JP2011164497A - Tempo value detecting device and tempo value detection method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To estimate a precise tempo value with consideration of a double-half tempo value in a short processing time when estimating a tempo value of a music piece. <P>SOLUTION: An amplitude envelope is calculated from an amplitude waveform of a received music piece, and frequency analysis is carried out on the amplitude envelope to extract a power spectrum. As to all frequencies over the whole music piece, power spectra of frequencies in a double-half relation are multiplied by the power spectrum of the frequency, and the frequencies are rearranged sequentially from a frequency with a strong power spectrum. A value of zero-crossing where an amplitude waveform of the music piece crosses a temporal axis is calculated, while a correlation between the value of zero-crossing and the tempo value is stored. A tempo value is extracted while considering the correlation based on the calculated zero-crossing, and a tempo value closest to the extracted tempo value is selected from rearranged tempo values. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a tempo value detection device capable of detecting a tempo value from a music piece. Here, the “tempo value” refers to a value (bpm = beat per minute) ”indicating how many beats that appear periodically in a music piece within one minute.

  2. Description of the Related Art Conventionally, in the field of processing information related to music, there has been a great deal of research that makes it possible to estimate the tempo value from the waveform information of the music.

  For example, in the following Patent Document 1 (Japanese Patent Laid-Open No. 5-027751), in order to estimate the tempo value of a song, the frequency of the music input from the music is analyzed to calculate the power spectrum and its differential value, A method is disclosed in which a measure time length is calculated by obtaining a maximum value of the autocorrelation function of the differential value, and a tempo value is calculated from the measure time length and the number of beats.

  The following Patent Document 2 (Japanese Patent Laid-Open No. 7-064544) also discloses a method for estimating the tempo value of a music piece. In this processing method, the level of a specific frequency band of the input acoustic signal is extracted, the peak value of the extracted level and the generation time are sequentially stored, and the average value of the peak values of the stored level is calculated. Of the peak values of the stored level, two peak value levels exceeding the calculated average value are selected. Then, for each occurrence interval of the two selected levels, a search is made as to whether or not the other stored level exists, and when another level exists for each occurrence interval, the occurrence interval is determined. Tempo data is generated based on the above, and when no other level exists at each occurrence interval, the other stored level is selected and the same search is performed to estimate the tempo value. .

  In addition to such a patent document, there is a tempo value estimation method as described in the following non-patent document 1 as a non-patent document. This tempo value estimation method uses 13-dimensional parameters as a preliminary preparation, learns classification with GMM (Gaussian mixture model), and selects tempo value candidates. Based on the above, the music is classified, and the tempo value candidate most suitable for the corresponding class is selected as the tempo value in the corresponding unit section. Thereafter, the median tempo value of the entire music is obtained from the tempo values obtained for each unit section, and output as the estimated tempo value.

JP-A-5-027751 Japanese Unexamined Patent Publication No. 7-064544

Xiao et al, "Using a statistic model to capture the association between timbre and perceived tempo", Proc. of 11th ISMIR pp. 659-662 2008.

  However, when estimating the tempo value using such a method, there are the following problems.

  That is, in the method according to Patent Document 1, the user inputs the beat in one measure in advance, the time length of the one measure is obtained, and the tempo value is obtained from the beat. You have to input the beat in the. For this reason, there is a problem that it takes a lot of time to automatically obtain tempo values for a large number of music pieces.

  Further, with the technique according to Patent Document 2, the tempo value can be estimated without inputting a beat in such a measure, but with the technique according to Patent Document 2, the time width of beats and beats, which are peak values, is estimated. Since the tempo value is obtained from the tempo value, there is a possibility that the tempo value will be greatly shifted depending on the performance method of the music having strong and weak beats. That is, if there is a strong beat at the 1st and 3rd beats in a 120 bpm song (ie when playing 4/4 time), there will be a peak value at the 1st and 3rd beats in 1 bar Therefore, depending on how the peak value is taken, there is a possibility that it will be judged as 60 bpm, which is half of 120 bpm. On the other hand, in the case of 120 bpm music, if there is a beat of another musical instrument between adjacent beats, there is a possibility that a so-called double half tempo value problem that it is judged to be 240 bpm may occur.

  On the other hand, in the method described in Non-Patent Document 1, the tempo value candidates are listed, and the tempo value candidates obtained based on the average MFCC are used to solve the double-half tempo value problem. The value can be estimated. However, in such a tempo value estimation method, a plurality of peak values are obtained based on an autocorrelation function, and the tempo value is obtained from the peak value and the reciprocal of the time length of the peak value. When the length is long, the calculation amount of the autocorrelation function becomes enormous.

