JP4685735B2 - Acoustic signal section detection method, apparatus, program, and recording medium therefor - Google Patents

Description

  The present invention relates to a method for extracting a time interval in which a target signal exists from an acoustic signal in which a target speech such as a speech signal or a musical sound signal is mixed with a noise signal, and in particular, predicts background noise from an acoustic signal and its prediction error. The present invention relates to a method, an apparatus, a program, and a recording medium for extracting a time interval in which a target signal exists based on a signal.

  When extracting speech from conversations recorded in the background of noise in a car or station, or encoding speech generated in noise with a mobile phone, the noise is suppressed and high accuracy and high speed are achieved. Therefore, a technique for estimating a section where a target signal such as conversation exists is required. As a typical example of such a conventional technique, for example, a VAD (Voice Activity Detector) algorithm can be exemplified. In such a conventional acoustic section estimation technology, the feature amount is obtained by using the frequency spectrum of the signal, the entire band energy, the energy of each band after the band division, the number of zero crossings of the signal waveform and their time derivatives, etc. And a predetermined threshold are compared, and the acoustic section is estimated (for example, see Non-Patent Document 1). In addition, a method has been proposed in which background noise is once modeled by an autoregressive model or the like, and a target signal is estimated using an autocorrelation function or higher-order statistics for the prediction error (for example, non-patent literature). 2 and 3).

In addition, in the field of finance, a distributed heterogeneous model (see, for example, Non-Patent Document 4) that has been studied to classify volatility (fluctuation) of the price / earnings ratio has recently been used in the field of audio signal processing. (For example, see Non-Patent Documents 5, 6, and 7.) Specifically, when a GARCH model is applied to the linear prediction error sequence, a target signal such as a speech signal exists in a section where a nonuniform dispersion estimated from the GARCH model changes rapidly, and does not change rapidly. There has been a technique for detecting an acoustic section by determining that there is no target signal in the section.
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 Communication Magazine, pp. 64-73, September, 1997. Markel, JD "The SHIFT algorithm for fundamental frequency estimation", IEEE Trans.Audio Electroacoust., Vol.AU-20, pp.367-377, 1972. Nmer, E., Goubran, R., and Mahmoud, S. "Robust voice activity detection using higher-order statistics in the LPC residual domain", IEEE Trans.SAP, 9, 3, March, 217-231, 2001. Bollerslev, T. "Generalized autoregressive conditional heteroscedasticity", Journal of Econometrics, 51, pp.307-327, 1986. Cohen, I., "SuperGaussian GARCH models for speech signals", Proceedings of INTERSPEECH, pp.2053-2056, 2005. Abdolahi, M., and Amindavar, H. "GARCH coefficients as feature for speech recognition in Persian isolated digit," Proceedings of the 30th International Conference on Acoustics, Speech, and Signal Processing, vol. I, pp. 957-960, 2005 . Hiroko Kato and Kentaro Ishizuka, “Proposal of speech interval detection method using GARCH model”, Acoustical Society of Japan Spring Research Conference, 107-108, 2006.

  However, when applying a variance heterogeneous model to linear prediction errors, it is necessary to use a so-called numerical optimization method to estimate the parameters of the model, but the calculation process takes a very long time. There was a problem that it took.

According to the present invention, the linear prediction error calculation means obtains a linear prediction error sequence for an acoustic signal sampled at a predetermined sampling frequency. Autocorrelation function value calculating _means, and _{k min} and _{k max,} a k min _<k natural number relationship holds the _max, the k and _{k min} or _{k max} following a natural number, squaring the values of the linear prediction error sequence The autocorrelation function value for each lag k is obtained for each square series of the linear prediction error. The counting means determines whether or not the value of the autocorrelation function for each lag k is greater than a predetermined value, and outputs the number of times determined to be large. When the number of times output in the counting step is greater than a predetermined number, the determination unit determines that a sound signal or a musical sound signal exists in the acoustic signal, or the number of times output in the counting step. Is less than the predetermined number, it is determined that there is no voice signal or musical sound signal in the acoustic signal.

  By extracting features directly from the variance of the linear prediction error sequence without modeling the linear prediction error sequence, it is possible to reduce the amount of calculation and shorten the time with high accuracy. Section detection can be performed.

[principle]
If an acoustic signal sequence in which a target signal such as a voice signal or a musical tone signal is buried in a noise signal is y _n (n = 1,..., N), a noise signal sequence x _n (n = 1,..., N), an error sequence z _n (n = 1,..., N) is
y _n = x _n + z _n (A)
Satisfy the relationship.
Here, the background noise is a stationary signal as compared with a target signal such as a voice signal or a musical tone signal, and therefore can be described by an AR model of the following order M. For example, M≈6 is preferable.
x _n = Σ _{m = 1} ^M A _m x _n−m + ε _n (B)
Here, it is assumed that ε _n (n = 1,..., N) has an average of 0 and follows a normal distribution with unknown variance σ _n ² .

