | If it is small, it indicates divergent vibration. Here, p is the model order, φ _k and π _k are autoregressive coefficients, and γ is a scale parameter.

  In other words, the ExpAR model expresses a natural expression for the assumption of Markov stationarity in which past information is aggregated in the previous signal x (t−1) for explaining the current signal x (t). It is the nonlinear model shown.

Table 1 summarizes the relationship among the intrinsic root λ _1,2 of AR (2), K _t that determines the pitch energy attenuation, and x (t−1) of the amplitude dependence term of the ExpAR model. It can be said that the characteristics that change with attenuation or divergence according to the change in amplitude of the signal are described.

The condition for the stationarity of the ExpAR model is that all the characteristic roots of Equation (22) fall inside the unit circle. When the ExpAR model is applied to speech data, the coefficients φ ₁ and φ ₂ usually correspond to vocal tract transfer characteristics estimated by applying LPC, that is, formants. Further, from the assumption of Markov property, the ExpAR model is stationary even when the characteristic root of Equation (27) is outside the unit circle.

The autoregressive coefficient π _{k in} equation (26) indicates the strength of nonlinearity and determines the above-described attenuation or divergence. In addition, the scale parameter γ in the equation (26) is also a portion showing non-linearity. However, in some cases, π _k and γ may be in a mutually canceling relationship (for example, when π _k = 0). Therefore, FIG. 29 shows the relationship among certain audio data, autoregressive coefficient π _k , and scale parameter γ in the case of model order p = 2.

As shown in FIG. 29, the autoregressive coefficient π _k does not take a value of 0 in all frames (in FIG. 29, there is a portion where the autoregressive coefficient π _k becomes 0 for convenience of display scale, but strictly, Is not 0.) Since the scale parameter γ is small in the speech interval, that is, the nonlinearity is increasing, the ExpAR model captures the speech nonlinearity and can extract the speech interval. It is shown that.

In other words, in the case of a signal having nonlinear vibration characteristics such as an audio signal, all the autoregressive coefficients π _k are constantly set to 0 or a value that impairs the effect of the scale parameter γ that affects the nonlinearity. Therefore, it can be said that the ExpAR model is effective for modeling such a signal.

<Relationship between formant, autoregressive coefficient π _k and scale parameter γ>
The relationship between the formant relating to the vocal tract transmission system, the autoregressive coefficient π _{k of the} exponential term (π _k exp (−γx (t−1) ² )) indicating the nonlinearity of the equation (26), and the scale parameter γ is shown. .

Autoregressive coefficients φ ₁ , φ ₂ indicating stationarity [Refer to formula (22). ], The peak group of this spectrum corresponds to a formant estimated by LPC. These peaks show the average formant height and phase between certain frames. On the other hand, in the ExpAR model, a characteristic root as shown in the equation (27) can also be obtained. The behavior of this root depends on the amplitude of the signal and indicates a state in which the phase moves in and out of the unit circle or the phase changes in the unit circle. Such root behavior indicates the fluctuation of the height of the formant and the frequency position which fluctuate every moment between frames. Therefore, in order to show this simply, Equation (26) is modeled with φ _k , π _k , x (t−1), and ε (t) as fixed values, and autoregressive coefficient π and scale parameter γ are set as follows. The spectrum when set to various values is shown in FIG. As can be seen from FIG. 30, even when the autoregressive coefficient π is fixed and the scale parameter γ is changed, or when the scale parameter γ is fixed and the autoregressive coefficient π is changed, the spectrum changes. The formant height and frequency position are changing.

  An embodiment in which an audio signal section is detected using an acoustic feature obtained by the signal determination method of the present invention for an audio signal in which an audio signal and a noise signal are mixed will be described. The acoustic signal data used in this example is uttered by one male voice included in the noisy speech recognition evaluation working environment (AURORA-2J). A signal obtained by adding station noise to a signal reading speech with a signal-to-noise ratio of 10 dB and discretely sampled with a sampling frequency of 8 kHz and a quantization bit number of 16 bits was used. FIG. 28A shows an audio signal not including a noise signal, and FIG. 28B shows an audio signal added with the noise signal. This is the same as that shown in FIGS. For this acoustic signal, the signal determination method according to the present invention sets the time length of one frame to 25 ms (200 sample points), moves the start point of the frame every 0.125 ms (1 sample point), and takes an exponent of model order 8 The parameters of the function autoregressive model were estimated and the moving average was taken at 200 sample points. The output acoustic features (scale parameter γ) are shown in FIG. As shown in the figure, it can be seen that the acoustic feature output by the signal determination method according to the present invention shows a small value in the section where the audio signal exists, and shows a large value in the other sections. For this acoustic feature, an audio signal section was detected when the average value (dotted line in FIG. 28C) was set as a threshold value. When the acoustic feature value exceeds the threshold value, the voice signal section is set, and when the acoustic feature value falls below the threshold value, the voice signal section is set. FIG. 28 (d) shows the result when the audio signal interval is 1 and the non-audio signal interval is 0. As shown in the figure, it has been confirmed that the method of the present invention is effective in estimating the existence section of a speech signal that is a target signal even in an environment with a noise signal.