  Therefore, the present invention has been made paying attention to the above-described problem, and an object thereof is to provide a tempo value estimation method capable of estimating a tempo value with high accuracy in consideration of a double half tempo value in a short processing time. .

  That is, in order to solve the above-described problem, the present invention provides a tempo value estimation device that estimates the tempo of the music from the input music, receiving means for receiving the music input, and frequency analysis of the amplitude waveform in the received music A frequency analysis means for extracting the power spectrum, and a half processing means for adding to the power spectrum of each frequency a value that takes into account the power spectrum of the frequency that is in a half-half relationship for all frequencies throughout the music; Means for calculating a zero cross value at which the amplitude waveform in the music received by the means crosses the time axis, a correlation storage means for storing a correlation between the zero cross value and the tempo value in advance, and the half-half processing means Corresponding in order from the maximum power spectrum among the power spectra obtained by The wave number is extracted as a tempo value candidate, and an estimation means for estimating one tempo value from the tempo value candidates by providing the correlation from the zero cross calculated by the zero cross calculation means is provided. It is.

  That is, in general, the zero-cross value of the amplitude waveform and the tempo value have a correlation, and the higher the tempo value, the higher the zero-cross value. Conversely, if the tempo value is low, the zero cross value is also low. Therefore, in the present invention, since the correct tempo value is extracted by using such a correlation, the tempo value can be accurately estimated with a reduced amount of calculation.

  Further, in such an invention, an envelope calculating means for calculating the amplitude envelope of the waveform from the amplitude waveform in the received music is provided, and the power spectrum of the entire music is extracted by performing frequency analysis on the calculated envelope waveform information. To.

  In this way, the received waveform itself is not frequency-analyzed, but the low-frequency envelope waveform is frequency-analyzed. The power spectrum can be extracted.

  Furthermore, when the amplitude envelope is calculated by the envelope calculating means, an average value of a waveform obtained by squaring the amplitude waveform is calculated for each predetermined smoothing section.

  In this way, downsampling can be performed while expressing the calculation of the amplitude envelope.

  In the present invention, in a tempo value estimation device for estimating the tempo of a music from the input music, a receiving means for receiving the music input, and a frequency analysis for extracting a power spectrum by frequency analysis of the amplitude waveform in the received music Means, a half processing means for adding to the power spectrum of each frequency a value that takes into account the power spectrum of the frequency that is in a half-half relationship for all frequencies throughout the music, and an amplitude waveform in the music received by the receiving means Zero-cross calculating means for calculating a zero-cross value that crosses the time axis, correlation storage means for storing the correlation between the zero-cross value and the tempo value in advance, and the maximum power spectrum obtained by the half-half processing means Extract corresponding frequencies as tempo value candidates in order from the power spectrum. In addition, since the correlation is taken into consideration from the zero cross calculated by the zero cross calculation means and an estimation means for estimating one tempo value from the tempo value candidates is provided, the calculation amount is reduced and the accuracy is increased. A good tempo value can be estimated.

Functional block diagram of a tempo value estimation apparatus according to an embodiment of the present invention The figure which shows the outline | summary of the smoothing downsampling in the same form The figure which shows the power spectrum and the double half processing in the same form The figure which shows distribution of the tempo value of the music which calculates the acoustic parameter The figure which shows the value of the zero cross in the form The figure which shows the correlation of the zero cross and tempo value in the form Flowchart for tempo value estimation in the same form

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The tempo value estimation device 1 in this embodiment is configured by an information processing device such as a personal computer, a portable music playback device, an electronic musical instrument, etc., which receives information on the amplitude waveform of the music, It is possible to electronically estimate the tempo value that indicates how many beats per minute are output. More specifically, the tempo position estimating apparatus 1 first receives an input of a music piece, calculates an amplitude envelope by data processing the amplitude waveform in the amplitude-time domain of the received music piece. Then, frequency analysis of the calculated amplitude envelope is performed to extract the power spectrum of the entire music, and among those power spectra, the power spectrum of the double frequency of each frequency is multiplied by the power spectrum of that frequency, and the maximum The corresponding frequencies are rearranged in order from the power spectrum. On the other hand, in parallel with this, a zero cross value at which the amplitude waveform in the amplitude-time domain crosses the time axis is calculated in advance, and information indicating the correlation between the zero cross value and the tempo value is stored in advance. The optimum tempo value is extracted from the rearranged tempo value candidates using the zero cross value while referring to the information indicating the correlation. Details of the function realizing means for performing these processes will be described below. Note that these function realizing means are configured by cooperating a program and the like with the CPU and storage medium of the information processing apparatus, the input device, the display output device, the communication device, and the like.