The parameter A _m (m = 1,..., M) in the AR model is obtained by using a method such as least square estimation. Then, the actually observed y _n, by subtracting the x _n to be estimated by the formula (B) using the parameters, the linear prediction error sequence _{z n (n = 1, ...} , N) Request .
Linear prediction error sequence _{z n (n = 1, ...} , N) squares sequence _z ⁿ 2 linear prediction error obtained by squaring the (n = 1, ..., N ) value of the autocorrelation function of the lag k for [rho _k
ρ _k = E (z _n ² z _n−k ² ) / (dispersion of z _n ² )
_Are determined for all natural numbers k in the range of, for example, _kmin <k < _kmax . The value of the autocorrelation function directly indicates the strength of the relationship between time points in the time series. In general, when k becomes larger than 0, the value of the autocorrelation function becomes closer to 0. The value of the sample autocorrelation function calculated from the data is not strictly 0, but most of those that fall within the range of ± 2 times the standard deviation of z _n ² (95% confidence interval) are regarded as 0. Suppose the sample autocorrelation function is not significant. However, it is assumed that the sample autocorrelation function has a certain tendency when there are many exceeding the range.

そして、すべてのｋについて上記の大小関係を比較した結果、ｃｏｕｎｔの値が多ければ、その音響区間には、音声信号や楽音信号等の目的音声が含まれていると判断することができる。 In the present invention, when there are many linear prediction error square series z _n ² (n = 1,..., N) exceeding a predetermined value c, it is considered that the tendency can be captured. The predetermined value c is, for example,
It is about twice the standard deviation of the value of the autocorrelation function for each lag k, or about 0.2 to 0.5. For example, the value ρ _{k of the} autocorrelation function of lag k is compared with a predetermined value c, and count is added as follows.

As a result of comparing the above magnitude relations for all k, if the count value is large, it can be determined that the target speech such as a voice signal or a musical tone signal is included in the acoustic section.

[Embodiment]
In FIG. 1, the example of a function structure of the acoustic area detection apparatus 77 is shown. FIG. 3 shows an example of processing of the acoustic section detection device 77.
The acoustic section detection device 77 includes, for example, the control unit 10, the storage unit 15, the signal reception unit 20, the signal extraction unit 25, the linear prediction error calculation unit 30, the autocorrelation function value calculation unit 35, the count unit 40, and the determination unit 45. Have The autocorrelation function value calculation unit 35 includes a square unit 351. The count unit 40 includes a comparison unit 401.
The storage unit 15 includes y _n (n = 1,..., N), A _m (m = 1,..., M), z _n (n = 1,..., N), z _n ² (n = 1, _{..., N), k min,} k max, k, ρ k, c, count, the area for storing the d is provided. _Assume that k _min , k _max , c, and d are preset and stored at desired values. For example, _kmin = 40 and _kmax = 200.
The storage unit 15 includes a control unit 10, a storage unit 15, a signal reception unit 20, a signal extraction unit 25, a linear prediction error calculation unit 30, an autocorrelation function value calculation unit 35, a count unit 40, a determination unit 45, and a data It is connected so that it can read and write.

<Step S0>
The signal receiving unit 20 of the acoustic section detecting device 77 receives an acoustic signal in which a voice signal, a musical sound signal, and a noise signal are mixed. The received acoustic signal is output to the signal cutout unit 25. Alternatively, the received acoustic signal is stored in the storage unit 15.
The signal extraction unit 25 samples the acoustic signal received by the signal receiving unit 20 or the acoustic signal read from the storage unit 15 with a predetermined sampling frequency and a predetermined number of quantization bits. For example, sampling is performed at a sampling frequency of 8 kHz and a quantization bit number of 16 bits. The signal cutout unit 25 further divides the sampled and discretized acoustic signal into frames. One of the frames, N samples _{y 1, ..., y n,} ..., consists of _{y N.} Hereinafter, a sequence of samples constituting one frame is referred to as an acoustic signal sequence y _n (n = 1,..., N). For example, the length of one frame is 250 ms, and is composed of 2000 samples y ₁ ,..., Y ₂₀₀₀ . The acoustic signal divided into frames is stored in the storage unit 15. The process of determining whether the received acoustic signal includes a target signal such as a voice signal or a musical tone signal is performed in units of frames.

