JP2011095509A - Acoustic signal analysis device, acoustic signal analysis method and acoustic signal analysis program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a technique to solve the problem of conventional arts which do not accurately detect characteristics of music. <P>SOLUTION: The acoustic signal analysis device 1 includes: an acquisition part 12 for acquiring an acoustic signal 2; a first feature quantity-calculating part 13 for calculating a first value of each sound volume of a plurality of sections including a first section of the acoustic signal 2; a second feature quantity-calculating part 14 for calculating a second value regarding each sound volume in a plurality of sections including a second section which is longer than the first section of the acoustic signal 2; an evaluation value-calculating part 15 for calculating an evaluation value of time sequence by using the first value and the second value, which becomes larger, as the first value is larger, and as a second value corresponding to the first value in time is larger; and a feature position-detecting part 16 for detecting a section where the evaluation value calculated by the evaluation value-calculating part 15 becomes a maximum, or a local maximum. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a technique for analyzing an acoustic signal.

  In recent years, a large amount of music data stored in a computer storage medium or the like has been widely used. Along with this, there is an increasing need for technology for easily and quickly grasping the contents of each of a large amount of stored music data. As one of the techniques, a technique for detecting music rust and excitement as a point of listening to music has been proposed.

  For example, Patent Document 1 discloses a technique for detecting a position where the volume is maximum in music data and reproducing a specific portion of the music data including the position. Patent Document 2 discloses a technique for detecting the degree of excitement and excitement of music using the ratio of the output values of the filters of the high band, the medium band, and the low band.

JP 2007-80304 A JP2003-228387A

  Although it is possible to detect a characteristic part of a music piece using the above-described conventional technique, there are various kinds of music pieces. There has been a demand for an acoustic signal analyzer that can detect errors and detect characteristic portions of music with higher accuracy.

  The present invention has been made in view of such problems, and an object of the present invention is to provide an acoustic signal analyzer and the like that can detect a characteristic portion of a music piece with high accuracy.

  In order to solve the above problems and achieve the above object, an acoustic signal analyzer of the present invention includes an acquisition unit that acquires an acoustic signal, and a plurality of sections having a first period of the acoustic signal acquired by the acquisition unit. A first volume information calculation unit that calculates a first value related to each volume, and a volume of each of a plurality of sections having a second period longer than the first period of the acoustic signal acquired by the acquisition unit; Using a second volume information calculation unit that calculates a second value, the first value, and the second value, the larger the first value, the longer the first value. An evaluation value calculation unit that calculates a time-series evaluation value that increases as the corresponding second value increases, and a position at which the evaluation value calculated by the evaluation value calculation unit is maximum or maximum is detected. And a feature position detection unit.

  Moreover, the acoustic signal analysis method of the present invention includes a step of acquiring an acoustic signal, a step of calculating a first value related to a volume of each of a plurality of sections having a first period of the acquired acoustic signal, and an acquired acoustic signal Calculating a second value relating to the volume of each of a plurality of sections having a second period longer than the first period of the signal, and using the first value and the second value, A step of calculating a time-series evaluation value that becomes a larger value as the value of is larger and the second value temporally corresponding to the first value is larger, and the calculated evaluation value is maximum or maximum Detecting a section that becomes.

  Furthermore, a program for causing a computer to realize the functions of the constituent elements of the acoustic signal analyzer of the present invention is also an aspect of the present invention.

  The present invention can provide an acoustic signal analyzing apparatus and the like that detect characteristic portions of music with high accuracy.

1 is a diagram illustrating a configuration of an acoustic signal analysis device according to a first embodiment. It is a figure which shows the relationship between frame time length Tf1 and frame shift time length Tg1. 4 is a flowchart illustrating steps of an operation of a first feature amount calculation unit of the acoustic signal analysis device according to the first embodiment. 6 is a flowchart illustrating steps of an operation of a second feature amount calculation unit of the acoustic signal analysis device according to the first embodiment. 4 is a flowchart illustrating steps of an operation of an evaluation value calculation unit of the acoustic signal analysis device according to the first embodiment. 4 is a flowchart illustrating steps of an operation of a feature position detection unit of the acoustic signal analysis device according to the first embodiment. It is a figure which shows the condition where an evaluation value changes with progress of time. It is a schematic diagram which shows the mode of a change of the 1st feature-value E1 calculated using a comparatively short section length. It is a schematic diagram which shows the mode of the change of the 2nd feature-value E2 calculated using comparatively long section length. It is a schematic diagram when the sum (E1 + E2) of the first feature value and the second feature value is used as an evaluation value. It is a schematic diagram when the product (E1 × E2) of the first feature value and the second feature value is used as an evaluation value. It is a figure which shows the structure of the acoustic signal analyzer of Embodiment 2. FIG. 6 is a flowchart illustrating steps of an operation of a beat time detection unit of the acoustic signal analysis device according to the second embodiment. It is a figure which shows the example of an autocorrelation. It is a figure which shows distribution of the existence probability of the time length of a beat. FIG. 6 is a configuration diagram of an acoustic signal analysis device according to a third embodiment. 10 is a flowchart illustrating steps of an operation of a frequency band data calculation unit of the acoustic signal analysis device according to the third embodiment. It is a figure which shows a frequency spectrum. 10 is a flowchart illustrating steps of an operation of a feature position detection unit of the acoustic signal analysis device according to the third embodiment. It is a figure which shows the condition where the width of a frequency band changes with progress of time. FIG. 10 is a configuration diagram of an acoustic signal analysis device according to a fourth embodiment. 10 is a flowchart showing steps of an operation of an evaluation value calculation unit of the acoustic signal analysis device according to the fourth embodiment. FIG. 10 is a configuration diagram of an acoustic signal analysis device according to a fifth embodiment. 10 is a flowchart illustrating steps of an operation of a volume data calculation unit of the acoustic signal analyzer of the fifth embodiment. FIG. 10 is a configuration diagram of an acoustic signal analysis device according to a sixth embodiment. 16 is a flowchart illustrating steps of an operation of a beat time detection unit of the acoustic signal analysis device according to the sixth embodiment.

  EMBODIMENT OF THE INVENTION Below, the form for implementing this invention is demonstrated with reference to drawings.

(Embodiment 1)
First, the acoustic signal analyzer 1 of Embodiment 1 will be described with reference to FIG. FIG. 1 is a configuration diagram of an acoustic signal analyzer 1 according to the first embodiment. As shown in FIG. 1, the acoustic signal analysis device 1 according to the first embodiment includes a control unit 11, an acquisition unit 12, a first feature value calculation unit 13, a second feature value calculation unit 14, and an evaluation. A value calculation unit 15 and a feature position detection unit 16 are included.

  The acoustic signal analyzer 1 acquires the acoustic signal 2 and outputs the characteristic position information 3.

  The acoustic signal 2 is an acoustic signal related to music. The acoustic signal 2 may be a digital signal or an analog signal. The sound signal 2 may be a signal including sound other than music such as DJ in addition to music, such as an audio signal of a music program such as radio or television, instead of a signal only of music. The acoustic signal 2 exists outside the acoustic signal analyzer 1. However, if the storage unit is provided in the acoustic signal analyzer 1, the acoustic signal 2 may be stored in the storage unit and exist inside the acoustic signal analyzer 1.

  The feature position information 3 is information for identifying a portion where the “total volume” of the acoustic signal 2 is large. The part often coincides with a part where the rust position of the music or the composition of the music or the organization of the musical instrument changes greatly, that is, a characteristic part of the music.

  The control unit 11 of the acoustic signal analysis device 1 controls each unit by exchanging information with other units constituting the acoustic signal analysis device 1.

  The acquisition unit 12 acquires the acoustic signal 2 and generates PCM (Pulse Code Modulation) data sampled at the sampling period Ts (sampling frequency Fs = 1 / Ts) from the acquired acoustic signal 2. When the acoustic signal 2 is an analog signal, the acquisition unit 12 converts the analog signal into a digital signal to generate PCM data. When the acoustic signal 2 is a digital compressed signal other than PCM, the acquisition unit 12 decodes the digital compressed signal. To generate PCM data. When the acoustic signal 2 is a digital signal and the sampling period is different from the sampling period Ts, the acquisition unit 12 converts the sampling rate and generates PCM data with the sampling period Ts.

  In the following description, PCM data generated by the acquisition unit 12 is described as acoustic data x [m] (m = 0 to M−1, where M is the total number of samples of acoustic data) or acoustic data. When the acquisition unit 12 finishes generating the acoustic data, the acquisition unit 12 notifies the control unit 11 to that effect.

  The first feature value calculation unit 13 calculates a first feature value related to sound volume from the acoustic data generated by the acquisition unit 12. The first feature amount calculation unit 13 calculates a feature amount related to the volume of a relatively short time interval. The first feature amount calculation unit 13 performs processing in units of frames. However, the unit of processing is not limited thereto.

  In the following, it is assumed that the time length of each frame processed by the first feature quantity calculation unit 13 is Tf1, and the time length of the frame shift is Tg1. At this time, the frame sample number N1 = Tf1 / Ts, and the frame shift sample number G1 = Tg1 / Ts. The frame shift is a time difference between the heads of adjacent frames. Adjacent frames may partially overlap or may not overlap.

  The time length of the frame and the time length of the frame shift will be described with reference to FIG. FIG. 2 is a diagram illustrating the relationship between the frame time length Tf1 and the frame shift time length Tg1. FIG. 2A shows a case where adjacent frames do not overlap and there is no gap between frames. FIG. 2B is a diagram illustrating a case where adjacent frames partially overlap. In this case, Tf1> Tg1. FIG. 2C is a diagram illustrating a case where there is a gap between adjacent frames. In this case, Tf1 <Tg1.

  The first feature quantity calculation unit 13 starts the operation shown in the flowchart of FIG. 3 according to the instruction of the control unit 11. FIG. 3 is a flowchart showing each step of the operation of the first feature quantity calculation unit 13.

  First, the first feature amount calculation unit 13 first calculates the total number H1 of frames according to the following equation (1) (S100).

ｆｌｏｏｒ（ ）は、小数点以下を切り捨てた整数を返す関数である。ＭとＮ１との関係は、Ｍ＞Ｎ１である。

floor () is a function that returns an integer with the decimal part truncated. The relationship between M and N1 is M> N1.

  Next, the first feature quantity calculation unit 13 sets “0” to the control variable i (S110).

  Next, the first feature amount calculation unit 13 generates i-th frame data (S120). The i-th frame data is data from the acoustic data x [i × G1] to the acoustic data x [i × G1 + N1-1]. The first feature amount calculation unit 13 generates, as the i-th frame data, a value obtained by multiplying the data from the acoustic data x [i × G1] to the acoustic data x [i × G1 + N1-1] by a window function. May be. The window function is a Hamming window function, a Hanning window function, a Blackman window function, a Gauss window function, or the like. It can be said that the method not using the window function is the same as the method of generating the i-th frame data by multiplying the acoustic data by the rectangular window.

  When a window function is used, the window function coefficient is usually maximized at the center of the frame and the window function coefficient is minimized at the beginning and end of the frame, but other methods may be used. For example, the window function coefficient is maximized at the beginning of the frame (x [i × G1]), then the window function coefficient is sequentially decreased, and the window function coefficient at the end of the frame (x [i × G1 + N1-1]). May be minimized. The i-th frame data is described as “D1 [i] [j] (j = 0 to ND1, where ND1 = N1-1)”.

  Next, the first feature value calculation unit 13 calculates the first feature value of the i-th frame using any of the methods described later (S130).

  Next, the first feature amount calculation unit 13 increases the value of the control variable i by “1” (S140).

  Next, the first feature amount calculation unit 13 determines whether or not the value of the control variable i is less than H1 (S150). If the value of the control variable i is less than H1 (Yes in S150), the first feature amount calculation unit 13 returns to step S120 and repeats the processing up to step S140. If the value of the control variable i is H1, (No in S150), the process ends.

  In this way, the first feature amount calculation unit 13 calculates the H1 time-series data E1 [i] (i = 0 to H1-1), which is the first feature amount related to the volume, and the processing is completed. This is notified to the control unit 11.

  Next, a method in which the first feature value calculation unit 13 calculates the first feature value of the i-th frame in step S130 will be described.

  (1) The first feature amount calculation method uses an absolute value of the amplitude of acoustic data. Specifically, as shown in the following equation (2), a value (sum) obtained by adding the absolute value of the amplitude by the number of samples of the frame is set as a feature amount E1 [i] corresponding to the i-th frame.

なお、下記の式（３）に示すように、総和の代わりに平均値を用いてもよい。

As shown in the following formula (3), an average value may be used instead of the sum.

（２）特徴量の第２の算出方法は、音響データの振幅の２乗を用いる方法である。具体的には、下記の式（４）に示すように、振幅の２乗の値をフレームのサンプル数だけ加算した値（総和）を、ｉ番目のフレームに対応する特徴量Ｅ１［ｉ］とする。

(2) The second feature amount calculation method uses the square of the amplitude of acoustic data. Specifically, as shown in the following equation (4), a value (sum) obtained by adding the square of the amplitude by the number of samples of the frame is defined as a feature amount E1 [i] corresponding to the i-th frame. To do.

なお、下記の式（５）に示すように、総和の代わりに平均値を用いてもよい。また、式（４）又は式（５）の右辺の平方根をとった値を、ｉ番目のフレームに対応する特徴量Ｅ１［ｉ］としてもよい。第１及び第２の算出方法は、計算量が少なくなるという効果が得られる。

As shown in the following formula (5), an average value may be used instead of the sum. Further, a value obtained by taking the square root of the right side of Expression (4) or Expression (5) may be used as the feature amount E1 [i] corresponding to the i-th frame. The first and second calculation methods have the effect of reducing the amount of calculation.