  First, the accepting unit 2 accepts an input of music pieces owned by the user. When accepting this music, the WAVE format digital signal, which is widely used as a CD sound source, the analog signal acquired via a microphone, etc., A / D converted signal, the information distributed on the Internet radio and the mp3 format Accept information. As shown in FIG. 2, the music information received in this way is stored as an amplitude waveform in an amplitude-time region in which the vertical axis is amplitude and the horizontal axis is time axis. When estimating the tempo value from this amplitude waveform, human beings generally specify the tempo value by counting the number of beats in unit time, but in order to extract the tempo value of this amplitude waveform, the appearance period of the periodically appearing beats It is necessary to ask. Normally, a strong sound as an accent is output at the time when a beat exists, and in the amplitude waveform, time having a large amplitude periodically exists. On the other hand, the waveform of music changes greatly depending on the type of musical instrument, and includes many waveform components that change at high frequencies. Therefore, in order to estimate the tempo value from the amplitude waveform, the low frequency resolution is increased. Here, “low frequency resolution” indicates the frequency resolution of the frequency analysis in the low frequency band of the waveform information. As the frequency band corresponding to the tempo value, the tempo value is practically included in the range of about 50 bpm-220 bpm, so the low frequency band is about 1 Hz-3 Hz. Therefore, in order to ensure this low frequency resolution, envelope information and the like are calculated so that a fine and accurate tempo value can be extracted.

  The envelope calculation means 3 calculates the amplitude envelope from the amplitude waveform in the amplitude-time domain of the music received by the reception means 2 and down-samples to ensure low frequency resolution. As a method for securing low frequency resolution, there are a method of increasing the number of sampling points of frequency analysis as a current method, or a method of down-sampling that lowers the sampling frequency as preprocessing. When increasing the number of sampling points for frequency analysis, if the precision of the required tempo estimation is 1 bpm and Fourier analysis is used for frequency analysis, and the waveform information to be analyzed is a sampling frequency of 44100 Hz, the number of sampling points is approximately 2.6 million or more. Necessary (required number of sampling points 2646000 = sampling frequency 44100Hz / (1bpm / 60). When downsampling is performed, the frequency band to be analyzed is limited to low frequency, and the samples required for frequency analysis according to the number of downsampling In this case, however, aliasing will occur if band limitation is not performed.When these two methods are compared, the number of sample points is relatively large when the number of sample points for frequency analysis is increased. Because it is not considered a realistic method, try downsampling and do aliasing. The band limitation for preventing the frequency from being generated is a low frequency band where the frequency required for tempo value estimation is 4 Hz or less, so by calculating the amplitude envelope of the waveform information, the effect is almost the same as a low-pass filter that passes that band. It is possible to calculate the amplitude envelope first before downsampling.

  Here, when calculating the amplitude envelope, smoothing downsampling is used. When performing this smoothing downsampling, as shown in FIG. 2, the average value of the acoustic power spectrum, which is the square value of the amplitude, is obtained in the smoothing section that is the section for each number of downsampling, and the value is downsampled. Is extracted as waveform information. Here, the smoothing section corresponds to the time length of the sampling period in the amplitude waveform after smoothing down-sampling, and obtaining the acoustic power spectrum corresponds to full-wave rectification. The outline of the smoothed downsampling will be described with reference to FIG. 2. FIG. 2 (a) shows the amplitude waveform received by the receiving means 2, and FIG. 2 (b) shows the result of smoothed downsampling. The waveform in FIG. 2B is obtained by squaring the amplitude waveform, and an average value of the waveform is calculated and stored as a value in the smoothed section in the waveform. These values are actually calculated by the following equation.

  As shown in FIG. 2B and Formula 1, in the smoothed downsampling, downsampling is performed while expressing the calculation of the amplitude envelope by obtaining the average value of the sound power spectrum for each fixed section. Here, considering aliasing, obtaining the average value of the acoustic power spectrum in the smoothed section corresponds to passing through a strong low-pass filter, and the necessary band is limited. When smoothed down-sampling is performed, all waveform information values become positive and the DC component (0 Hz) value increases. Therefore, when performing frequency analysis on the waveform information after smoothed down-sampling, an average such as Z value conversion is used. Perform normalization to set the value to “0”.