<Step S1>
The linear prediction error calculation unit 30 reads out the acoustic signal sequence y _n (n = 1,..., N) constituting a certain frame from the storage unit 15 and linear prediction error sequence z _n (n = 1,..., N). Calculate
FIG. 2 is a diagram illustrating a functional configuration of the linear prediction calculation unit 30, and FIG. 5 is a diagram illustrating each process constituting step S1.
The parameter estimation unit 301 (FIG. 2) substitutes the acoustic signal sequence y _n (n = 1,..., N) in the above equation (B) instead of the noise signal sequence x _n (n = 1,..., N). By doing so, the parameter A _m (m = 1,..., M) is obtained. That is, the acoustic signal sequence y _n (n = 1,..., N) of the first frame generally does not include the target signal. Therefore, these acoustic signal sequences y _n (n = 1,..., N) can be regarded as noise signal sequences x _n (n = 1,..., N). Further, even if the acoustic signal sequence y _n (n = 1,..., N) used for estimating the parameter of the above formula (B) includes the target signal, the acoustic signal sequence y _n including the target signal. The acoustic section estimation method according to the present embodiment is used for the linear prediction error sequence z _n (n = 1,..., N) calculated by the parameters estimated based on (n = 1,..., N). It has been experimentally confirmed that it is possible to verify whether or not the target signal is included. For this reason, there is no problem even if the parameters of the above equation (B) are estimated using the acoustic signal sequence y _n (n = 1,..., N).
The parameter estimation unit 301 obtains the parameter A _m (m = 1,..., M) using, for example, a method based on least square estimation (step S11, FIG. 5). Hereinafter, a method of obtaining the parameter parameter A _m (m = 1,..., M) will be described using a method based on least square estimation.

上記式（Ｂ）は、下記のように記述することができる。
Ｘ＝Ｚθ＋Ε …（Ｃ）
最小二乗推定では、誤差ε_ｎの二乗の和を最小にするパラメータＡ_ｍ（ｍ＝１,…，Ｍ）を求める。すなわち、・^Ｔを、行列・の転置とすると、
Ｊ＝Σ_{ｎ＝Ｍ＋１} ^Ｎε_ｎ ^２＝（Ｘ−Ｚθ）^Ｔ（Ｘ−Ｚθ）
を最小にするθを求めることによって、パラメータＡ_ｍ（ｍ＝１,…，Ｍ）を求める。
具体的には、・^−１を、行列・の逆行列とすると、

という処理を行うことによって、パラメータ推定部３０１は、Ｊを最小にするθ’＝（Ａ_１,…,Ａ_Ｍ）^Ｔを求める（ステップＳ１１、図５）。計算されたパラメータＡ_ｍ（ｍ＝１,…，Ｍ）は、記憶部１５に格納される。 ≪Least squares estimation≫
If X, Z, θ, and E are defined as follows,

The above formula (B) can be described as follows.
X = Zθ + Ε (C)
In the least square estimation, a parameter A _m (m = 1,..., M) that minimizes the sum of squares of the error ε _n is obtained. That is, if ^T is the transpose of the matrix
J = Σ _{n = M + 1} ^N ε _n ² = (X−Zθ) ^T (X−Zθ)
By obtaining θ that minimizes the parameter A _m , the parameter A _m (m = 1,..., M) is obtained.
Specifically, when ⁻¹ is an inverse matrix of a matrix,

By performing the processing, the parameter estimation unit 301 obtains θ ′ = (A ₁ ,..., A _M ) ^T that minimizes J (step S11, FIG. 5). The calculated parameters A _m (m = 1,..., M) are stored in the storage unit 15.

The calculation of the parameter A _m (m = 1,..., M) can be omitted by using the AR model parameter A _m (m = 1,..., M) estimated in a certain frame as it is in other frames. May be.
The first residual calculation unit 302 calculates the linear prediction residual series z _n using the parameter A _m (m = 1,..., M) read from the storage unit 15 (step S12). Specifically, subtracting x _n (n = 1,..., N) obtained by the above equation (C) from the acoustic signal sequence y _n (n = 1,..., N) read from the storage unit 15. To obtain a linear prediction residual series z _n (n = 1,..., N). The calculated linear prediction residual series z _n (n = 1,..., N) is stored in the storage unit 15.
The Kalman filter unit 303 applies the Kalman filter using the parameters A _m (m = 1,..., M) obtained by the parameter estimation unit 301, so that the linear prediction residual sequence z _n (n = 1,. , N) may be obtained. By using the Kalman filter, there is a merit that the amount of the linear prediction residual z _n (n = 1,..., N) can be adjusted.