（３）特徴量の第３の算出方法は、特定の周波数成分を用いる方法である。ｉ番目のフレームデータＤ１［ｉ］［ｊ］に対して離散フーリエ変換（ＤＦＴ)を行い、出力の実数部Ｒｅ［ｋ］と虚数部Ｉｍ［ｋ］（ｋ＝０〜（Ｎ１／２））とを用いて、下記の式（６）又は式（７）式により、特徴量Ｅ１［ｉ］を算出する。

(3) The third feature amount calculation method uses a specific frequency component. The discrete Fourier transform (DFT) is performed on the i-th frame data D1 [i] [j], and the real part Re [k] and the imaginary part Im [k] (k = 0 to (N1 / 2)) of the output And the feature quantity E1 [i] is calculated by the following formula (6) or formula (7).

式（６）は、音響データの振幅スペクトルの特定の周波数成分を用いて特徴量Ｅ１［ｉ］を算出するための式であり、式（７）は、音響データのパワースペクトルの特定の周波数成分を用いて特徴量Ｅ１［ｉ］を算出するための式である。これらの式において、ＦＬは利用する周波数成分の下限を示す所定の定数であり、ＦＨは利用する周波数成分の上限を示す所定の定数であって、０≦ＦＬ≦ＦＨ≦Ｎ１／２の関係が満たされており、ＦＬとＦＨとの間の周波数成分の総和を算出して特徴量Ｅ１［ｉ］としている。ＦＬ及びＦＨは、例えば、高い周波数成分（例えば８ＫＨz以上）が除外されるように設定される。なお、下記の式（８）に示すように、周波数成分毎に定められた重み係数ｗ［ｋ］をスペクトルの周波数成分と掛け合わせて特徴量Ｅ１［ｉ］を算出してもよい。

Expression (6) is an expression for calculating the feature quantity E1 [i] using a specific frequency component of the amplitude spectrum of the acoustic data, and Expression (7) is a specific frequency component of the power spectrum of the acoustic data. Is a formula for calculating the feature quantity E1 [i]. In these equations, FL is a predetermined constant indicating the lower limit of the frequency component to be used, FH is a predetermined constant indicating the upper limit of the frequency component to be used, and the relationship 0 ≦ FL ≦ FH ≦ N1 / 2 is satisfied. The sum of the frequency components between FL and FH is calculated and used as the feature amount E1 [i]. FL and FH are set such that, for example, high frequency components (for example, 8 kHz or more) are excluded. Note that, as shown in the following equation (8), the feature amount E1 [i] may be calculated by multiplying the weighting factor w [k] determined for each frequency component by the frequency component of the spectrum.

  In the third method, only a specific frequency component is selected. Thereby, compared with the case where all frequency components are used, the effect of improving the correspondence between the feature amount and the sense of volume felt by humans can be obtained. In particular, by setting the weighting coefficient w [k] for each frequency component according to the auditory characteristics, a feature value close to a volume feeling can be obtained.

上述した第３の算出方法では、離散フーリエ変換（ＤＦＴ)を用いるが、これに限定される訳ではなく、ＤＦＴに代えて、例えば、デジタルフィルタやアナログフィルタを用いて特定の周波数成分を抽出してもよい。

In the third calculation method described above, discrete Fourier transform (DFT) is used. However, the present invention is not limited to this. For example, a specific frequency component is extracted using a digital filter or an analog filter instead of DFT. May be.

  (4) A fourth feature amount calculation method is a method in which the i-th frame data is divided into two partial sections (groups) before and after in time, and a difference in numerical values related to sound volume calculated for each partial section is used. It is. Numerical values related to the volume of each partial section are calculated using the first to third calculation methods described above.

  As an example, a case in which the first feature amount calculation method is used will be described. First, the i-th frame data D1 [i] [j] (j = 0 to ND1) is divided into two partial sections before and after in time. The partial interval 1 before in time is described as “Da [i] [j] (j = 0 to N1 / 2-1)”, and the partial interval 2 after in time is referred to as “Db [i] [j]. (J = N1 / 2 to ND1) ”. Next, the data of each of the partial section 1 and the partial section 2 are substituted into Expression (2). However, in Expression (2), the range of addition of j = 0 to ND1 is changed to the start point and end point of each partial section. The result of substituting the previous partial interval 1 into the expression (2) in terms of time is Ea [i], and the result of substituting the subsequent partial interval 2 in the expression (2) is defined as Eb [i]. The difference is defined as a feature quantity E1 [i]. That is, E1 [i] = Eb [i] −Ea [i] is calculated as the feature amount.

  When E1 [i] becomes a negative value, a process for setting the feature amount to “0” may be performed. In the above example, there is no gap between the partial section 1 and the partial section 2, but there may be a gap between the partial section 1 and the partial section 2. Further, a part of the partial section 1 and a part of the partial section 2 may overlap.

  Further, as described above, frame data may be created using a Hamming window or a Gaussian window. At this time, the difference may be calculated after matching the boundary point that divides the two partial sections with the center point (location where the coefficient is maximum) such as a Hamming window or a Gaussian window. In this case, the acoustic data closer to the boundary between the two partial sections is weighted with a larger weighting factor. That is, when the partial section is divided before and after the N1 / 2th sample data as in the above example, the acoustic data corresponding to (N1 / 2-1) and N1 / 2 closest to the boundary is multiplied by the largest coefficient. In addition, the difference is calculated by multiplying the acoustic data corresponding to 0 and ND1 farthest from the boundary by the smallest coefficient.

  (5) A fifth feature amount calculation method uses a difference in numerical values indicating the volume of two adjacent frames. The numerical value indicating the volume of the frame is a feature amount obtained by any one of the first calculation method to the third calculation method. For example, when using the feature amount obtained by the first calculation method, the calculation result obtained by substituting the acoustic data corresponding to the (i−1) -th frame into Equation (2) is E1 ′ [i−1]. In addition, the calculation result obtained by substituting the acoustic data corresponding to the i-th frame into Equation (2) is held as E1 ′ [i]. Then, the difference between E1 '[i] and E1' [i-1] is calculated. That is, E1 [i] = E1 ′ [i] −E1 ′ [i−1] is calculated as the feature amount. In addition, when the fourth and fifth calculation methods are used, an effect that it is easy to detect a portion where the volume rapidly changes can be obtained.

  In the above first to fifth calculation methods, for example, the obtained data may be normalized so that the maximum value of the feature amount is 1 and the minimum value is 0.

  Here, the characteristics of the volume of the acoustic signal related to music will be described. The volume of the sound signal related to music is closely related to the multi-layered structure of music such as individual notes, decorative sounds of notes such as tremolo and vibrato, beats, measures, phrases, intros and rusts. In such a multi-layered structure of music, ornamental sounds of notes such as tremolo and vibrato and individual notes bring about a very short time unit volume change, whereas the big composition of music such as intro and rust is Bring volume changes in very long time units. In the first embodiment, ornamental sounds of notes such as tremolo and vibrato and individual notes cause a change in volume in a very short time unit, and a large composition of music such as intro and rust causes a change in volume in a very long time unit. It focuses on a feature that has not been considered in the past.

  For example, the characteristic amount related to the volume of the sound signal related to music is the position where the volume is maximum in the case where the volume is calculated for each section of 1 second and in the case where the volume is calculated for each section of 10 seconds. It can be quite different. If a position where the volume is maximized for each type of section is detected as in the prior art, a characteristic location such as rust is often erroneously detected. On the other hand, the acoustic signal analysis device 1 according to the first embodiment calculates a feature amount related to the sound volume for two types of sections having different lengths as described below.

  The first feature amount calculation unit 13 sets the frame time length Tf1 so as to detect a relatively short time volume in the multi-layered structure described above. For example, the first feature amount calculation unit 13 sets the time length Tf1 of the frame to a time length substantially corresponding to the length of one note.

  For example, if the music has a 4/4 time signature, the tempo of general music is often in the range of about 60 to 240 beats per minute, from a 16th note (1/4 beat) to a whole note (4 (Beats) is often used. Since the time length of the notes in the range is from 62.5 msec (16th notes at tempo 240) to 4 sec (all notes at tempo 60), the first feature amount calculation unit 13 The frame time length Tf1 is set in the range. For example, the first feature amount calculation unit 13 sets 500 msec corresponding to a quarter note at the tempo 120 as the time length Tf1 of the frame.

  Next, the second feature amount calculation unit 14 will be described. The second feature amount calculation unit 14 calculates a second feature amount related to volume.

  The second feature value calculation unit 14 calculates a second feature value related to the sound volume from the acoustic data generated by the acquisition unit 12. The second feature amount calculation unit 14 calculates a feature amount related to the volume of a relatively long time interval. The time length Tf2 of the frame processed by the second feature value calculation unit 14 is longer than the time length Tf1 of the frame processed by the first feature value calculation unit 13. The number N2 of frame samples processed by the second feature quantity calculation unit 14 is N2 = Tf2 / Ts, which is larger than the number N1 of frame samples processed by the first feature quantity calculation unit 13.

  The frame shift time length Tg2 when the second feature value calculation unit 14 operates may be the same as the frame shift time length Tg1 when the first feature value calculation unit 13 operates. , May be different. In the following, the frame shift time length Tg2 = Q × Tg1, and the number of frame shift samples G2 = Q × G1 when the second feature amount calculator 14 operates (Q is an integer of 1 or more). However, the frame shift time length Tg2 and the frame shift sample number G2 are not limited thereto.

  The second feature quantity calculation unit 14 starts the operation shown in the flowchart of FIG. 4 in accordance with an instruction from the control unit 11. FIG. 4 is a flowchart showing each step of the operation of the second feature amount calculation unit 14. As is apparent from a comparison between FIG. 4 and FIG. 3, the second feature quantity calculation unit 14 operates in the same manner as the first feature quantity calculation unit 13.

  First, the second feature quantity calculation unit 14 calculates the total number H2 of frames using the above equation (1) (S200). That is, the second feature amount calculation unit 14 replaces N1 in Equation (1) with N2, replaces G1 with G2, replaces H1 with H2, and calculates the total number H2 of frames (S200). The relationship between M and N2 is M> N2. The total number of frames H2 processed by the second feature quantity calculation unit 14 is equal to or less than the total number of frames H1 processed by the first feature quantity calculation unit 13.

  Next, the second feature quantity calculation unit 14 sets “0” to the control variable i (S210).

  Next, the second feature amount calculation unit 14 generates i-th frame data (S220). The i-th frame data is data from acoustic data x [i × G2] to acoustic data x [i × G2 + N2-1]. Note that the second feature amount calculation unit 14 generates, as the i-th frame data, a value obtained by multiplying the data from the acoustic data x [i × G2] to the acoustic data x [i × G2 + N2-1] by a window function. May be. The window function is a Hamming window function, a Hanning window function, a Blackman window function, a Gauss window function, or the like. It can be said that the method not using the window function is the same as the method of generating the i-th frame data by multiplying the acoustic data by the rectangular window.

  When a window function is used, the window function coefficient is usually maximized at the center of the frame and the window function coefficient is minimized at the beginning and end of the frame, but other methods may be used. For example, the window function coefficient is maximized at the beginning of the frame (x [i × G2]), then the window function coefficient is sequentially decreased, and the window function coefficient at the end of the frame (x [i × G2 + N1-1]). May be minimized. The i-th frame data is described as “D2 [i] [j] (j = 0 to ND2, where ND2 = N2-1)”.

  The head D1 [i × Q] [0] of the (i × Q) -th frame data processed by the first feature value calculation unit 13 and the i-th frame data processed by the second feature value calculation unit 14 The heads D1 [i] [0] coincide with each other as x [i × G2]. However, the heads of the frames do not necessarily have to be matched in this way. For example, the center of the frame may be matched, or the end of the frame may be matched.

  Next, the second feature quantity calculation unit 14 calculates the feature quantity of the i-th frame data as if the first feature quantity calculation unit 13 calculated the feature quantity of the i-th frame data (S230). ). The second feature amount calculation unit 14 replaces ND1 used by the first feature amount calculation unit 13 with ND2, and replaces D1 with D2 to calculate the feature amount.

  Next, the second feature amount calculation unit 14 increases the value of the control variable i by “1” (S240).

  Next, the second feature amount calculation unit 14 determines whether or not the value of the control variable i is less than H2 (S250). If the second feature quantity calculation unit 14 determines that the value of the control variable i is less than H2 (Yes in S250), the process returns to Step S220 and repeats the process up to Step S240, and the value of the control variable i is H2. If it is determined that there is any (No in S250), the process is terminated.

  The second feature amount calculation unit 14 calculates H2 time-series data E2 [i] (i = 0 to H2-1), which is a feature amount related to volume, by the above-described processing, and indicates that the processing is completed. Notify the control unit 11.

  Next, the time length Tf2 of the frame processed by the second feature quantity calculation unit 14 will be described. As described above, in general music, the volume changes due to the multi-layered structure having various time scales. The second feature amount calculator 14 sets the frame time length Tf2 so as to detect a relatively long sound volume. For example, the second feature amount calculation unit 14 sets the frame time length Tf2 to a length of one bar or more.

  In particular, rust, which is the point of listening to music, is often repeated in units of about 4 to 8 bars, and there is a high possibility that the volume is high from about 4 to 8 bars from the start position of the rust. Since the general music tempo often ranges from 60 beats to 240 beats per minute, the second feature quantity calculation unit 14 sets the frame time length Tf2 to 4 seconds to 32 seconds corresponding to 4 bars. Set to the range. For example, the second feature amount calculation unit 14 sets 8 seconds corresponding to 4 bars at the tempo 120 as the frame time length Tf2.

  The evaluation value calculation unit 15 uses the first feature value calculated by the first feature value calculation unit 13 and the second feature value calculated by the second feature value calculation unit 14 to evaluate the evaluation value. Is calculated. The evaluation value calculation unit 15 calculates the evaluation value so that the larger the first feature amount and the larger the second feature amount corresponding to the first feature amount, the larger the value.

  When the control unit 11 detects the end of the processes of the first feature value calculation unit 13 and the second feature value calculation unit 14, the control unit 11 instructs the evaluation value calculation unit 15 to start the operation. The evaluation value calculation unit 15 starts the operation shown in the flowchart of FIG. FIG. 5 is a flowchart showing each step of the operation of the evaluation value calculation unit 15.

  The evaluation value calculation unit 15 first sets “0” to the control variable i (S300).