  The signal determination apparatus and method of the present invention can be used for acoustic signal processing in speech and musical sound encoding, noise signal suppression, dereverberation, automatic speech recognition, etc., taking the case where the signal is an acoustic signal as an example. .

(A) An example of a voice signal in which only a silent section and a voice section exist. (B) The value of the scale parameter γ calculated by the signal determination apparatus / method of the present invention for the audio signal illustrated in FIG. (C) A station noise added to the audio signal illustrated in FIG. 1A with a signal-to-noise ratio of 10 dB. (D) The value of the scale parameter γ calculated by the signal determination apparatus / method of the present invention for the audio signal illustrated in FIG. The block diagram which illustrated the hardware constitutions of the signal determination apparatus (1) concerning 1st Embodiment. The function block diagram of the signal determination apparatus (1) concerning 1st Embodiment. The flowchart which shows the flow of a process of the conventional autoregressive coefficient calculation method. The functional block diagram which shows the processing function in the autoregressive coefficient calculation method which the present inventors created. The flowchart which shows the flow of a process of the autoregressive coefficient calculation method which the present inventors created. The functional block diagram which shows the processing function in the autoregressive coefficient calculation method (modification 1) which the present inventors created. The flowchart which shows the flow of a process of the autoregressive coefficient calculation method (modification 1) which the present inventors created. The functional block diagram which shows the processing function in the autoregressive coefficient calculation method (modification 2) which the present inventors created. The flowchart which shows the flow of a process of the autoregressive coefficient calculation method (modification 2) which the present inventors created. The functional block diagram which shows the processing function in the autoregressive coefficient calculation method (modification 3) which the present inventors created. The flowchart which shows the flow of a process of the autoregressive coefficient calculation method (modification 3) which the present inventors created. The flowchart which shows the process of the signal determination method based on a threshold value. The flowchart which shows the process of the signal determination method based on a support vector machine. The function block diagram of the signal determination apparatus (1) in 2nd Embodiment. The function block diagram of the signal determination apparatus (2) of 3rd Embodiment. The flowchart which shows the flow of the signal determination process in the signal determination apparatus (2) of 3rd Embodiment. The function block diagram of the signal determination apparatus (3) of 4th Embodiment. The flowchart which shows the flow of the signal determination process in the signal determination apparatus (3) of 4th Embodiment. The function block diagram of the signal determination apparatus (4) of 5th Embodiment. The flowchart which shows the flow of the signal determination process in the signal determination apparatus (4) of 5th Embodiment. The function block diagram of the signal determination apparatus (5) of 6th Embodiment. The flowchart which shows the flow of the signal determination process in the signal determination apparatus (5) of 6th Embodiment. The function block diagram of the signal determination apparatus of 7th Embodiment. The figure which shows the structure and connection relation of HMM for target signal area recognition, and HMM for non-target signal area recognition. The figure which shows the correspondence of learning data, HMM for target signal area recognition, and HMM for non-target signal area recognition in learning. It is a figure which shows the signal area determination of the time series pattern of the scale parameter (gamma) by HMM (signal determination model) obtained by learning. (A) The same audio signal as in FIG. (B) The same acoustic signal as in FIG. (C) Acoustic features (scale parameter γ) calculated in the verification experiment by the signal determination method of the present invention. (D) A signal determination result based on the threshold value. The figure which showed the change of the autoregressive coefficient (pi) _k and scale parameter (gamma) value for every flame | frame at the time of applying the ExpAR model of model order p = 2 to audio | voice data. In the ExpAR model in which φ _k , π _k , x (t−1), and ε (t) are fixed values, a diagram showing spectra when the autoregressive coefficient π and the scale parameter γ are set to various values (however, , X (t-1) = 0.5).

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Signal determination apparatus 20 Autoregressive coefficient calculation part 21 Signal determination part 22 Normalization part 23 Smoothing part

Claims (15)