≪Kalman filter≫
Here, the expression (A) is expressed in the state space expression as follows.
Observation equation: y _n = Hξ _n + z _n (D)
Equation of state: ξ _n = Fξ _n−1 + γε _n (E)
Here, H is a unit matrix, z _n is a noise parameter N (0, δ) indicating the amplitude of noise according to a normal distribution with mean 0 and variance δ, and ξ _n = (x ₁ , x ₂ ,... x _m ) ^T , γ is γ = (1, 0,..., 0), and F is

の状態遷移行列である。
この場合、カルマンフィルタのアルゴリズムは以下のようになる。
一期先予測過程：
ξ_{ｎ｜ｎ−１}＝Ｆξ_{ｎ−１｜ｎ−１} …（Ｇ）
Ｖ_{ｎ｜ｎ−１}＝ＦＶ_{ｎ−１｜ｎ−１}Ｆ^Ｔ＋γＱγ^Ｔ …（Ｈ）
フィルタリング過程：
Ｋ＝Ｖ_{ｎ｜ｎ−１}Ｈ^Ｔ（ＨＶ_{ｎ｜ｎ−１}Ｈ^Ｔ＋Δ）^−１ …（Ｉ）
ξ_ｎ｜ｎ＝ξ_{ｎ｜ｎ−１}＋Ｋ（ｙ_ｎ‐Ｈξ_{ｎ｜ｎ−１}） …（Ｊ）
Ｖ_ｎ｜ｎ＝（Ｉ‐ＫＨ）Ｖ_{ｎ｜ｎ−１} …（Ｋ）
なお、・_α|βはβ時点の情報を用いて生成されたα時点の・を意味する。また、Δは、分散δを要素とする対角行列である。Δにおけるδの値が、そのまま線形予測誤差ｚ_ｎの分散σになるため、利用者は、このΔにおけるδの値を任意に設定することにより、所望の分散σを有する線形予測誤差ｚ_ｎを求めることができる。また、Ｉは単位行列を意味し、Ｖは共分散行列であり、Ｑは、システムノイズの分散であり、ＡＲモデルをフィッティングしたときに得られる予測誤差ε_ｎの分散である。式（Ｇ）〜（Ｋ）の処理を逐次的に繰り返すことにより、ξ_ｎ｜ｎ，Ｖ_ｎ｜ｎ（ｎ＝１，…，Ｎ）が得られる。

This is a state transition matrix.
In this case, the Kalman filter algorithm is as follows.
First term forecast process:
ξ _{n | n−1} = Fξ _{n−1 | n−1} (G)
_{V n | n-1 = FV} n-1 | n-1 F T + γQγ T ... (H)
Filtering process:
^{_{K = V n | n-1}} H T (HV n | n-1 H T + Δ) -1 ... (I)
ξ _{n | n} = ξ _{n | n−1} + K (y _n −Hξ _{n | n−1} ) (J)
V _{n | n} = (I−KH) V _{n | n−1} (K)
Note that · _{α | β} means · at time α generated using information at time β. Δ is a diagonal matrix having the variance δ as an element. Since the value of δ in Δ becomes the variance σ of the linear prediction error z _n as it is, the user can arbitrarily set the value of δ in Δ to obtain the linear prediction error z _n having a desired variance σ. Can be sought. Further, I means a unit matrix, V is a covariance matrix, Q is a variance of system noise, and is a variance of a prediction error ε _n obtained when fitting an AR model. By sequentially repeating the processes of equations (G) to (K), ξ _{n | n} , V _{n | n} (n = 1,..., N) can be obtained.

The first-term prediction unit 3031 of the Kalman filter unit 303 calculates the matrix F based on the above formula (F) using the parameter A _m (m = 1,..., M) read from the storage unit 15 (step S13). FIG. 5). The calculated matrix F is stored in the temporary storage unit 3033. The control unit 10 increments the value of n stored in the temporary storage unit 3033 by 1 (step S14).
The one-term ahead prediction unit 3031 reads n, F, ξ _{n−1 | n−1} from the temporary storage unit 3033, performs the calculation of the above-described equations (G) and (H), and ξ _{n | n−1} and V _{n | n-1} is obtained (step S15). In this embodiment, arbitrary initial values are used for ξ _{0 | 0} and V _{0 | 0} . It is assumed that ξ _{0 | 0} and V _{0 | 0} are set in advance and stored in the temporary storage unit 3033. The generated V _{n | n−1} is stored in the temporary storage unit 3033, and ξ _{n | n−1} is stored in the temporary storage unit 3033 and the storage unit 15.