  Next, the evaluation value calculation unit 15 calculates a value to be set in the control variable j according to the following equation (9) (S310).

ｆｌｏｏｒ（ ）は、小数点以下を切り捨てた整数を返す関数である。Ｑは、第１の特徴量算出部１３が動作する際のフレームシフト時間長に対する、第２の特徴量算出部１４が動作する際のフレームシフト時間長の倍率であり、１以上の整数である。

floor () is a function that returns an integer with the decimal part truncated. Q is a magnification of the frame shift time length when the second feature quantity calculation unit 14 operates with respect to the frame shift time length when the first feature quantity calculation unit 13 operates, and is an integer of 1 or more. .

  Next, the evaluation value calculation unit 15 calculates an evaluation value α [i] corresponding to the control variable i according to a method described later (S320).

  Next, the evaluation value calculation unit 15 increases the value of the control variable by “1” (S330).

  Next, the evaluation value calculation unit 15 determines that the control variable i is H2 (the number of time-series data of feature amounts calculated by the second feature amount calculation unit 14) and Q (Q × H2). ) Is determined (S340). If the evaluation value calculation unit 15 determines that the control variable i is less than (Q × H2) (Yes in S340), the process returns to Step S310 and repeats the processing up to Step S330, and the control variable i is (Q × H2). If determined to be (No in S340), the process is terminated.

  The evaluation value calculation unit 15 calculates the evaluation value α [i] (i = 0 to Q × H2-1), which is (Q × H2) pieces of time series data, by the above-described processing, and ends the processing. Is notified to the control unit 11.

  The evaluation value calculation unit 15 calculates the evaluation value α [i] by any of the following methods.

  (1) As shown in the following formula (10), the first evaluation value calculation method uses a feature quantity E1 [i] calculated by the first feature quantity calculation unit 13 and a feature quantity E1 [i]. Is added to the feature quantity E2 [j] calculated by the second feature quantity calculation unit 14 corresponding to the time.

なお、特徴量Ｅ１［ｉ］と特徴量Ｅ１［ｉ］に時間的に対応する特徴量Ｅ２［ｊ］とを加算した値に所定値を乗算した値を評価値としてもよい。

Note that a value obtained by multiplying a feature value E1 [i] and a feature value E2 [j] temporally corresponding to the feature value E1 [i] by a predetermined value may be used as the evaluation value.

  (2) The second calculation method of the evaluation value temporally corresponds to the value obtained by multiplying the feature value E1 [i] by the coefficient β1 and the feature value E1 [i] as shown in the following equation (11). This is a method using an addition value of a value obtained by multiplying the feature amount E2 [j] by a coefficient β2. However, β1> 0 and β2> 0. In the second calculation method, the feature amount E1 and the feature amount E2 are respectively weighted and added.

（３）評価値の第３の算出方法は、下記の式（１２）に示すように、特徴量Ｅ１［ｉ］の対数値に係数β１を乗じた値と、特徴量Ｅ１［ｉ］に時間的に対応する特徴量Ｅ２［ｊ］の対数値に係数β２を乗じた値との加算値を用いる方法である。なお、第１から第３の算出方法は、Ｅ１とＥ２のどちらかが小さい箇所で、評価値をあまり小さくしたくない場合に用いる。第３の算出方法は、それに加えて、Ｅ１とＥ２のそれぞれの値の範囲が大きく異なる場合に適している。

(3) As the third calculation method of the evaluation value, as shown in the following equation (12), a value obtained by multiplying the logarithmic value of the feature quantity E1 [i] by the coefficient β1 and the feature quantity E1 [i] are timed. This is a method using an addition value of a value obtained by multiplying a logarithmic value of the corresponding feature quantity E2 [j] by a coefficient β2. The first to third calculation methods are used when one of E1 and E2 is small and it is not desired to make the evaluation value too small. In addition to this, the third calculation method is suitable when the ranges of the values of E1 and E2 are greatly different.

（４）評価値の第４の算出方法は、下記の式（１３）に示すように、特徴量Ｅ１［ｉ］と特徴量Ｅ１［ｉ］に時間的に対応する特徴量Ｅ２［ｊ］との積を用いる方法である。なお、式（１３）の右辺にさらに所定値を乗算した値を評価値としてもよい。

(4) A fourth calculation method of the evaluation value is as follows, as shown in the following equation (13), the feature quantity E1 [i] and the feature quantity E2 [j] temporally corresponding to the feature quantity E1 [i] This method uses the product of Note that a value obtained by further multiplying the right side of Expression (13) by a predetermined value may be used as the evaluation value.

（５）評価値の第５の算出方法は、下記の式（１４）に示すように、特徴量Ｅ１［ｉ］を基数としてγ１を指数とした累乗値と、特徴量Ｅ１［ｉ］に時間的に対応する特徴量Ｅ２［ｊ］を基数としてγ２を指数とした累乗値との積を用いる方法である。第４及び第５の方法は、Ｅ１とＥ２のどちらかが小さければ、評価値も小さくしたい場合に用いる。第５の方法は、それに加えて、Ｅ１とＥ２の評価値への影響力に重みを付けたい場合に適している。なお、式（１４）の右辺にさらに所定値を乗算した値を評価値としてもよい。

(5) The fifth calculation method of the evaluation value is as follows. As shown in the following equation (14), the characteristic value E1 [i] is a radix and γ1 is an exponent, and the characteristic value E1 [i] is time. This is a method of using a product of a characteristic value E2 [j] corresponding to a base and a power value with γ2 as an index. The fourth and fifth methods are used when it is desired to reduce the evaluation value if either E1 or E2 is small. In addition, the fifth method is suitable when it is desired to weight the influence on the evaluation values of E1 and E2. Note that a value obtained by further multiplying the right side of Expression (14) by a predetermined value may be used as the evaluation value.

（６）評価値の第６の算出方法は、下記の式（１５）に示すように、特徴量Ｅ１［ｉ］を基数としてγ１を指数とした累乗値と係数β１の積と、特徴量Ｅ１［ｉ］に時間的に対応する特徴量Ｅ２［ｊ］を基数としてγ２を指数とした累乗値と係数β２の積との和を用いる方法である。なお、式（１５）の右辺にさらに所定値を乗算した値を評価値としてもよい。

(6) The sixth method for calculating the evaluation value is as shown in the following equation (15): the product of the power value with the characteristic amount E1 [i] as the base and γ1 as the exponent and the coefficient β1, and the characteristic amount E1 This is a method using the sum of a product of a power value with a characteristic value E2 [j] temporally corresponding to [i] and γ2 as an index and a coefficient β2. Note that a value obtained by further multiplying the right side of Expression (15) by a predetermined value may be used as the evaluation value.

評価値算出部１５は、Ｅ１［ｉ］≧θ１かつＥ２［ｊ］≧θ２（θ１、θ２は所定値）の条件が成立する場合、上述した第１から第６の算出方法を用いて評価値を算出し、その条件が成立しない場合、評価値を「０」に設定してもよい。また、評価値算出部１５は、評価値α［ｉ］を計算した後に、α［ｉ］＜θ３（θ３は所定値）である場合、α［ｉ］を「０」にする処理を行ってもよい。

When the condition of E1 [i] ≧ θ1 and E2 [j] ≧ θ2 (θ1 and θ2 are predetermined values) is satisfied, the evaluation value calculation unit 15 uses the first to sixth calculation methods described above to evaluate the evaluation value If the condition is not satisfied, the evaluation value may be set to “0”. Further, after calculating the evaluation value α [i], the evaluation value calculation unit 15 performs a process of setting α [i] to “0” when α [i] <θ3 (θ3 is a predetermined value). Also good.

  The evaluation value α [i] calculated by the above-described method is the second feature amount E2 corresponding to the first feature amount E1 [i] in terms of time as the first feature amount E1 [i] is larger. The larger the value [j], the larger the value. In the multi-layered structure of music with various time scales, the time interval of the first feature value corresponds to a time scale such as one note or one beat, and the time interval of the second feature value is Corresponds to time scales longer than one measure. In the “characteristic position” of the song, such as the start position of the chorus, the changing point where the melody changes greatly, the position suitable for the audition, or the position that gives a strong impression to the listener, the volume at the beginning is large and 4 to 8 measures from there. During this period, the average sound volume is often large, and the evaluation value of such a portion is a large value. Therefore, by detecting the maximum value or the maximum value of the evaluation values, it is possible to accurately detect a feature position such as a rust start position.

  When the control unit 11 detects that the process of the evaluation value calculation unit 15 has been completed, the control unit 11 instructs the feature position detection unit 16 to start the operation.

  The feature position detection unit 16 detects a characteristic position such as a rust start position in the acoustic signal 2 using the evaluation value calculated by the evaluation value calculation unit 15.

  The feature position detection unit 16 detects the feature position using one of the following methods.

  (1) The first feature position detection method is a method for detecting a frame (position) having the maximum evaluation value. Among the evaluation values α [i] (i = 0 to Q × H2-1), the largest evaluation value is searched for and the index Imax corresponding to it is detected. Then, the time corresponding to Imax (Tg1 × Imax) is set as the feature position.

  Instead of searching for the maximum value for all the calculated evaluation values, the range for searching for the maximum value may be limited. That is, the position where the evaluation value calculated by the evaluation value calculation unit 15 is maximum may be detected for a continuous portion of the acoustic signal 2. Specifically, the maximum value may be searched for α [i] (i = H3 to H4, where H3 and H4 are integers satisfying 0 ≦ H3 <H4 <Q × H2-1). For example, H3 = 0 and H4 is set to a value corresponding to about 70% of the music length. Further, the first feature value E1 and the second feature value E2 are calculated from the acoustic signal 2 corresponding to a continuous part of the song, for example, about 70% of the song, and the evaluation value α calculated based on these is the maximum. May be detected. When the method of searching for the maximum value for the evaluation value corresponding to a continuous part of the acoustic signal 2 is used as described above, the processing amount can be reduced and the feature position detection accuracy can be improved for the following reason. Can do.

  The rust of a song is often repeated multiple times in one song, but the nuances of performance and singing are not always the same each time and are often slightly different. That is, the rust at the front position in time is not completely swelled compared to the rust at the back position, and there are many cases where there is still a margin in the degree of swell. Considering the case of playing a part of a song for audition, it is desirable that the part should be a part that makes the listener think "I want to listen to this whole piece of music". Rust in the front position in time that gives a sense of expectation to the future excitement is more suitable for audition than rust in the rear position in time. By limiting the range in which the maximum evaluation value is detected to about 70% of the first half of the music piece, rust at the previous position is easily detected, and the detection accuracy of the characteristic position for trial listening is improved.

  Also, an appropriate positive value may be set in H3 so that the intro part of the music is not included in the feature position detection target. The position at which the evaluation value is maximum is not used as the feature position, but a position that is a predetermined time before the position at which the evaluation value is maximum, or a position that is before the position at which the evaluation value is maximum. Only the position where the evaluation value becomes smaller may be set as the feature position. As a result, it is possible to prevent detection omission of rust out. The first detection method is suitable when it is desired to detect one feature position in the music. When the first detection method is used, an effect that the processing amount is reduced can be obtained.

  (2) The second feature position detection method is a method for detecting a position where the evaluation value is maximum according to the flowchart shown in FIG. FIG. 6 is a flowchart showing the steps of the operation in which the feature position detector 16 executes the second feature position detection method.

  The feature position detector 16 first sets an initial value “H5” in the control variable i (S400). H5 is a predetermined integer that satisfies 1 ≦ H5 <Q × H2-2. When searching for the maximum position for all the evaluation values calculated by the evaluation value calculation unit 15, H5 = 1 is set. If the intro of music is not included in the maximum position detection target, H5> 1.

  Next, the feature position detection unit 16 determines whether α [i] is a maximum value (S410). This determination method is, for example, a method of determining α [i] as a maximum value if α [i]> α [i−1] and α [i]> α [i + 1]. If the characteristic position detection unit 16 determines that α [i] is a maximum value (Yes in S410), the evaluation value α [i] at the maximum position and the value (index, time information) i of the control variable at the maximum position. Is stored in the working memory inside the feature position detector 16 (S420).

  Next, the feature position detection unit 16 increases the value of the control variable i by “1” (S430). If the characteristic position detection unit 16 determines in step S410 that α [i] is not the maximum value (No in S410), the feature position detection unit 16 increases the value of the control variable i by “1” (S430).

  Next, the feature position detection unit 16 determines whether or not the control variable i is equal to or less than the predetermined value H6 (S440). H6 is a predetermined integer that satisfies H5 <H6 <Q × H2-1. When searching for the maximum position for all the evaluation values calculated by the evaluation value calculation unit 15, H6 = Q × H2-2 is set, and the portion after the music is excluded from the detection target of the maximum position for the reasons described above. In this case, H6 <Q × H2-2, for example, a value corresponding to 70% of the length of the music. If the characteristic position detection unit 16 determines that the control variable i is equal to or less than the predetermined value H6 (Yes in S440), the feature position detection unit 16 returns to Step S410 and repeats the processing up to Step S430.

  When the characteristic position detection unit 16 determines that the control variable i has exceeded the predetermined value H6 (No in S440), the characteristic position detection unit 16 selects a predetermined number of local maximum positions from the local maximum information stored in the work memory ( S450). For example, the feature position detection unit 16 selects a predetermined number of maximum positions in descending order. The P maximum positions (time) selected in descending order of evaluation value are described as Ip [v] (v = 0 to P−1). At this time, α [Ip [0]] ≧ α [Ip [1]] ≧ α [Ip [2]] ≧. . . ≧ α [Ip [P-1]]. For example, when the evaluation value changes with the passage of time as shown in FIG. 7, the feature position detection unit 16 has a maximum position A having the maximum value, a maximum position B having the second value, and a value of 3 And the local maximum position C is selected.