A signal determination device for determining whether a discrete signal is a target signal,
Storage means for storing discrete signals;
An autoregressive coefficient calculating means for obtaining and outputting an autoregressive coefficient of a nonlinear function autoregressive model expressing the discrete signal based on the discrete signal stored in the storage means;
A signal determining means for receiving the autoregressive coefficient output by the autoregressive coefficient calculating means, determining whether the discrete signal is the target signal based on the autoregressive coefficient, and outputting the signal determination result; A signal determination device comprising: The autoregressive coefficient calculation means is
2. The signal according to claim 1, wherein an autoregressive coefficient, which is an index in an exponential function autoregressive model expressing the discrete signal, is obtained and output based on the discrete signal stored in the storage means. Judgment device. The signal determination means is
A comparison determination between the autoregressive coefficient output by the autoregressive coefficient calculation means and a predetermined threshold is performed, and a signal determination result associated in advance for each comparison determination result is output based on the comparison determination result. The signal determination device according to claim 1, wherein the signal determination device is a signal determination device. The signal determination means is
For the autoregressive coefficient output by the autoregressive coefficient calculating means, pattern recognition based on the learning content obtained by learning with the autoregressive coefficient obtained using a known signal in advance is performed, and based on the pattern recognition result 3. The signal determination apparatus according to claim 1, wherein a signal determination result associated in advance for each pattern recognition result is output. The storage means
Storing discrete signals of a plurality of frames,
The autoregressive coefficient calculating means is:
For the discrete signals of a plurality of frames stored in the storage means, an autoregressive coefficient is calculated for each discrete signal of each frame,
Autoregressive coefficient normalizing means for normalizing the autoregressive coefficient of the discrete signal of each frame output by the autoregressive coefficient calculating means,
The signal determination means includes
5. The signal determination apparatus according to claim 1, wherein an autoregressive coefficient of the discrete signal of each frame normalized by the autoregressive coefficient normalizing means is input. The storage means
Storing discrete signals of a plurality of frames,
The autoregressive coefficient calculating means is:
For the discrete signals of a plurality of frames stored in the storage means, an autoregressive coefficient is calculated for each discrete signal of each frame,
Autoregressive coefficient normalizing means for normalizing the autoregressive coefficient of the discrete signal of each frame output by the autoregressive coefficient calculating means;
A normalized coefficient smoothing means for smoothing the autoregressive coefficient of the discrete signal of each frame normalized by the autoregressive coefficient normalizing means,
The signal determination means includes
5. The signal determination apparatus according to claim 1, wherein an autoregressive coefficient of the discrete signal of each frame smoothed by the normalized coefficient smoothing means is input. The storage means
Storing discrete signals of a plurality of frames,
The autoregressive coefficient calculating means is:
For the discrete signals of a plurality of frames stored in the storage means, an autoregressive coefficient is calculated for each discrete signal of each frame,
Autoregressive coefficient smoothing means for smoothing the autoregressive coefficient of the discrete signal of each frame output by the autoregressive coefficient calculating means,
The signal determination means includes
5. The signal determination apparatus according to claim 1, wherein an autoregressive coefficient of the discrete signal of each frame smoothed by the autoregressive coefficient smoothing means is input. The storage means
Storing discrete signals of a plurality of frames,
The autoregressive coefficient calculating means is:
For the discrete signals of a plurality of frames stored in the storage means, an autoregressive coefficient is calculated for each discrete signal of each frame,
Autoregressive coefficient smoothing means for smoothing the autoregressive coefficient of the discrete signal of each frame output by the autoregressive coefficient calculating means;
Smoothed coefficient normalizing means for normalizing the autoregressive coefficient of the discrete signal of each frame smoothed by the autoregressive coefficient smoothing means,
The signal determination means includes
5. The signal determination apparatus according to claim 1, wherein the autoregressive coefficient of the discrete signal of each frame normalized by the smoothed coefficient normalizing means is input. A signal determination method for determining whether a discrete signal is a target signal,
An autoregressive coefficient calculating means for obtaining and outputting an autoregressive coefficient of a nonlinear function autoregressive model expressing the discrete signal based on the discrete signal stored in the storage means;
The signal determination means includes a signal determination step of determining whether or not the discrete signal is a target signal based on the autoregressive coefficient output in the autoregressive coefficient calculation step, and outputting the signal determination result. A signal determination method characterized by the above. The autoregressive coefficient calculation step is
The signal determination method according to claim 9, wherein an autoregressive coefficient that is an index in an exponential function autoregressive model that expresses the discrete signal is obtained and output based on the input discrete signal. The signal determination step
A comparison determination between the autoregressive coefficient output in the autoregressive coefficient calculation step and a predetermined threshold is performed, and a signal determination result associated in advance for each comparison determination result is output based on the comparison determination result. The signal determination method according to claim 9 or 10, wherein: The signal determination step
For the autoregressive coefficient output in the autoregressive coefficient calculation step, pattern recognition based on the learning content obtained by learning with the autoregressive coefficient obtained using a known signal in advance is performed, and based on this pattern recognition result The signal determination method according to claim 9 or 10, wherein a signal determination result associated in advance for each pattern recognition result is output. A signal determination method for determining whether or not discrete signals of a plurality of frames are respective target signals,
An autoregressive coefficient calculating unit is configured to generate a non-linear function autoregressive model for expressing a discrete signal based on the discrete signal for each discrete signal of each frame for the discrete signal of a plurality of frames stored in the storage unit. An autoregressive coefficient calculating step for obtaining and outputting the autoregressive coefficient;
Based on the autoregressive coefficient of the discrete signal of each frame output in the autoregressive coefficient calculating step, the signal determining means determines whether or not the discrete signal of each frame is the target signal, and determines the signal determination result. And a signal determining step for outputting the signal.       A signal determination program for causing a computer to function as the signal determination device according to claim 1.       A computer-readable program recording medium on which the signal determination program according to claim 14 is recorded.

Images