Next, the filtering unit 3032 reads V _{n | n−1} from the temporary storage unit 3033, calculates the above-described formula (I), obtains K, and stores it in the temporary storage unit 3033. Further, the filtering unit 3032, xi] _n from the temporary storage unit 3033 _| reads _n-1, reads the acoustic signal series _{y n} from the storage unit 15, xi] _n by performing the calculation of the formula (J) _| seeking _n, one o'clock The data is stored in the storage unit 3033 and the storage unit 15. Further, the filtering unit 3032 reads K and V _{n | n−1} from the temporary storage unit 3033, calculates the equation (K), obtains V _{n | n} , and stores it in the temporary storage unit 3033 (step S16). ).
Then, the second residual calculation section 304, the sound signal from the storage unit 15 discrete mixing signal _{y n} and xi] _{n |} reads _n-1,
z _n = y _n −Hξ _{n | n−1} (L)
The linear prediction error sequence z _n is generated by the calculation of (Step S17). The generated linear prediction error signal z _n is stored in the storage unit 15.
Next, the control unit 10 determines whether or not n stored in the temporary storage unit 3033 is N. Here, if n = N is not satisfied, the control unit 10 stores n + 1 as a new n in the temporary storage unit 3033, and returns the process to step S15. On the other hand, if n = N, the process of step S1 is terminated.

<Step S2, FIG. 3>
Returning to the description of FIGS.
The square unit 351 of the autocorrelation function value calculation unit 35 obtains the linear prediction error sequence z _n ^{2 obtained} by squaring the linear prediction error sequence z _n (n = 1,..., N) read from the storage unit 15. . The square sequence z _n ² of the linear prediction error is stored in the storage unit 15.
<Step S2 '>
The control unit 10 stores k in the storage unit 15 as k = k _min .

例えば、Ｅ｛ｚ_ｎ ^２ｚ_ｎ−ｋ ^２｝は、ｚ_ｎ ^２の平均値を（ｚ_ｎ ^２）_ａｖｅとし、

を計算することにより求めることができる。 <Step S3>
The autocorrelation function value calculation unit 35 obtains the value ρ _{k of the} autocorrelation function of lag k for the square series z _n ² of the linear prediction error obtained by the square unit 351. The calculated ρ _k is stored in the storage unit 15.
The value ρ _{k of the} autocorrelation function is defined as follows, _where σ ′ ² is the variance of z _n ² and γ _k is the autocovariance of lag k.

For ^{_{example, E {z n 2 z n}} -k 2} _is the mean value of _z ^{n 2} and ^_(z n _{2) ave,}

Can be obtained by calculating.

<Step S4>
The control unit 10 compares the magnitude of k read from the storage unit 15 and k _max read from the storage unit 15. If k ≦ k _max , the process of step S5 is performed. If k> k _max , the process of step S6 is performed.

<Step S5>
The control unit 10 increments k read from the storage unit 15 by 1 and stores it in the storage unit 15. Then, the process of step S3 is performed.
Thus, by repeating the process of step S5 from step S3, the autocorrelation function value calculation unit _35, a _{k min} and _{k max,} a k min _<k natural number relationship holds the _max, the k _{k min} or more A value ρ _{k of an} autocorrelation function for each lag k is obtained for a square series z _n ² of linear prediction errors obtained by squaring each value of the linear prediction error series z _n with a natural number equal to or less than k _max .

<Step S6>
The control unit 10 sets k stored in the storage unit 15 to _kmin . Also, the count stored in the storage unit 15 is set to 0.

<Step S7>
The comparison unit 401 of the count unit 40 determines whether or not ρ _k read from the storage unit 15 is greater than a predetermined value c. The predetermined value c is, for example, a value _obtained by approximately doubling the standard deviation SE (ρ _k ) of the autocorrelation function value or a value of approximately 0.2 to 0.5. If ρ _k > c, a signal to that effect is output to the count unit 40. When the comparison unit 401 uses a value obtained by multiplying the standard deviation SE (ρ _k ) of the autocorrelation function value by about twice as a threshold value, the comparison unit 401 calculates the value prior to the above determination. There are various calculation methods for calculating the standard deviation SE (ρ _k ) (see, for example, Reference 1), but the standard deviation SE (ρ _k ) of the autocorrelation function value is the length of the time series. Since N is known to approximate (1 / N) ^1/2 , SE (ρ _k ) can be calculated using this approximate expression.
[Reference 1] Bartlett, MS, "On the theoretical specification and sampling properties of autocorrelated time-series", JRStat.Sci, 1946.

<Step S8>
When receiving a signal indicating that ρ _k > c, the count unit 40 increments the count value read from the storage unit 15 by one.
Note that the count unit 40 may increment the count value read from the storage unit 15 when ρ _k <c instead of ρ _k > c.
In other words, the count may be added as follows.

<Step S9>
The control unit 10 increments the value of k stored in the storage unit 15 by 1.
<Step S10>
The control unit 10 compares the magnitude of k read from the storage unit 15 and k _max read from the storage unit 15. If k ≦ k _max , the process of step S7 is performed. If k> k _max , the process of step S11 is performed.
Thus, by repeating the processing from step S7 to step S10, the counting unit 40 determines whether the value of the autocorrelation function for each lag k is greater than a predetermined value c, and determines that it is large. The number of times counted is output.