  Note that the feature position detection unit 16 may exclude those that are close in time to the already selected maximum positions when selecting a predetermined number of maximum positions in descending order of the maximum values. For example, the feature position detection unit 16 selects only the maximum value that is separated from the already selected maximum position by a predetermined time or more. In addition, the feature position detection unit 16 may detect a position where the evaluation value calculated by the evaluation value calculation unit 15 is maximum for a continuous part of the acoustic signal 2. The above is the description of the second feature position detection method. The second detection method is suitable when it is desired to detect a plurality of feature positions from the music.

  The feature position detection unit 16 outputs the maximum position Imax or the maximum position Ip [v] (v = 0 to P−1) of the evaluation value detected in this way as feature position information 3 to the outside of the acoustic signal analyzer 1. To do. By reproducing the acoustic signal 2 using the characteristic position information 3, it is possible to reproduce a characteristic portion such as a rust of the music.

  The acoustic signal analysis device 1 according to Embodiment 1 described above detects a characteristic location using two different section lengths. Below, the effect is demonstrated using FIGS. 8-11.

  FIG. 8 is a schematic diagram illustrating a change in the first feature amount E1 calculated using a relatively short section length. The horizontal axis in FIG. 8 indicates the frame number (time). In FIG. 8, the section from frame number 8 to frame number 16 is the chorus section. In general, the volume of the chorus section tends to be higher than that of other parts. However, even in the chorus section, the volume may be slightly reduced due to a break in the vocal, as indicated by point X in frame number 10 in FIG. Also, as in the point S of frame number 2 in FIG. 8, the volume may be instantaneously high except in the chorus section where a percussion instrument is played strongly or a vocal shout or the like enters. . In such a case, when the maximum feature amount position is detected as the feature position, an S point that is not actually a chorus section is detected as a chorus section. That is a false detection.

  FIG. 9 is a schematic diagram illustrating a change state of the second feature amount E2 calculated using a relatively long section length. The frame number in FIG. 9 and the frame number in FIG. 8 correspond to each other, and the same frame number indicates the same position. Also in FIG. 9, the section from frame number 8 to frame number 16 is the chorus section. As is apparent from a comparison between FIG. 8 and FIG. 9, the second feature quantity E2 shown in FIG. 9 changes more gently than the first feature quantity E1.

  The point S having the maximum value in FIG. 8 is not so large in FIG. In the rust section, the second feature amount E2 often takes a large value. The second feature amount E2 may be maximized in the middle of the chorus section instead of at the head of the chorus section. In the example of FIG. 9, the second feature amount E2 has the maximum at the Y point of the frame number 12. The location is included in the chorus section, but is not the head of the chorus section (frame number 8).

  As the trial listening start position of the music, it is most desirable to detect the head (T point) of the chorus section. However, if one section length is used, even if the section length is short as shown in FIG. Thus, even if the section length is long, the head of the chorus section may not be detected.

  FIG. 10 is a schematic diagram when the sum (E1 + E2) of the first feature value and the second feature value is used as the evaluation value. FIG. 10 shows the same range as FIG. 8 and FIG. In FIG. 10, the sum of the feature values (E1 + E2) is relatively large at the S point other than the chorus section and the Z point (frame number 13) in the middle of the chorus section, but at the leading T point of the chorus section. Maximum.

  FIG. 11 is a schematic diagram when the product (E1 × E2) of the first feature value and the second feature value is used as the evaluation value. FIG. 11 shows the same range as FIG. 8 to FIG. In FIG. 11, the product (E1 × E2) of the feature quantity is relatively large at the S point other than the chorus section and the Y point (frame number 12) in the middle of the chorus section, but at the beginning of the chorus section. Maximum at point T.

  As is apparent from FIGS. 10 and 11, by calculating the evaluation value by combining the feature amounts having different section lengths, the detection accuracy of the chorus section (the head of the chorus section) is improved. Therefore, in order to detect the feature position with high accuracy, the acoustic signal analysis device 1 according to Embodiment 1 detects the feature position by calculating an evaluation value by combining feature amounts having different section lengths.

  In the first embodiment, two types of feature amounts are calculated using two types of time length sections, and an evaluation value is calculated using them. However, the present invention is not limited to this. For example, three or more types of feature amounts may be calculated using three or more types of time length sections, and an evaluation value may be calculated using them.

(Embodiment 2)
Next, the acoustic signal analyzer 1 according to the second embodiment will be described with reference to FIG. FIG. 12 is a configuration diagram of the acoustic signal analyzer 1 according to the second embodiment. As shown in FIG. 12, the acoustic signal analysis device 1 according to the second embodiment includes a control unit 11, an acquisition unit 12, a first feature value calculation unit 13, a second feature value calculation unit 14, and an evaluation. A value calculation unit 15, a feature position detection unit 16, and a beat time detection unit 17 are included.

  The acoustic signal analysis device 1 according to the second embodiment includes a beat time detection unit 17 in addition to the components included in the acoustic signal analysis device 1 according to the first embodiment. This is the difference between the first embodiment and the second embodiment.

  When the control unit 11 detects that the acoustic data is generated by the acquisition unit 12, the control unit 11 determines whether the first feature value calculation unit 13 and the second feature value calculation unit 14 are instructed to start the operation. Instructs the time detector 17 to start the operation.

  The beat time detector 17 performs processing in units of frames. The time length of the frame processed by the beat time detection unit 17 is Tf3, and the time length of the frame shift when the beat time detection unit 17 operates is Tg3. The frame sample number N3 processed by the beat time detection unit 17 is N3 = Tf3 / Ts, and the frame shift sample number G3 is G3 = Tg3 / Ts. In order to accurately calculate the beat time, Tf3 and Tg3 are set to a time considerably shorter than the length of one beat. In general music, the tempo is 60 to 240 and the time length of one beat is often in the range of 250 msec to 1 sec. Therefore, Tf3 and Tg3 are set to appropriate values in the range of about 5 msec to 50 msec. The

  The beat time detector 17 performs processing according to the flowchart shown in FIG. FIG. 13 is a flowchart showing each step of the operation of the beat time detection unit 17.

  First, the beat time detection unit 17 calculates the total number H7 of frames using Equation (1) (S500). Specifically, the beat time detection unit 17 replaces N1 in Equation (1) with N3, replaces G1 with G3, replaces H1 with H7, and calculates the total number H7 of frames.

  Next, the beat time detection unit 17 sets “0” in the control variable i (S510).

  Next, the beat time detection unit 17 generates i-th frame data (S520). Specifically, the beat time detection unit 17 generates acoustic data x [i × G3 + N3-1] as the i-th frame data from the acoustic data x [i × G3]. The beat time detection unit 17 may generate a value obtained by multiplying the data from the acoustic data x [i × G3] to the acoustic data x [i × G3 + N3-1] by a window function as the i-th frame data. Good. The window function is a Hamming window function, a Hanning window function, a Blackman window function, a Gauss window function, or the like. It can be said that the method described first is the same as the method of generating the i-th frame data by multiplying the acoustic data by a rectangular window. The i-th frame data is described as “D3 [i] [j] (j = 0 to ND3, where ND3 = N3-1)”.

  Next, the beat time detector 17 calculates the feature amount of the i-th frame (S530). Specifically, the beat time detection unit 17 calculates the feature amount by using the fourth or fifth calculation method used when the first feature amount calculation unit 13 calculates the feature amount. In other words, the beat time detection unit 17 calculates the difference E3 [i] within the frame or between the frames by using the amplitude of the acoustic data or the specific frequency component of the acoustic data.

  Next, the beat time detecting unit 17 increases the value of the control variable i by “1” (S540).

  Next, the beat time detector 17 determines whether or not the value of the control variable i is less than H7 (S550). When determining that the value of the control variable i is less than H7 (Yes in S550), the beat time detection unit 17 returns to step S520 and repeats the processing up to step S540.

  When determining that the value of the control variable i is H7 (No in S550), the beat time detection unit 17 calculates the autocorrelation of the feature amount E3 [i] (i = 0 to H7-1) (S560). The beat time detector 17 calculates the autocorrelation Y (Δ) according to the following equation (16) while sequentially changing the autocorrelation index difference Δ within a predetermined tempo range.

Ｈ８は、０≦Ｈ８＜Ｈ９を満たす所定値であり、Ｈ９は、Ｈ８＜Ｈ９≦Ｈ７−１−Δを満たす所定値である。例えば、テンポの検出範囲が６０から２４０である場合、Ｅ３はＴｇ３の時間間隔で生成されているので、Δ＝（２５０／Ｔｇ３）から（１０００／Ｔｇ３）の範囲でΔは変えられる。Ｔｇ３は、ｍｓｅｃ単位の値である。

H8 is a predetermined value that satisfies 0 ≦ H8 <H9, and H9 is a predetermined value that satisfies H8 <H9 ≦ H7-1−Δ. For example, when the detection range of tempo is 60 to 240, E3 is generated at a time interval of Tg3, and therefore Δ can be changed in the range of Δ = (250 / Tg3) to (1000 / Tg3). Tg3 is a value in units of msec.

  Next, the beat time detector 17 detects the peak position of the autocorrelation Y (Δ) and calculates the beat time length τ (S570). The autocorrelation Y (Δ) calculated in step S560 has several peaks as shown in FIG. The beat time detection unit 17 detects the maximum value position Δmax between the shortest beat of the detection target and the longest beat of the detection target, and sets τ = Tg3 × Δmax as the time length of one beat. In FIG. 14, “P” is Δ corresponding to the shortest beat to be detected, and “R” is Δ corresponding to the longest beat to be detected.

  Further, as shown in FIG. 15, a distribution Ω (Δ) indicating the existence probability of the beat time length is prepared, and the beat time detecting unit 17 calculates the autocorrelation Y (Δ) and the distribution Ω (Δ). After calculating the product (Ω (Δ) Y (Δ)), the position of the maximum value may be detected, thereby detecting the time length of one beat. The beat time detector 17 can calculate the beat length with higher accuracy by using Ω (Δ). In FIG. 15, “P” is Δ corresponding to the shortest beat to be detected, “U” is Δ that maximizes the existence probability of the beat, and “R” is the longest beat to be detected. The corresponding Δ.

  The beat time detection unit 17 notifies the control unit 11 of the beat time length τ thus detected.

  The control unit 11 calculates two numerical values of τ1 = λ1 × τ and τ2 = λ2 × τ. λ1 and λ2 are predetermined coefficients that satisfy λ1 <λ2.

  Then, the control unit 11 instructs the first feature value calculation unit 13 to set Tf1 = τ1, and also instructs the second feature value calculation unit 14 to set Tf2 = τ2. Thereafter, the control unit 11 instructs the first feature value calculation unit 13 and the second feature value calculation unit 14 to start the operation. The first feature value calculator 13 sets τ1 based on the time length of one beat detected by the beat time detector 17 to the frame time length Tf1, and the second feature value calculator 14 detects the beat time. Τ2 based on the time length of one beat detected by the unit 17 is set as the time length Tf2 of the frame. The subsequent operation of each unit is the same as the operation described in the first embodiment.

  Since the acoustic signal analysis apparatus 1 according to the second embodiment sets the section length when calculating the feature quantity related to the volume based on the time length of the beat of the music, it can be applied to music of various genres and types. The feature position can be detected with high accuracy.

  Note that the functions of the components of the acoustic signal analysis apparatus 1 according to each of the above-described embodiments are performed by, for example, hardware such as a computer CPU (processor) and a memory, and a computer program for realizing the functions. It is realized by doing. However, each of the above functions may be realized in any form such as realized by a dedicated circuit. Moreover, the computer program for implement | achieving the function of each structure part of the acoustic signal analyzer 1 may be stored in a recording medium.

(Embodiment 3)
In general, multiple musical instruments and singing are often performed simultaneously at characteristic parts of the music such as rust and excitement of the music, and the acoustic signal at the characteristic part has a wide frequency band. There are many. In other words, the acoustic signal at a characteristic location often includes a wide range of frequency components from a low range to a high range. The acoustic signal analyzing apparatus according to the third embodiment focuses on the characteristic of the frequency band of the acoustic signal at the above characteristic part, which has not been conventionally considered, and accurately detects the characteristic part of the music.

  First, the acoustic signal analysis apparatus 101 of Embodiment 3 is demonstrated using FIG. FIG. 16 is a configuration diagram of the acoustic signal analyzer 101 according to the third embodiment. As shown in FIG. 16, the acoustic signal analysis apparatus 101 according to the third embodiment includes a control unit 111, an acquisition unit 112, a frequency band data calculation unit 113, a smoothing unit 114, and a feature position detection unit 115. Have.

  The acoustic signal analyzer 101 acquires the acoustic signal 102 and outputs the characteristic position information 103.

  The acoustic signal 102 is an acoustic signal related to music. The acoustic signal 102 may be a digital signal or an analog signal. The sound signal 102 may be a signal including sound other than music such as DJ in addition to music, such as an audio signal of a music program such as radio or television, instead of a signal only of music. The acoustic signal 102 exists outside the acoustic signal analyzer 101. However, if the acoustic signal analyzer 101 is provided with a storage unit, the acoustic signal 102 may be stored in the storage unit and exist inside the acoustic signal analyzer 101.

  The feature position information 103 is information that identifies a location where the frequency band of the acoustic signal 102 is wide. The part often coincides with a part where the rust position of the music or the composition of the music or the organization of the musical instrument changes greatly, that is, a characteristic part of the music.

  The control unit 111 of the acoustic signal analysis device 101 exchanges information with other units constituting the acoustic signal analysis device 101 to control each unit.

  The acquisition unit 112 acquires the acoustic signal 102 and generates PCM (Pulse Code Modulation) data sampled at the sampling period Ts (sampling frequency Fs = 1 / Ts) from the acquired acoustic signal 102. When the acoustic signal 102 is an analog signal, the acquisition unit 112 converts the analog signal into a digital signal to generate PCM data. When the acoustic signal 102 is a digital compressed signal other than PCM, the acquisition unit 112 decodes the digital compressed signal. To generate PCM data. When the acoustic signal 102 is a digital signal and the sampling period is different from the sampling period Ts, the acquisition unit 112 converts the sampling rate and generates PCM data with the sampling period Ts.