<Step S11 '>
When the count value read from the storage unit 15 is greater than the predetermined number d read from the storage unit 15, the determination unit 45 includes a sound signal or a musical sound signal in the acoustic signal, or the storage unit If the count value read from 15 is smaller than the predetermined number d read from the storage unit 15, it is determined that no sound signal or musical sound signal exists in the acoustic signal, and the determination result is output. The predetermined number d is appropriately set according to the accuracy of section detection required by the practitioner of the present invention. When it is desired to completely remove the noise, the predetermined number d can be set high. In addition, when the target signal is completely picked up while allowing noise to be mixed, the predetermined number d can be set low. The predetermined number d, for example, comparison unit 401, the number multiplied by about 0.3 to 0.5 to the number of times the comparing the size relationship of [rho _k and c. Prior to the above determination process, the determination unit 45 calculates k _max −k _min using k _max and _kmin read from the storage unit 15 to obtain the number of times that the magnitude relationship between ρ _k and c is compared. .
The determination unit 45, divided by the number of comparing the magnitude relation of the value of count [rho _k and c is greater than a certain value _{e 1,} the acoustic signal series _y n (n = 1, ... , N) includes an audio signal or a musical sound signal, or if the value obtained by dividing the value of count by the number of comparisons of the magnitude relationship between ρ _k and c is smaller than a certain value e ₁ , the acoustic signal sequence It may be determined that y _n (n = 1,..., N) does not include a voice signal or a musical sound signal, and the determination result may be output. The constant value e ₁ can be set as appropriate according to the accuracy required by the practitioner of the present invention, similarly to the predetermined number d. For example, the constant value e ₁ can be set to about 0.3 to 0.5.

Further, in the process of step S8, when the count unit 40 increments the count value read from the storage unit 15 by 1 when ρ _k <c instead of ρ _k > c, the determination unit 45 When the count value read from the storage unit 15 is larger than the predetermined number d read from the storage unit 15, there is no audio signal or musical sound signal in the acoustic signal, or read from the storage unit 15. When the count value is smaller than the predetermined number d read from the storage unit 15, it is determined that a sound signal or a musical sound signal exists in the acoustic signal, and the determination result is output. The predetermined number d in this case can also be set as appropriate in accordance with the interval detection accuracy required by the practitioner of the present invention. When it is desired to completely remove noise, the predetermined number d can be set low. In addition, when the target signal is completely picked up while allowing noise to be mixed, the predetermined number d can be set high. The predetermined number d, the number multiplied by about 0.5 to 0.7 to the number of times the comparing the size relationship of [rho _k and c.

In this case, if the value obtained by dividing the count value by the number of comparisons between the magnitudes of ρ _k and c is greater than a certain value e ₂ , the sound signal sequence y _n (n = 1,..., N) does not include an audio signal or musical sound signal, or if the value obtained by dividing the count value by the number of comparisons of the magnitude relationship between ρ _k and c is smaller than a certain value e ₂ , It may be determined that the sound signal series y _n (n = 1,..., N) includes a voice signal or a musical sound signal, and the determination result may be output. The constant value e ₂ can be set as appropriate according to the accuracy required by the practitioner of the present invention, similarly to the predetermined number d. For example, the constant value e ₂ can be set to about 0.5 to 0.7.
As described above, the sound section detection device 77 detects whether or not the target sound / musical sound signal is included in the observed sound signal.

[Modifications, etc.]
In the above embodiment, after obtaining the values [rho _k of the autocorrelation function for all lag k (step S3 to S5, FIG. 3), the magnitude relationship between the value [rho _k with a predetermined value c of the autocorrelation function I investigated. However, as shown in FIG. 4, each time the autocorrelation function value ρ _k for lag k is obtained, the comparison unit 401 of the counting unit 40 determines the magnitude relationship between the autocorrelation function value ρ _k and the predetermined value c. You may make it check.
Moreover, the processing function of the acoustic signal section detection device can be realized by a computer. In this case, the contents of the processing function of the acoustic signal section detection device are described by a program. Then, by executing this program on a computer as shown in FIG. 6, for example, each processing function of the acoustic signal section detecting device shown in FIG. 1 is realized on the computer.

The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.
The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, this 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 a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is provided 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 acoustic signal section detecting device is configured by executing a predetermined program on the computer. However, at least a part of these processing contents may be realized by hardware. .
In addition to the above embodiments, the acoustic signal section detection method, apparatus, program, and recording medium according to the present invention are not limited to the above-described embodiments, and can be appropriately changed without departing from the spirit of the present invention. Is possible.

[Experimental example]
1. Data used in the experiment The acoustic signal including background noise was obtained by adding environmental pressure sound with SNR 0 dB to a waveform discretized with a sampling frequency of 8 kHz and a quantization bit number of 16 bits generated from the ATR travel dialogue database. . The environmental noise was recorded independently on the street. The time frame size was shifted by 2000 points (250 ms), and the overlap of frames was shifted by 80 points (10 ms).

最も低いＦａｌｓｅ ＡｃｃｅｐｔとＦａｌｓｅ Ｒｅｊｅｃｔからグリッド単位ごとに（Ｆａｌｓｅ Ａｃｃｅｐｔ，Ｆａｌｓｅ Ｒｅｊｅｃｔ）を２次元座標にプロットするとＲＯＣ曲線を描くことができる。 2. Evaluation of section detection performance In order to examine the accuracy of section detection, a correct voice section in units of 10 ms based on transcription by a human was used. In order to show how much speech section is detected from the feature amount proposed by the present invention (autocorrelation function value of the residual square series), a Receiver Operating Characteristics (ROC) curve is applied (for example, reference Reference 2). In calculating the ROC curve, it is necessary to calculate False Accept and False Reject. False Accept and False Reject are defined as follows.

A ROC curve can be drawn by plotting (False Accept, False Reject) in two-dimensional coordinates for each grid unit from the lowest False Accept and False Reject.

[Reference 2]
Marzinzik, M., and Kollmeier, B. "Speech pause detection for noise spectrum estimation by tracking power envelope dynamics," IEEE Trans. Speech Audio Processing, vol. 10, pp. 109-118, 2002.
3. Result First, an AR model of about 6 dimensions is applied to an acoustic signal overlapping street background noise with an SN ratio of 0 dB, and an error sequence is extracted. For comparison, the sample autocorrelation function for the normal error sequence (only linear prediction error) and the proposed error square sequence is applied, and the number of counts in a certain lag range is obtained from the above equation (a). Here, the range of the lag to be noted is a range from 50 Hz to 200 Hz where the fundamental frequency of voice is expressed, and is set to 40 to 200. In addition, the evaluation is also performed for 250 to 1950 in the range of 4 Hz to 32 Hz (for example, refer to Reference 3) in which the amplitude variation characteristic corresponding to the syllable speed of the voice that affects the intelligibility of the voice appears. This is because the frequency characteristic of the amplitude fluctuation of the speech can be considered as a physical phenomenon corresponding to the causality of the dispersion fluctuation, which is a feature captured by the proposed method.

[Reference 3]
Greenberg, S. "Temporal properties of spoken language," Proceedings of the 18th International Congress. On Acoustics, vol. I, pp. 441-444, 2004.
FIG. 7 shows a target acoustic signal (noise-added speech signal), a noiseless speech signal (noise-free speech signal), and a correct speech section for reference. The diagram shown in the upper part of FIG. 7 is a noise-added speech signal waveform including street background noise with an S / N ratio of 0 dB. The diagram shown in the middle part is a speech signal waveform without noise. Show time. In the vertical axis of the diagram shown in the lower part of FIG. 7, 1 indicates the correct answer and 0 indicates the incorrect answer. FIG. 8 shows an ROC curve focused on the range of lugs 40 to 200, and FIG. 9 shows an ROC curve focused on the range of lugs 250 to 1950. The solid line shows the result of using the method according to the present invention, and the broken line shows the result of the method using only the conventional linear prediction error. The vertical axis represents False Reject, and the horizontal axis represents False Accept. According to FIG. 8 and FIG. 9, the lower False Accept and False Reject are obtained by using the autocorrelation function value of the linear prediction error square series than by using the autocorrelation function value of the linear prediction error as in the conventional method. It can be realized that the method according to the present invention is an effective method with few false detections.

The figure which illustrates the function structure of the acoustic area detection apparatus 77. The figure which illustrates the function structure of the linear prediction error calculation part. The figure which illustrated the process of the acoustic area detection apparatus 77. The figure which illustrated another process of the acoustic area detection apparatus 77. The figure which illustrates the process of the linear prediction error calculation part 30 in step S1. The figure which assists description of how to obtain | require a target track position. The figure which shows the noise addition audio | voice, the noiseless audio | voice, and the correct audio | voice area used for the experiment for verifying the effectiveness of this invention. The figure which shows the ROC curve which paid its attention to the range of the lugs 40-200. The figure which shows the ROC curve which paid its attention to the range of lugs 250-1950.