  In the following description, PCM data generated by the acquisition unit 112 is described as acoustic data x [m] (m = 0 to M−1, where M is the total number of samples of acoustic data) or acoustic data. When the acquisition unit 112 finishes generating the acoustic data, the acquisition unit 112 notifies the control unit 111 accordingly. In Embodiment 3, the frequency band data calculation unit 113, the smoothing unit 114, and the feature position detection unit 115 start operating after the acquisition unit 112 generates all of the acoustic data. However, the frequency band data calculation unit 113, the smoothing unit 114, and the feature position detection unit 115 may start operating after the acquisition unit 112 generates part of the acoustic data.

  The frequency band data calculation unit 113 calculates time-series data related to the width of the frequency band from the acoustic data generated by the acquisition unit 112. The frequency band data calculation unit 113 performs processing in units of frames. However, the unit of processing is not limited thereto.

  In the following, it is assumed that the time length of each frame processed by the frequency band data calculation unit 113 is Tf11 and the time length of the frame shift is Tg11. At this time, the frame sample number N11 = Tf11 / Ts, and the frame shift sample number G11 = Tg11 / Ts. Frame shift is the time difference between the heads of adjacent frames. Adjacent frames may partially overlap or may not overlap.

  The frequency band data calculation unit 113 starts the operation shown in the flowchart of FIG. 17 in accordance with an instruction from the control unit 111. FIG. 17 is a flowchart showing each step of the operation of the frequency band data calculation unit 113.

  First, the frequency band data calculation unit 113 calculates the total number H11 of frames according to the following equation (17) (S600).

ｆｌｏｏｒ（ ）は、小数点以下を切り捨てた整数を返す関数である。ＭとＮ１１との関係は、Ｍ＞Ｎ１１である。

floor () is a function that returns an integer with the decimal part truncated. The relationship between M and N11 is M> N11.

  Next, the frequency band data calculation unit 113 sets “0” to the control variable i (S610).

  Next, the frequency band data calculation unit 113 generates i-th frame data (S620). The i-th frame data is data from the acoustic data x [i × G11] to the acoustic data x [i × G11 + N11-1]. The frequency band data calculation unit 113 generates a value obtained by multiplying the data from the acoustic data x [i × G11] to the acoustic data x [i × G11 + N11-1] by a window function as the i-th frame data. Also good. The window function is a Hamming window function, a Hanning window function, a Blackman window function, a Gauss window function, or the like. It can be said that the method described first is the same as the method of generating the i-th frame data by multiplying the acoustic data by a rectangular window. The i-th frame data is described as “D11 [i] [j] (j = 0 to ND11, where ND11 = N11-1)”.

  Next, the frequency band data calculating unit 113 calculates the frequency spectrum by analyzing the frequency of the i-th frame data using a known discrete Fourier transform (DFT) (S630). The frequency spectrum may be either an amplitude spectrum or a power spectrum. The intensity of the frequency spectrum may be expressed by a linear (linear) scale or a logarithmic scale.

  The frequency band data calculation unit 113 may use a method such as a wavelet transform or a filter bank instead of the discrete Fourier transform. The frequency spectrum of the i-th frame is described as “S [i] [k] (k = 0 to N11 / 2)”. k = 0 corresponds to the DC component with the lowest frequency, k = N11 / 2 corresponds to half the sampling frequency Fs, which is the highest frequency, and the component therebetween is the k × (Fs / N11) frequency. Corresponding to The axis indicating the frequency may be a logarithmic scale instead of a linear scale.

  Next, the frequency band data calculation unit 113 calculates an index (an index indicating the width of the frequency band) E11 [i] related to the frequency spectrum bandwidth of the i-th frame using a method described later (S640).

  Next, the frequency band data calculation unit 113 increases the value of the control variable i by “1” (S650).

  Next, the frequency band data calculation unit 113 determines whether or not the value of the control variable i is less than H11 (S660). If the value of the control variable i is less than H11 (Yes in S660), the frequency band data calculation unit 113 returns to step S620 and repeats the processing up to step S650, and if the value of the control variable i is H11 (S660). No), the process is terminated.

  In this way, the frequency band data calculation unit 113 calculates H11 time-series frequency band data E11 [i] (i = 0 to H11-1) related to the bandwidth of the frequency spectrum, and the processing is completed. This is notified to the control unit 111.

  Next, a method in which the frequency band data calculation unit 113 calculates the index E11 [i] related to the bandwidth of the frequency spectrum of the i-th frame in step S640 will be described.

  (1) The first calculation method of the index related to the bandwidth is a method of detecting a minimum frequency and a maximum frequency having a spectrum intensity equal to or higher than a predetermined value in a frequency spectrum, and calculating a difference between those frequencies. In general, the frequency spectrum is expressed as shown in FIG. A threshold λa for a low frequency and a threshold λb for a high frequency are prepared, and a frequency Ka that is the minimum k that satisfies S [i] [k] ≧ λa and a maximum that satisfies S [i] [k] ≧ λb The frequency Kb, which is k, is detected. Then, the difference between the frequency Ka and the frequency Kb, that is, (Kb−Ka) is defined as a bandwidth, and this is used as an index related to the bandwidth. The index related to the bandwidth may be (Fs / N11) × (Kb−Ka). In the first calculation method, the minimum frequency and the maximum frequency that satisfy the condition are obtained in a frequency range of about several tens Hz to several KHz or 10 KHz. The minimum frequency is detected in steps of several tens Hz (accuracy), and the maximum frequency is detected in steps of several hundred Hz (accuracy).

  (2) The second method for calculating the bandwidth-related index is a method using the shape of the frequency spectrum. Specifically, the second calculation method uses a sum of products (values of product-sum operation) of values related to the difference between each frequency value of the frequency spectrum and a predetermined value and the intensity of the frequency. It is. More specifically, the index E11 [i] related to the bandwidth of the frequency spectrum is calculated using the following formula (18) or formula (19).

式（１８）は、周波数スペクトルの各周波数の値と所定値との差の２乗値とその周波数の強度との積の総和を用いて帯域幅に関する指標Ｅ１１［ｉ］を算出するための式であり、式（１９）は、周波数スペクトルの各周波数の値と所定値との差の絶対値とその周波数の強度との積の総和を用いて帯域幅に関する指標Ｅ１１［ｉ］を算出するための式である。

Expression (18) is an expression for calculating the bandwidth index E11 [i] using the sum of the products of the square value of the difference between each frequency value of the frequency spectrum and the predetermined value and the intensity of the frequency. Equation (19) is for calculating the bandwidth index E11 [i] using the sum of products of the absolute value of the difference between each frequency value of the frequency spectrum and the predetermined value and the intensity of the frequency. It is an expression.

  In Expression (18) and Expression (19), K1 is an integer indicating the lower limit of the frequency to be processed, K2 is an integer indicating the upper limit of the frequency to be processed, and 0 ≦ K1 <K2 ≦ (N11 / 2) ) Relationship is satisfied. ω is a predetermined value, and the relationship of K1 ≦ ω ≦ K2 is satisfied. η is a value set in a range of 0 ≦ η ≦ 1. When η = 1, information on the intensity of the frequency spectrum does not enter E11 [i], so that a pure index regarding the bandwidth can be obtained. When η = 0, the denominator is 1, and E11 [i] is expressed only by the numerator, so that an index taking into account the intensity of the frequency spectrum is obtained. The value of η may be an intermediate value between 0 and 1, for example, 0.5.

  In Expression (18), μ is a predetermined value greater than zero. For example, μ = 1 or μ = 0.5. In Equation (18) and Equation (19), k is not a frequency itself but a number for identifying a frequency component. Instead of (k−ω), (k × Fs / N11−ω) of the frequency itself is used. ) May be used.

  (3) A third method for calculating the bandwidth-related index is a method using the sum of products of values related to the difference between the frequency of each component of the frequency spectrum and the average frequency and the component. Specifically, first, the average frequency ωa is calculated according to the following equation (20).

次に、式（１８）又は式（１９）のωに、ωａを代入してＥ１１［ｉ］を算出する。第３の算出方法を用いる場合、事前にωを決定しておく必要がないので、第２の算出方法を用いる場合よりも、多様なジャンルや音楽スタイルの楽曲に対応して帯域幅に関する指標Ｅ１１を算出することができる。

Next, E11 [i] is calculated by substituting ωa into ω in formula (18) or formula (19). When the third calculation method is used, it is not necessary to determine ω in advance. Therefore, the bandwidth index E11 corresponds to music of various genres and music styles, compared to the case where the second calculation method is used. Can be calculated.

  In the second and third calculation methods, a frequency spectrum in a frequency range of about several tens Hz to several KHz or 10 KHz is calculated. The resolution of the frequency spectrum is set to several tens Hz to several hundreds Hz. Alternatively, a spectrum may be calculated in which the frequency resolution is not equal (linear), but the frequency resolution in the low band is fine and the resolution becomes coarser (logarithmically on the frequency axis) as the frequency increases. Moreover, you may calculate the frequency spectrum corresponding to the temperament (average temperament etc.) used by the music. In the equal temperament, the frequencies of each scale, de, de #, re, re # are logarithmically arranged at equal intervals. In the second calculation method, for example, a value of about 1 K to 2 KHz is set as the predetermined value ω. Further, an average frequency in general music may be set as the predetermined value ω.

  (4) The fourth calculation method of the index related to the bandwidth is a method of using a numerical difference regarding the bandwidth of the frequency spectrum of two adjacent frames. The numerical value related to the bandwidth is a value obtained by any one of the first to third calculation methods.

  For example, when using the value obtained by the second calculation method, the result of substituting the acoustic data corresponding to the (i−1) -th frame into the equation (18) or the equation (19) is E11 ′ [i−1]. And holding the result of substituting the acoustic data corresponding to the i-th frame into the equation (18) or the equation (19) as E11 ′ [i]. Then, a difference E11 [i] = E11 ′ [i] −E11 ′ [i−1] between E11 ′ [i] and E11 ′ [i−1] is calculated, and this is calculated as an index related to the bandwidth of the frame i. To do. This index is not the bandwidth itself but the amount of change in bandwidth. Since the bandwidth often increases rapidly at the start position of the chorus of the music, the value of this index becomes large at such a location.

  In the first to fourth calculation methods of the bandwidth-related index, for example, the obtained data may be normalized so that the maximum value is 1 and the minimum value is 0.

  When detecting the end of the processing of the frequency band data calculation unit 113, the control unit 111 instructs the smoothing unit 114 to start the operation.

  Next, the smoothing unit 114 will be described. Since the frequency band data E11 [i] (i = 0 to H11-1) generated by the frequency band data calculation unit 113 often includes minute fluctuations (noise), the smoothing unit 114 Then, noise is removed by filtering with a low-pass filter. For example, the smoothing unit 114 uses the following equation (21) to multiply the frequency band data E11 of three adjacent frames by the coefficient (1, 2, 1) to obtain a smoothed output E [i] ( i = 0 to H11-1) is calculated. Of course, a low-pass filter with other coefficients may be used.

なお、両側の隣接フレームデータが揃わないＥ［０］及びＥ［Ｈ１１−１］については、揃っていないデータに対する係数を「０」に設定する。このように、周波数帯域データを平滑化することにより、特徴位置検出部１１５による特徴位置の検出精度が向上する。なお、平滑化部１１４は、省略されてもよい。

For E [0] and E [H11-1] where adjacent frame data on both sides are not aligned, the coefficient for the unaligned data is set to “0”. Thus, by smoothing the frequency band data, the feature position detection accuracy by the feature position detector 115 is improved. Note that the smoothing unit 114 may be omitted.

  When detecting the end of the process of the smoothing unit 114, the control unit 111 instructs the feature position detection unit 115 to start the operation.

  The feature position detection unit 115 detects a characteristic position such as a rust start position in the acoustic signal 102 using the value obtained by the smoothing unit 114. One of the following methods is used as a method for detecting the feature position. However, when the smoothing unit 114 is omitted, the feature position detection unit 115 processes the frequency band data E11 [i] calculated by the frequency band data calculation unit 113. Also, E [i] in the following description is replaced with E11 [i].

  (1) The first feature position detection method is a method for detecting a frame (position) at which the smoothed output is maximized. The index i (hereinafter referred to as “Imax”) at the position where the smoothed output E [i] (i = 0 to H11-1) is maximum is detected, and the time from the beginning of the music corresponding to Imax ( Tg11 × Imax) is defined as the feature position.

  Note that, instead of searching for the maximum value from all of the smoothed outputs, the range for searching for the maximum value may be limited. That is, the position where the value obtained by the smoothing unit 114 is maximum may be detected for a continuous portion of the acoustic signal 102. Specifically, the maximum value may be searched for E [i] (i = Ha to Hb, Ha and Hb are integers satisfying 0 ≦ Ha <Hb <H11-1). For example, Ha = 0 is set, and Hb is set to a value corresponding to about 70% of the music length. Further, the frequency band data E11 may be calculated from the acoustic signal 102 corresponding to a continuous part of the music, for example, about 70% of the music, and the position where the smoothed output E calculated based on this is maximized may be detected. . Using the method of searching for the maximum value for the smoothed output corresponding to a continuous portion of the acoustic signal 102 as described above can reduce the processing amount and improve the detection accuracy of the feature position for the following reason. be able to.

  The rust of a song is often repeated multiple times in one song, but the nuances of performance and singing are not always the same each time and are often slightly different. That is, the rust at the front position in time is often not completely raised as compared with the rust at the rear position. Considering the case of playing a part of a piece of music for trial listening, it is desirable that the part is a part that makes the listener think that “I want to listen to this whole piece of music”. Therefore, the rust in the position before the music that gives a sense of expectation to the future excitement is more suitable for trial listening than the rust in the position behind the music in the fully excited state. Limiting the range of detecting the maximum smoothed output to about 70% of the first half of the music makes it easier to detect rust at the position in front of the music and improves the detection accuracy of the characteristic position for trial listening.

  Also, an appropriate value may be set for Ha so that the intro part of the music is not included in the feature position detection target.