Claims (13)

A linear prediction error calculating means for calculating a linear prediction error sequence for an acoustic signal sampled at a predetermined sampling frequency;
Autocorrelation function value calculating _means, and _{k min} and _{k max,} a k min _<k natural number relationship holds the _max, the k and _{k min} or _{k max} following a natural number, squaring the values of the linear prediction error sequence An autocorrelation function value calculating step for obtaining a value of an autocorrelation function for each lag k with respect to a square series of the linear prediction error,
A counting step in which the counting means determines whether or not the value of the autocorrelation function for each lag k is greater than a predetermined value, and outputs the number of times determined to be large;
When the number of times output in the counting step is greater than a predetermined number, the determination unit determines that a sound signal or a musical sound signal exists in the acoustic signal, or the number of times output in the counting step. Is less than a predetermined number, a determination step of determining that there is no voice signal or musical sound signal in the acoustic signal;
A method for detecting an acoustic signal section. The acoustic signal section detection method according to claim 1,
The linear prediction error calculation step includes:
A parameter estimation step of applying an AR model to the acoustic signal and calculating a parameter in the AR model;
A first residual calculation step for obtaining a linear prediction error sequence by subtracting a linear prediction value calculated using the above parameters from the acoustic signal;
Composed of,
An acoustic signal section detection method characterized by the above. The acoustic signal section detection method according to claim 1,
The linear prediction error calculation step includes:
A parameter estimation step of applying an AR model to the acoustic signal and calculating a parameter in the AR model;
A Kalman filter step in which the Kalman filter means performs the Kalman filter process using the parameter and the variance value of the desired linear prediction error;
A second residual calculation means for calculating a linear prediction error sequence using the result of the Kalman filter processing;
Composed of,
An acoustic signal section detection method characterized by the above. The acoustic signal section detection method according to any one of claims 1 to 3,
The determination step determines that an audio signal or a musical sound signal exists in the acoustic signal if a value obtained by dividing the number of times output in the counting step by the determined number of times is equal to or greater than a first predetermined value. Or means for determining that there is no audio signal or musical sound signal in the acoustic signal if a value obtained by dividing the number of times output in the counting step by the determined number of times is equal to or less than a first predetermined value. Is,
An acoustic signal section detection method characterized by the above. The first constant value according to claim 4 is 0.3 to 0.5.
An acoustic signal section detection method characterized by the above. The acoustic signal section detection method according to any one of claims 1 to 3,
The counting step is a step of determining whether or not the value of the autocorrelation function for each lag k is smaller than a predetermined value, and outputting the number of times determined to be small,
In the determination step, when the number of times output in the counting step is less than a predetermined number, it is determined that a sound signal or a musical sound signal exists in the acoustic signal, or the output in the counting step When the number of times is greater than a predetermined number, it is a step of determining that there is no audio signal or musical sound signal in the acoustic signal.
An acoustic signal section detection method characterized by the above. It is an acoustic signal section detection method according to claim 5,
The determination step determines that a sound signal or a musical sound signal exists in the acoustic signal if a value obtained by dividing the number of times output in the counting step by the determined number of times is equal to or less than a second predetermined value. Or means for determining that no audio signal or musical sound signal exists in the acoustic signal if a value obtained by dividing the number of times output in the counting step by the determined number of times is equal to or greater than a second predetermined value. Is,
An acoustic signal section detection method characterized by the above. The second constant value according to claim 7 is 0.5 to 0.7.
An acoustic signal section detection method characterized by the above. The acoustic signal section detection method according to any one of claims 1 to 8,
The predetermined value is approximately twice the standard deviation of the value of the autocorrelation function for each lag k.
An acoustic signal section detection method characterized by the above. The acoustic signal section detection method according to any one of claims 1 to 8,
The predetermined value is 0.2 to 0.5.
An acoustic signal section detection method characterized by the above. Linear prediction error calculation means for obtaining a linear prediction error sequence for an acoustic signal sampled at a predetermined sampling frequency;
and k _min and _{k max,} a k min _<k natural number relationship holds the _max, the k and _{k min} or _{k max} following a natural number, for the square series of linear prediction error obtained by squaring the values of the linear prediction error sequence , Autocorrelation function value calculating means for determining the value of the autocorrelation function for each lag k,
Counting means for determining whether or not the value of the autocorrelation function for each lag k is greater than a predetermined value, and outputting the number of times determined to be large;
When the number of times output by the counting means is greater than a predetermined number, it is determined that a sound signal or a musical sound signal exists in the acoustic signal, or the number of times output by the counting means is a predetermined number. If there are less than, determination means for determining that there is no audio signal or musical sound signal in the acoustic signal,
An acoustic signal section detecting device having   An acoustic signal section detection program for causing a computer to execute each step of the acoustic signal section detection method according to claim 1.   A computer-readable recording medium on which the acoustic signal section detection program according to claim 12 is recorded.

Images