  Note that the position where the smoothed output is maximized is not set as the feature position, but the smoothed output is smoothed before the position where the smoothed output is maximized for a predetermined time or before the position where the smoothed output is maximized. A position where the normalized output is smaller than the maximum value by a predetermined value may be set as the feature position. As a result, it is possible to prevent detection omission of rust out.

  (2) The second feature position detection method is a method for detecting a position where the smoothed output is maximized according to the flowchart shown in FIG. FIG. 19 is a flowchart showing each step of the operation when the feature position detection unit 115 executes the second feature position detection method.

  The feature position detection unit 115 first sets an initial value “Hc” in the control variable i (S700). Hc is a predetermined integer that satisfies 1 ≦ Hc <H11-1. When searching for the maximum position from all the smoothed outputs, Hc = 1. When the intro of the music is not included in the detection target of the maximum position, Hc> 1.

  Next, the feature position detection unit 115 determines whether E [i] is a local maximum (S710). This determination method is, for example, a method of determining E [i] as a maximum value if E [i]> E [i-1] and E [i]> E [i + 1]. If the characteristic position detection unit 115 determines that E [i] is a maximum value (Yes in S710), the smoothed output E [i] at the maximum position and the value of the control variable (index, time information) at the maximum position. i is stored in the working memory inside the feature position detector 115 (S720).

  Next, the feature position detection unit 115 increases the value of the control variable i by “1” (S730). When it is determined in step S710 that E [i] is not the maximum value (No in S710), the feature position detection unit 115 performs the process of step S730.

  Next, the feature position detection unit 115 determines whether or not the control variable i is equal to or less than the predetermined value Hd (S740). Hd is a predetermined integer that satisfies Hc <Hd <H11-1. When searching for the maximum position for all of the smoothed outputs, Hd = H11-2. For the reasons described above, when the portion behind the music is excluded from the detection target of the maximum position, Hd <H11-2 is set, and the detection range of the maximum value of the smoothed output is, for example, 70% of the length of the music Limited to.

  If the characteristic position detection unit 115 determines that the control variable i is equal to or less than the predetermined value Hd (Yes in S740), the feature position detection unit 115 returns to step S710 and repeats the processing up to step S730.

  If the characteristic position detection unit 115 determines that the control variable i is greater than the predetermined value Hd (No in S740), the characteristic position detection unit 115 selects a predetermined number of local maximum positions from the local maximum information stored in the work memory (S750). . Specifically, the feature position detection unit 115 selects a predetermined number of local maximum positions in descending order of local maximum values. The P maximum positions (time) selected in descending order are described as Ip [v] (v = 0 to P−1). At this time, E [Ip [0]] ≧ E [Ip [1]] ≧ E [Ip [2]] ≧. . . ≧ E [Ip [P-1]]. For example, when the width of the frequency band changes as time passes as shown in FIG. 20, the feature position detection unit 115 sets the maximum position A ′ having the maximum value and the maximum position B ′ having the second value. The local maximum position C ′ having the third value is selected.

  Note that the feature position detection unit 115 may exclude a position close in time to the already selected maximum position when selecting a predetermined number of maximum positions in descending order of the maximum value. For example, the feature position detection unit 115 may select only the maximum value that is separated from the already selected maximum position by a predetermined time or more. In addition, the feature position detection unit 115 may detect a position where the value obtained by the smoothing unit 114 is maximum for a continuous portion of the acoustic signal 102. The above is the description of the second feature position detection method.

  The feature position detection unit 115 outputs the maximum position Imax or the maximum position Ip [v] (v = 0 to P−1) detected in this way to the outside of the acoustic signal analysis apparatus 101 as the feature position information 103. By reproducing the acoustic signal 102 using the characteristic position information 103, it is possible to reproduce a characteristic portion of the music such as rust.

  As described above, the acoustic signal analysis apparatus 101 according to Embodiment 3 calculates the width of the frequency band of each section constituting the acoustic signal 102 or data directly related thereto, and detects the section in which the maximum or maximum is obtained. To do. As a result, it is possible to accurately detect feature positions such as rust and excitement of music.

(Embodiment 4)
Next, the acoustic signal analysis apparatus 101 of Embodiment 4 is demonstrated using FIG. FIG. 21 is a configuration diagram of the acoustic signal analysis device 101 according to the fourth embodiment. As shown in FIG. 21, the acoustic signal analysis apparatus 101 according to the fourth embodiment includes a control unit 111, an acquisition unit 112, a frequency band data calculation unit 113, a smoothing unit 114a, a feature position detection unit 115, A second frequency band data calculation unit 116 and an evaluation value calculation unit 117 are included.

  The acoustic signal analysis device 101 according to the fourth embodiment includes a second frequency band data calculation unit 116 and an evaluation value calculation unit 117 in addition to the components included in the acoustic signal analysis device 101 according to the third embodiment. . The acoustic signal analysis apparatus 101 according to the fourth embodiment includes a smoothing unit 114a instead of the smoothing unit 114 included in the acoustic signal analysis apparatus 101 according to the third embodiment. This is the difference between the fourth embodiment and the third embodiment.

  The operations of the acquisition unit 112 and the frequency band data calculation unit 113 are the same as those described in the third embodiment.

  The operation of the second frequency band data calculation unit 116 is substantially the same as the operation of the frequency band data calculation unit 113. However, the second frequency band data calculation unit 116 processes a frame having a time length Tf12 different from the time length Tf11 of the frame processed by the frequency band data calculation unit 113. The reason will be described below.

  The frequency components of acoustic signals related to music vary in various time scales, such as individual notes that make up music, decorative sounds of notes such as vibrato, beats, measures, phrases, and general composition such as intro and rust Change due to various factors (multi-layered structure of music). In such a multi-layered structure of music, the ornamental sound of one note changes its frequency on a relatively short time scale, whereas the global structure such as intro and rust has a relatively long time scale. Change the frequency with.

  For example, at the starting point of chorus, a plurality of instruments and singing with different sound ranges are often played simultaneously, and particularly percussion instruments with a wide frequency band and a short decay time are often played. There is a strong tendency to spread the frequency band in a relatively short time corresponding to a half note. In addition, normal chorus has a length of several bars or more, and both low-frequency parts and high-frequency parts often continue to be played, so the frequency band can be extended in a relatively long time corresponding to several bars from the start point of chorus. There is a wide tendency. Since rust has such characteristics, rust detection accuracy can be improved by calculating a plurality of frequency band data having different time scales.

  The time length Tf12 of the frame processed by the second frequency band data calculation unit 116 is longer than the time length Tf11 of the frame processed by the frequency band data calculation unit 113. Specifically, the frequency band data calculation unit 113 processes a frame of Tf11 corresponding to one musical note or a time length of one beat or less of the music, and the second frequency band data calculation unit 116 is longer than one beat. A frame of Tf12 corresponding to a time length of about 1 bar to 8 bars is processed. For example, Tf11 is set to 125 msec corresponding to a sixteenth note of a music piece having a 4/4 time signature and a tempo of 120, and Tf12 is set to 2 sec corresponding to one measure.

  The frame shift time length Tg12 when the second frequency band data calculation unit 116 operates and the frame shift time length Tg11 when the frequency band data calculation unit 113 operates may be the same, May be different. In the fourth embodiment, Tg12 = Q1 × Tg11, and the number of frame shift samples G12 = Q1 × G11 when the second frequency band data calculation unit 116 operates (Q1 is an integer equal to or greater than 1). However, Tg12 and G12 are not limited to this.

  In addition, the total number of frames processed by the second frequency band data calculation unit 116 is H12.

  Under such conditions, the second frequency band data calculation unit 116 performs the same operation as the frequency band data calculation unit 113 of the third embodiment, and the second frequency band data E12 [j] (j = 0 to H12-1) is calculated.

  Next, the evaluation value calculation unit 117 will be described. The evaluation value calculation unit 117 includes frequency band data E11 [i] calculated by the frequency band data calculation unit 113, and second frequency band data E12 [j] calculated by the second frequency band data calculation unit 116. Is used to calculate an evaluation value. The evaluation value calculation unit 117 uses the frequency band data E11 [i] and the second frequency band data E12 [j], E11 [i] is large, and E12 [i] is temporally corresponding to E11 [i]. The evaluation value is calculated such that the larger j] is, the larger the value is.

  Upon detecting the end of the processing of the frequency band data calculation unit 113 and the second frequency band data calculation unit 116, the control unit 111 instructs the evaluation value calculation unit 117 to start operation, and the evaluation value calculation unit 117 starts the operation shown in the flowchart of FIG. FIG. 22 is a flowchart showing the steps of the operation of the evaluation value calculation unit 117.

  First, the evaluation value calculation unit 117 sets “0” to the control variable i (S800).

  Next, the evaluation value calculation unit 117 calculates a value to be set in the control variable j according to the following equation (22) (S810).

ｆｌｏｏｒ（ ）は、小数点以下を切り捨てた整数を返す関数である。Ｑ１は、周波数帯域データ算出部１１３が動作する際のフレームシフト時間長を基準とした、第２の周波数帯域データ算出部１１６が動作する際のフレームシフト時間長の倍率であり、１以上の整数である。

floor () is a function that returns an integer with the decimal part truncated. Q1 is a magnification of the frame shift time length when the second frequency band data calculation unit 116 is operated, based on the frame shift time length when the frequency band data calculation unit 113 is operated, and is an integer of 1 or more It is.

  Next, the evaluation value calculation unit 117 calculates an evaluation value α [i] corresponding to the control variable i according to a method described later (S820).

  Next, the evaluation value calculation unit 117 increases the value of the control variable i by “1” (S830).

  Next, the evaluation value calculation unit 117 determines whether the control variable i is less than the product value (Q1 × H12) of H12 (the total number of frames processed by the second frequency band data calculation unit 116) and Q1. It is determined whether or not (S840). If the evaluation value calculation unit 117 determines that the control variable i is less than (Q1 × H12) (Yes in S840), the process returns to Step S810 and repeats the processing up to Step S830, and the control variable i is (Q1 × H12). If it is determined that (No in S840), the process is terminated.

  The evaluation value calculation unit 117 calculates the evaluation value α [i] (i = 0 to Q1 × H12-1), which is (Q1 × H12) pieces of time series data, by the above-described processing. The evaluation value calculation unit 117 notifies the control unit 111 that the processing has been completed.

  The evaluation value calculation unit 117 calculates the evaluation value α [i] by any of the following methods.

  (1) The first evaluation value calculation method includes frequency band data E11 [i] and a second frequency corresponding in time to frequency band data E11 [i] as shown in the following equation (23). This is a method of adding the band data E12 [j].

なお、周波数帯域データＥ１１［ｉ］と、Ｅ１１［ｉ］に時間的に対応する第２の周波数帯域データＥ１２［ｊ］とを加算した値に所定値を乗算した値を評価値としてもよい。

Note that a value obtained by multiplying a value obtained by adding the frequency band data E11 [i] and the second frequency band data E12 [j] temporally corresponding to E11 [i] by a predetermined value may be used as the evaluation value.

  (2) As shown in the following equation (24), the second method of calculating the evaluation value is a value obtained by multiplying E11 [i] by a coefficient β3, and E12 [j] temporally corresponding to E11 [i]. ] Is added to a value obtained by multiplying the coefficient β4 by a coefficient β4. However, β3> 0 and β4> 0. In the second calculation method, E11 [i] and E12 [j] are respectively weighted and added.

（３）評価値の第３の算出方法は、下記の式（２５）に示すように、Ｅ１１［ｉ］の対数値に係数β３を乗じた値と、Ｅ１１［ｉ］に時間的に対応するＥ１２［ｊ］の対数値に係数β４を乗じた値との加算値を用いる方法である。なお、第１から第３の算出方法は、Ｅ１１とＥ１２のどちらかが小さい箇所で、評価値をあまり小さくしたくない場合に用いる。第３の算出方法は、それに加えて、Ｅ１１とＥ１２のそれぞれの値の範囲が大きく異なる場合に適している。

(3) The third calculation method of the evaluation value corresponds to the value obtained by multiplying the logarithmic value of E11 [i] by the coefficient β3 and E11 [i] in time, as shown in the following equation (25). In this method, an addition value of a value obtained by multiplying the logarithmic value of E12 [j] by a coefficient β4 is used. The first to third calculation methods are used when one of E11 and E12 is small and it is not desired to make the evaluation value too small. In addition to this, the third calculation method is suitable when the ranges of the values of E11 and E12 are greatly different.

（４）評価値の第４の算出方法は、下記の式（２６）に示すように、Ｅ１１［ｉ］とＥ１１［ｉ］に時間的に対応するＥ１２［ｊ］との積を用いる方法である。なお、式（２６）の右辺にさらに所定値を乗算した値を評価値としてもよい。

(4) The fourth calculation method of the evaluation value is a method using a product of E11 [i] and E12 [j] temporally corresponding to E11 [i] as shown in the following equation (26). is there. Note that a value obtained by further multiplying the right side of Expression (26) by a predetermined value may be used as the evaluation value.

（５）評価値の第５の算出方法は、下記の式（２７）に示すように、Ｅ１１［ｉ］を基数としγ３を指数とした累乗値と、Ｅ１１［ｉ］に時間的に対応するＥ１２［ｊ］を基数としγ４を指数とした累乗値との積を用いる方法である。第４及び第５の算出方法は、Ｅ１１とＥ１２のどちらかが小さければ、評価値も小さくしたい場合に用いる。第５の算出方法は、それに加えて、Ｅ１１とＥ１２の評価値への影響力に重みを付けたい場合に適している。なお、式（２７）の右辺にさらに所定値を乗算した値を評価値としてもよい。

(5) The fifth method of calculating the evaluation value temporally corresponds to the power value with E11 [i] as the base and γ3 as the exponent, and E11 [i] as shown in the following equation (27). This is a method of using a product of a power value with E12 [j] as a radix and γ4 as an exponent. The fourth and fifth calculation methods are used when it is desired to reduce the evaluation value if either E11 or E12 is small. In addition, the fifth calculation method is suitable when it is desired to weight the influence on the evaluation values of E11 and E12. Note that a value obtained by further multiplying the right side of Expression (27) by a predetermined value may be used as the evaluation value.

（６）評価値の第６の算出方法は、下記の式（２８）に示すように、Ｅ１１［ｉ］を基数としγ３を指数とした累乗値と係数β３の積と、Ｅ１１［ｉ］に時間的に対応するＥ１２［ｊ］を基数としγ４を指数とした累乗値と係数β４の積との和を用いる方法である。なお、式（２８）の右辺にさらに所定値を乗算した値を評価値としてもよい。

(6) A sixth method for calculating the evaluation value is as follows: E11 [i] is a radix and γ3 is an exponential product with a coefficient β3 and E11 [i] This is a method of using the sum of the product of the power value and the coefficient β4 with the temporally corresponding E12 [j] as the radix and γ4 as the exponent. Note that a value obtained by further multiplying the right side of Expression (28) by a predetermined value may be used as the evaluation value.

なお、評価値算出部１１７は、Ｅ１１［ｉ］≧θ１かつＥ１２［ｊ］≧θ２（θ１、θ２は所定値）の条件が成立する場合、上述した第１から第６の算出方法により評価値を算出し、その条件が成立しない場合、評価値を「０」に設定してもよい。また、評価値算出部１１７は、評価値α［ｉ］を算出した後に、α［ｉ］＜θ３（θ３は所定値）である場合、α［ｉ］を「０」に設定してもよい。

Note that the evaluation value calculation unit 117 evaluates the evaluation value using the first to sixth calculation methods described above when the conditions of E11 [i] ≧ θ1 and E12 [j] ≧ θ2 (θ1 and θ2 are predetermined values) are satisfied. If the condition is not satisfied, the evaluation value may be set to “0”. Moreover, after calculating the evaluation value α [i], the evaluation value calculation unit 117 may set α [i] to “0” when α [i] <θ3 (θ3 is a predetermined value). .

  The evaluation value α [i] calculated by the above-described method becomes larger as E11 [i] is larger and E12 [j] corresponding to E11 [i] is larger in time. In the multi-layered structure of music having various time scales, E11 [i] represents a relatively short time change such as one note or one beat, and E12 [j] represents a longer time change.

  The “feature position” of a song, such as the start position of the chorus, the point where the tune changes greatly, the position suitable for auditioning, and the position that gives a strong impression to the listener, often has a very wide frequency band at the beginning. From 1 to 8 bars, the average frequency band is often wide, and the evaluation value of such a portion is a large value. Therefore, by detecting the maximum value or the maximum value of the evaluation values, it is possible to accurately detect a feature position such as a rust start position.

  When detecting the end of the processing of the evaluation value calculation unit 117, the control unit 111 instructs the smoothing unit 114a to start the operation. The smoothing unit 114a performs the same operation as the smoothing unit 114 of the third embodiment. However, the smoothing unit 114a processes the evaluation value α [i] (i = 0 to Q1 × H12-1) instead of the frequency band data E11 [i] (i = 0 to H11-1). The smoothed output E [i] (i = 0 to Q1 × H12-1) is calculated. Note that the smoothing unit 114a may be omitted. Similarly to the third embodiment, the frequency band data may be smoothed by providing a smoothing unit 114 after the frequency band data calculation unit 113. Further, the second frequency band data may be smoothed.

  When detecting the end of the process of the smoothing unit 114a, the control unit 111 instructs the feature position detection unit 115 to start the operation. The feature position detection unit 115 performs processing similar to the processing described in the third embodiment, and outputs the feature position information 103 to the outside of the acoustic signal analyzer 101.

  As described above, the acoustic signal analysis apparatus 101 according to Embodiment 4 uses two time periods having different time lengths in order to accurately detect changes in frequency bands at different time scales due to the multi-layered structure of music. Two types of frequency band data are calculated using the sections, and an evaluation value is calculated by combining them. As a result, even when there is a change in frequency bands with different temporal scales, it is possible to accurately detect a characteristic position such as a rust position.

(Embodiment 5)
Next, the acoustic signal analysis apparatus 101 of Embodiment 5 is demonstrated using FIG. FIG. 23 is a configuration diagram of the acoustic signal analyzer 101 according to the fifth embodiment. As shown in FIG. 23, the acoustic signal analysis apparatus 101 according to the fifth embodiment includes a control unit 111, an acquisition unit 112, a frequency band data calculation unit 113, a smoothing unit 114a, a feature position detection unit 115, An evaluation value calculation unit 117a and a volume data calculation unit 118 are included.

  The acoustic signal analysis device 101 according to the fifth embodiment includes a volume data calculation unit 118 instead of the second frequency band data calculation unit 116 included in the acoustic signal analysis device 101 according to the fourth embodiment. The acoustic signal analysis apparatus 101 according to the fifth embodiment includes an evaluation value calculation unit 117a instead of the evaluation value calculation unit 117 included in the acoustic signal analysis apparatus 101 according to the fourth embodiment. This is the difference between the fifth embodiment and the fourth embodiment.

  The operations of the acquisition unit 112 and the frequency band data calculation unit 113 are the same as those described in the third embodiment.

  The volume data calculation unit 118 calculates data related to volume for each predetermined time interval. The time length Tf13 of the frame processed by the volume data calculation unit 118 and the time length Tf11 of the frame processed by the frequency band data calculation unit 113 may be the same or different. In the fifth embodiment, Tf13> Tf11 is set, but the present invention is not limited to this. In this case, the number N13 of frames processed by the sound volume data calculation unit 118 is N13 = Tf13 / Ts, and thus is larger than the number N11 of frames processed by the frequency band data calculation unit 113.

  As described in the fourth embodiment, the frequency component of an acoustic signal related to music varies depending on various factors (multi-layered structure of music) having different time scales, but the same can be said for sound volume.

  For example, at the beginning of the chorus of a song, in addition to playing multiple instruments and singing at the same time, each instrument is often played “strongly (forte)”. There is a strong tendency that the frequency band is widened and the volume is increased in a relatively short time interval corresponding to a half note. In addition, normal chorus has a length of several bars or more, and both the low-frequency part and the high-frequency part continue to be played, so the frequency band in a relatively long time corresponding to several bars from the start point of the chorus Tend to be wide and loud. Since rust has such characteristics, the accuracy of detecting a characteristic position such as rust can be improved by combining frequency band data having different time scales and volume data.

  Tf11 is set to a length of about 16th note to half note or less, and Tf13 is set to a time length of about 1 bar to about 8 bars. For example, Tf11 is set to 125 msec corresponding to a sixteenth note of a music piece having a 4/4 time signature and a tempo of 120, and Tf13 is set to 8 sec corresponding to four measures.

  The frame shift time length Tg13 when the volume data calculation unit 118 operates and the frame shift time length Tg11 when the frequency band data calculation unit 113 operates may be the same or different. Good. In the fifth embodiment, Tg13 is Tg13 = R1 × Tg11, and the frame shift sample number G13 when the volume data calculation unit 118 operates is G13 = R1 × G11 (R1 is an integer of 1 or more). . However, Tg13 and G13 are not limited to this.

  The sound volume data calculation unit 118 starts the operation shown in the flowchart of FIG. 24 in accordance with an instruction from the control unit 111. FIG. 24 is a flowchart showing each step of the operation of the sound volume data calculation unit 118.

  First, the volume data calculation unit 118 calculates the total number H13 of frames using the equation (17) (S900). That is, the sound volume data calculation unit 118 calculates the total number of frames H13 by replacing N11 in Equation (17) with N13, replacing G11 with G13, and replacing H11 with H13. In the fifth embodiment, M> N13. The total number of frames H13 processed by the volume data calculation unit 118 is equal to or less than the total number of frames H11 processed by the frequency band data calculation unit 113.

  Next, the volume data calculation unit 118 sets “0” to the control variable i (S910).

  Next, the volume data calculation unit 118 generates i-th frame data (S920). Specifically, the volume data calculation unit 118 generates the acoustic data x [i × G13 + N13-1] as the i-th frame data from the acoustic data x [i × G13]. Note that the volume data calculation unit 118 may generate a value obtained by multiplying the data from the acoustic data x [i × G13] to the acoustic data x [i × G13 + N13-1] by the window function as the i-th frame data. Good. The window function is, for example, a Hamming window function, a Hanning window function, a Blackman window function, or a Gauss window function. It can be said that the method described first is the same as the method of generating the i-th frame data by multiplying the acoustic data by a rectangular window.

  By the way, when the window function is used, the window function coefficient is usually maximized at the center of the frame and the window function coefficient is minimized at the beginning and end of the frame, but other methods may be used. For example, the volume data calculation unit 118 maximizes the window function coefficient at the beginning of the frame (x [i × G13]), then sequentially decreases the coefficient of the window function, and ends the frame (x [i × G13 + N13-1). ]), The window function coefficient may be minimized. The i-th frame data is described as “D13 [i] [j] (j = 0 to ND13, where ND13 = N13-1)”.

  The head D11 [i × R1] [0] of the i × R1 frame data processed by the frequency band data calculation unit 113 and the head D13 [i] [0] of the i th frame data processed by the volume data calculation unit 118 ] Are matched as x [i × G13], but it is not always necessary to match the heads of the frames in this way. For example, the center of the frame may be matched, or the end of the frame may be matched.

  Next, the volume data calculation unit 118 calculates volume data using the i-th frame data according to a method described later (S930).

  Next, the volume data calculation unit 118 increases the value of the control variable i by “1” (S940).

  Next, the volume data calculation unit 118 determines whether or not the value of the control variable i is less than H13 (S950). When determining that the value of the control variable i is less than H13 (Yes in S950), the sound volume data calculation unit 118 returns to step S920 and repeats the processing up to step S940, and determines that the value of the control variable i is H13. Then (No in S950), the process ends.

  The volume data calculation unit 118 calculates H13 volume data E13 [i] (i = 0 to H13-1) by the above-described process, and notifies the control unit 111 that the process is completed.

  Next, details of the processing performed by the volume data calculation unit 118 in step S930 will be described.

  (1) The first calculation method of the volume data is a method using the absolute value of the amplitude of the acoustic data. Specifically, as shown in the following equation (29), a value (sum) obtained by adding the absolute value of the amplitude by the number of samples of the frame is set as volume data corresponding to the i-th frame.

なお、下記の式（３０）に示すように、総和の代わりに平均値を用いてｉ番目のフレームに対応する音量データを算出してもよい。

Note that, as shown in the following equation (30), volume data corresponding to the i-th frame may be calculated using an average value instead of the sum.

（２）音量データの第２の算出方法は、音響データの振幅の２乗を用いる方法である。具体的には、下記の式（３１）に示すように、振幅の２乗の値をフレームのサンプル数だけ加算した値（総和）をｉ番目のフレームに対応する音量データとする。

(2) The second calculation method of volume data is a method using the square of the amplitude of acoustic data. Specifically, as shown in the following equation (31), a value (sum) obtained by adding the square value of the amplitude by the number of samples of the frame is set as volume data corresponding to the i-th frame.

なお、下記の式（３２）に示すように、総和の代わりに平均値を用いてｉ番目のフレームに対応する音量データを算出してもよい。また、式（３１）又は式（３２）の右辺の平方根をとった値をｉ番目のフレームに対応する音量データＥ１３［ｉ］としてもよい。

As shown in the following equation (32), the volume data corresponding to the i-th frame may be calculated using an average value instead of the sum. Further, a value obtained by taking the square root of the right side of Expression (31) or Expression (32) may be used as the volume data E13 [i] corresponding to the i-th frame.

（３）音量データの第３の算出方法は、所定の範囲の周波数成分を用いる方法である。ｉ番目のフレームデータＤ１３［ｉ］［ｊ］に対して、離散フーリエ変換（ＤＦＴ)を行い、周波数スペクトルＳ１３［ｉ］［ｋ］（ｋ＝０〜Ｎ１３／２）を算出する。周波数スペクトルは、振幅スペクトルとパワースペクトルのいずれでもよい。そして、所定の範囲の各周波数の強度の総和をＥ１３［ｉ］とする。

(3) The third calculation method of volume data is a method using frequency components in a predetermined range. A discrete Fourier transform (DFT) is performed on the i-th frame data D13 [i] [j], and a frequency spectrum S13 [i] [k] (k = 0 to N13 / 2) is calculated. The frequency spectrum may be either an amplitude spectrum or a power spectrum. Then, the sum of the intensities of the respective frequencies within a predetermined range is set to E13 [i].

  (4) A fourth calculation method of volume data is a method using a difference in numerical values indicating the volume of two adjacent frames. The numerical value indicating the volume of the frame is a value obtained by any one of the first to third calculation methods described above. For example, when the value obtained by the first calculation method is used, the calculation result when the acoustic data corresponding to the (i−1) th frame is substituted into the equation (29) is held as E13 ′ [i−1]. At the same time, the calculation result when the acoustic data corresponding to the i-th frame is substituted into Expression (29) is held as E13 ′ [i]. Then, the difference E13 [i] = E13 ′ [i] −E13 ′ [i−1] between E13 ′ [i] and E13 ′ [i−1] is calculated as volume data. This method is a method for calculating the amount of change in volume.

  In the first to fourth calculation methods described above, for example, the obtained data may be normalized so that the maximum value of the volume data is 1 and the minimum value is 0.

  When detecting the end of the processing of the frequency band data calculation unit 113 and the volume data calculation unit 118, the control unit 111 instructs the evaluation value calculation unit 117a to start the operation. The evaluation value calculation unit 117a performs the same operation as the evaluation value calculation unit 117 of the fourth embodiment. However, in the fourth embodiment, the evaluation value calculation unit 117 calculates the evaluation value using the frequency band data E11 and the second frequency band data E12. However, in the fifth embodiment, the evaluation value calculation unit 117a includes The evaluation value α is calculated using the frequency band data E11 and the volume data E13.

  When detecting the end of the process of the evaluation value calculation unit 117a, the control unit 111 instructs the smoothing unit 114a to start the operation. The smoothing unit 114a performs the same operation as in the fourth embodiment.

  When detecting the end of the process of the smoothing unit 114a, the control unit 111 instructs the feature position detection unit 115 to start the operation. The feature position detection unit 115 performs the same operation as that described in the third embodiment, and outputs the feature position information 103 to the outside of the acoustic signal analyzer 101.

  As described above, the acoustic signal analysis apparatus 101 according to the fifth embodiment has a time length in order to accurately detect a change in frequency band and a change in volume at different time scales due to the multi-layered structure of music. Frequency band data and volume data are calculated using two different time intervals, and an evaluation value is calculated by combining them. For this reason, the feature position can be detected with higher accuracy.

(Embodiment 6)
Next, the acoustic signal analysis apparatus 101 according to the sixth embodiment will be described with reference to FIG. FIG. 25 is a configuration diagram of the acoustic signal analyzer 101 according to the sixth embodiment. As shown in FIG. 25, the acoustic signal analysis apparatus 101 according to the sixth embodiment includes a control unit 111, an acquisition unit 112, a frequency band data calculation unit 113, a smoothing unit 114a, a feature position detection unit 115, A second frequency band data calculation unit 116, an evaluation value calculation unit 117, and a beat time detection unit 119 are included.

  Embodiment 6 The acoustic signal analysis device 101 includes a beat time detection unit 119 in addition to the components included in the acoustic signal analysis device 101 of the fourth embodiment. This is the difference between the sixth embodiment and the fourth embodiment.

  When the control unit 111 detects that the acquisition unit 112 has generated the acoustic data, the control unit 111 determines the beat time before instructing the frequency band data calculation unit 113 and the second frequency band data calculation unit 116 to start the operation. The detection unit 119 is instructed to start the operation.

  The beat time detector 119 performs processing in units of frames. The time length of the frame processed by the beat time detecting unit 119 is Tf14, and the time length of the frame shift when the beat time detecting unit 119 is operated is Tg14. The frame sample number N14 processed by the beat time detection unit 119 is N14 = Tf14 / Ts, and the frame shift sample number G14 is G14 = Tg14 / Ts. In order to accurately calculate the beat time, Tf14 and Tg14 are set to a time considerably shorter than the length of one beat. In general music, since the tempo is 60 to 240 and the time length of one beat is often in the range of 250 msec to 1 sec, Tf14 and Tg14 are set to appropriate values of about 5 msec to 50 msec.

  The beat time detection unit 119 performs processing according to the flowchart shown in FIG. FIG. 26 is a flowchart showing each step of the operation of the beat time detection unit 119.

  First, the beat time detection unit 119 calculates the total number H14 of frames using Expression (17) (S1000). Specifically, the beat time detection unit 119 calculates the total number H14 of frames by replacing N11 in Equation (17) with N14, replacing G11 with G14, and replacing H11 with H14.

  Next, the beat time detection unit 119 sets “0” to the control variable i (S1010).

  Next, the beat time detection unit 119 generates i-th frame data (S1020). Specifically, the beat time detection unit 119 generates acoustic data x [i × G14 + N14-1] as the i-th frame data from the acoustic data x [i × G14]. The beat time detection unit 119 generates a value obtained by multiplying the data from the acoustic data x [i × G14] to the acoustic data x [i × G14 + N1-1] by the window function as the i-th frame data. Good. The window function is, for example, a Hamming window function, a Hanning window function, a Blackman window function, or a Gauss window function. It can be said that the method described first is the same as the method of generating the i-th frame data by multiplying the acoustic data by a rectangular window. The i-th frame data is described as “D14 [i] [j] (j = 0 to ND14, where ND14 = N14-1)”.

  Next, the beat time detection unit 119 calculates the amount of change in volume corresponding to the i-th frame (S1030). Specifically, the beat time detection unit 119 calculates the volume change amount E14 [i] by using the fourth volume data calculation method used by the volume data calculation unit 118 of the fifth embodiment.

  Next, the beat time detection unit 119 increases the value of the control variable i by “1” (S1040).

  Next, the beat time detection unit 119 determines whether or not the value of the control variable i is less than H14 (S1050). When determining that the value of the control variable i is less than H14 (Yes in S1050), the beat time detecting unit 119 returns to Step S1020 and repeats the process up to Step S1040.

  When determining that the value of the control variable i is H14 (No in S1050), the beat time detecting unit 119 calculates the autocorrelation of the volume change amount E14 [i] (i = 0 to H14-1) (S1060). ). The beat time detector 119 calculates the autocorrelation Y (Δ) according to the following equation (33) while sequentially changing the autocorrelation index difference Δ within a predetermined tempo range.

式（３３）において、Ｈｅ及びＨｆは、０≦Ｈｅ＜Ｈｆ≦Ｈ１４−１−Δ、を満たす所定の整数である。例えば、テンポの検出範囲が６０から２４０（１拍の時間２５０ｍｓｅｃから１０００ｍｓｅｃ）である場合、Ｅ１４はＴｇ１４の時間間隔で生成されているので、Δ＝（２５０／Ｔｇ１４）から（１０００／Ｔｇ１４）の範囲でΔは変えられる。Ｔｇ１４は、ｍｓｅｃ単位の値である。

In the formula (33), He and Hf are predetermined integers that satisfy 0 ≦ He <Hf ≦ H14-1-Δ. For example, if the tempo detection range is 60 to 240 (one beat time is 250 msec to 1000 msec), E14 is generated at a time interval of Tg14, so Δ = (250 / Tg14) to (1000 / Tg14) Δ can be changed in the range. Tg14 is a value in units of msec.

  Next, the beat time detector 119 detects the peak position of the autocorrelation Y (Δ) and calculates the beat time length τ (S1070). The autocorrelation Y (Δ) calculated in step S1060 has several peaks as shown in FIG. The beat time detection unit 119 detects the position Δmax of the maximum value between the shortest beat of the detection target and the longest beat of the detection target, and sets τ = Tg14 × Δmax as the time length of one beat. In FIG. 14, “P” is Δ corresponding to the shortest beat to be detected, and “R” is Δ corresponding to the longest beat to be detected.

  Further, as shown in FIG. 15, a distribution Ω (Δ) indicating the existence probability of the beat time length is prepared, and the beat time detection unit 119 calculates the autocorrelation Y (Δ) and the distribution Ω (Δ). After calculating the product (Ω (Δ) Y (Δ)), the position of the maximum value may be detected, thereby detecting the time length of one beat. The beat time detector 119 can calculate the beat length with higher accuracy by using Ω (Δ). In FIG. 15, “P” is Δ corresponding to the shortest beat to be detected, “U” is Δ that maximizes the existence probability of the beat, and “R” is the longest beat to be detected. The corresponding Δ.

  The beat time detection unit 119 notifies the control unit 111 of the beat duration τ detected in this way.

  The control unit 111 calculates two numerical values of τ11 = λ11 × τ and τ12 = λ12 × τ. λ11 and λ12 are predetermined coefficients that satisfy λ11 <λ12. For example, λ11 is a value from “0.25” to “1”, and λ12 is a value from about “4” to “8”.

  Then, the control unit 111 instructs the frequency band data calculation unit 113 to set Tf11 = τ11 and also instructs the second frequency band data calculation unit 116 to set Tf12 = τ12. Thereafter, the control unit 111 instructs the frequency band data calculation unit 113 and the second frequency band data calculation unit 116 to start operation. The frequency band data calculating unit 113 sets τ11 based on the time length of one beat detected by the beat time detecting unit 119 to the frame time length Tf11, and the second frequency band data calculating unit 116 is a beat time detecting unit. Τ12 based on the time length of one beat detected by 119 is set as the time length Tf12 of the frame. The subsequent operation of each part is the same as the operation described in the fourth embodiment.

  Since the time length of one beat of the music differs depending on the music genre and style, the optimum section length for calculating the frequency band data and volume data also differs depending on the music genre and style. It is difficult to determine the optimal section length in advance. The acoustic signal analysis apparatus 101 according to the sixth embodiment detects a time length of one beat, and sets a section length when calculating frequency band data and second frequency band data based on the time length. Thereby, it is possible to accurately detect the feature position for various genres and types of music.

  Note that the volume data calculation unit 118 described in the fifth embodiment may also set Tf13 based on the time length of the beat.

  Moreover, you may combine each method of Embodiment 1- Embodiment 6 mentioned above. For example, the frequency band data, the second frequency band data, and the volume data are calculated by combining the method of the fourth embodiment and the method of the fifth embodiment, and the evaluation value is calculated using these three. May be. Further, by combining the method of the first embodiment and the method of the fourth embodiment, a first feature amount related to volume, a second feature amount related to volume, frequency band data, and second frequency band data And the evaluation value may be calculated using these four. Further, three or more types of frequency band data may be calculated. In this way, by calculating an evaluation value by combining different types of feature quantities (including volume-related feature quantities and data related to frequency bands), it is possible to further determine the characteristic parts of the song for a wide variety of songs. It can be detected with high accuracy.

  Furthermore, the functions of the components of the acoustic signal analysis apparatus 101 according to each of the above-described embodiments are performed by, for example, hardware such as a computer CPU (processor) and a memory, and a computer program for realizing the functions. It is realized by doing. However, each of the above functions may be realized in any form such as realized by a dedicated circuit. Moreover, the computer program for implement | achieving the function of each structure part of the acoustic signal analyzer 101 may be stored in a recording medium.

  DESCRIPTION OF SYMBOLS 1 Acoustic signal analyzer, 2 Acoustic signal, 3 Feature position information, 11 Control part, 12 Acquisition part, 13 1st feature-value calculation part, 14 2nd feature-value calculation part, 15 Evaluation value calculation part, 16 Feature position Detection unit, 17 beat time detection unit, 101 acoustic signal analyzer, 102 acoustic signal, 103 feature position information, 111 control unit, 112 acquisition unit, 113 frequency band data calculation unit, 114 smoothing unit, 114a smoothing unit, 115 A feature position detection unit, 116 a second frequency band data calculation unit, 117 evaluation value calculation unit, 117a evaluation value calculation unit, 118 volume data calculation unit, 119 beat time detection unit.

Claims (8)

An acquisition unit for acquiring an acoustic signal;
A first volume information calculation unit that calculates a first value related to the volume of each of a plurality of sections having a first period of the acoustic signal acquired by the acquisition unit;
A second volume information calculation unit that calculates a second value related to the volume of each of a plurality of sections having a second period longer than the first period of the acoustic signal acquired by the acquisition unit;
Using the first value and the second value, the larger the first value and the larger the second value corresponding to the first value in time, the larger the time. An evaluation value calculation unit for calculating the evaluation value of the series;
An acoustic signal analyzer comprising: a feature position detection unit that detects a position where the evaluation value calculated by the evaluation value calculation unit is maximum or maximum. Furthermore, a beat time detection unit that detects a time length of one beat of the acoustic signal acquired by the acquisition unit,
The first volume information calculation unit sets the first period based on a time length of one beat detected by the beat time detection unit,
The acoustic signal analysis device according to claim 1, wherein the second volume information calculation unit sets the second period based on a time length of one beat detected by the beat time detection unit. The evaluation value calculation unit
An added value of the first value and the second value;
An added value of a value obtained by multiplying the first value by a first coefficient and a value obtained by multiplying the second value by a second coefficient;
An addition value of a value obtained by multiplying a logarithmic value of the first value by a third coefficient and a value obtained by multiplying the logarithmic value of the second value by a fourth coefficient; A multiplication value of the value of 1 and the second value;
A multiplication value of a first power value having the first value as a radix and a fifth coefficient as an index and a second power value having the second value as a radix and a sixth coefficient as an index;
Using either one of a value obtained by multiplying the first power value by a seventh coefficient and a value obtained by multiplying the second power value by an eighth coefficient The acoustic signal analyzer according to claim 1 or 2, wherein the evaluation value is calculated. The first volume information calculation unit is configured to determine whether each section having the first period includes:
A sum of absolute values of the amplitudes of the acoustic signals;
A sum of square values of amplitudes of the acoustic signals;
A sum of predetermined frequency components of the acoustic signal;
A difference in the sum of absolute values of the amplitude of the acoustic signal in each of the first half and the second half,
A difference in the sum of squares of the amplitude of the acoustic signal in each of the first half and the second half;
The acoustic signal analyzer according to any one of claims 1 to 3, wherein any one of the difference between the sum totals of the predetermined frequency components of the acoustic signal in each of the first half and the second half is the first value of the section. The second sound volume information calculation unit, in each section having the second period,
A sum of absolute values of the amplitudes of the acoustic signals;
A sum of square values of amplitudes of the acoustic signals;
A sum of predetermined frequency components of the acoustic signal;
A difference in the sum of absolute values of the amplitude of the acoustic signal in each of the first half and the second half,
A difference in the sum of squares of the amplitude of the acoustic signal in each of the first half and the second half;
5. The acoustic signal analyzer according to claim 1, wherein any one of the difference between the sum totals of the predetermined frequency components of the acoustic signal in each of the first half and the second half is the second value of the section. The acoustic signal according to any one of claims 1 to 5, wherein the characteristic position detection unit detects a position where the evaluation value calculated by the evaluation value calculation unit is maximum or maximum for a continuous part of the acoustic signal. Analysis equipment. Obtaining an acoustic signal;
Calculating a first value relating to the volume of each of a plurality of sections having a first period of the acquired acoustic signal;
Calculating a second value related to the volume of each of a plurality of sections having a second period longer than the first period of the acquired acoustic signal;
Using the first value and the second value, the larger the first value and the larger the second value corresponding to the first value in time, the larger the time. Calculating a series evaluation value;
An acoustic signal analysis method comprising: detecting a section where the calculated evaluation value is maximum or maximum. A function to acquire an acoustic signal;
A function of calculating a first value relating to a volume of each of a plurality of sections having a first period of the acquired acoustic signal;
A function of calculating a second value related to the volume of each of a plurality of sections having a second period longer than the first period of the acquired acoustic signal;
Using the first value and the second value, the larger the first value and the larger the second value corresponding to the first value in time, the larger the time. A function to calculate the evaluation value of the series;
An acoustic signal analysis program for causing a computer to realize a function for detecting a section where the calculated evaluation value is maximum or maximum.

Images