  The frequency analysis unit 4 calculates the power spectrum in the power spectrum-frequency domain by performing Fourier transform (FFT) on the envelope information calculated by the envelope calculation unit 3. Since this Fourier transform is performed at every analysis time, a plurality of power spectra-frequency domain power spectra are calculated for the entire music. By the way, the power spectrum analyzed in this way basically has a large power spectrum at the frequency position where the beat is strong, but in reality, it is within one measure like a 4/4 time signature. In many cases, the frequency having the maximum power spectrum does not always correspond to the tempo value as in the case where strong beats exist at the first and third beats. Therefore, tempo value candidates for tempo value estimation are extracted by the next half-half processing means 5.

  In the double-half processing means 5, first, a tempo value is obtained from the frequency by obtaining a frequency having a high dominance in the tempo value estimation. Here, the “highly dominant frequency” indicates a frequency having a large power spectrum over the entire music, and indicates a frequency having a power spectrum larger than a predetermined threshold. At this time, since there is a double / half tempo value problem in the tempo value estimation, it is considered that frequencies having a relationship of double or half of the frequency having a large power spectrum are also included in the highly dominant frequency. Therefore, when enumerating tempo value candidates that can be correct, enumerate tempo values corresponding to frequencies with high dominance in consideration of the double-half relationship. Further, when enumerating, since it is considered that the dominance of the tempo value is a dominant tempo value candidate, ranking is performed using, for example, a power spectrum corresponding to the tempo value (frequency) as an index. The flow of this process is shown in FIG. 3. In short, the amplitude waveform of the music is subjected to frequency analysis for a short time, and all the frequencies in the analysis result are multiplied by the value of the power spectrum corresponding to the frequency of the half-fold relationship ( FIG. 3 (a)). That is, the 100Hz power spectrum is multiplied by the power spectrum of 50Hz and 200Hz, the 101Hz power spectrum is multiplied by the power spectrum of 50.5Hz and 202Hz, and so on. Multiply by the value of. Next, these power spectra are summed on the time axis (FIG. 3 (b)) to obtain a reference frequency (hereinafter referred to as a reference frequency) corresponding to a large power spectrum peak (FIG. 3 (c)). Is the standard tempo value. As a result, the power spectra corresponding to the tempo values of the half-half relationship are emphasized to each other, and tempo value candidates having a frequency with a large power spectrum over the entire music can be obtained. If there is a halve-related tempo value among the enumerated tempo value candidates, the power spectrum of the corresponding tempo value is added, and the power spectrum is rearranged in descending order (FIG. 3 (d)). As a result, the power spectrum of the tempo value candidate having the double / half relationship is increased, so that the tempo value having a large power spectrum and the double / half relationship is positioned in a high order.

  Next, after ranking the candidates for the tempo in this way, the correct tempo value is extracted from these candidates using another index. In general, for the index of music speed, it is considered that the “acoustic parameter” calculated from the tempo value and the waveform information of the music corresponds to this, and the correlation between this acoustic parameter and the tempo value is obtained, and the acoustic parameter having a high correlation is obtained. Is an index of the speed of music. The waveform information of the music for which the acoustic parameters are calculated is a group of 100 music, and the tempo value is a value that is experienced by a music experienced person for each music. The distribution of tempo values of the used music group is shown in FIG. 4 with the horizontal axis representing the tempo value class and the vertical axis representing the number of songs. As shown in FIG. 4, the tempo values of the music group used are normally distributed centering on the average tempo value of 100 bpm, and it can be confirmed that there is no bias in the tempo values. At this time, the sound source of the music used was waveform information in WAVE format, and was a monaural sound source with a sampling frequency of 44100 Hz and a time length of 4-6 min. Table 1 shows the correlation between the calculated acoustic parameter and the tempo value.

When the calculation of the acoustic parameters shown in Table 1 is described, the average value Γ _ave1 and standard deviation Γ _1SD of the acoustic power spectrum is the acoustic power spectrum Γ _{1 at} time x (x = 0, 1, 2,... X). [x], and the sound power spectrum Γ ₁ [x] is obtained by the following equation 2 with the original waveform information at time x (x = 0, 1, 2, ... X) as f [x]. It is done.

Fluxの平均値Γ_ave2 及び標準偏差Γ_2SD は、時刻x"(x" =0, 1, 2, ・・・X")におけるFluxΓ₂[x"]から得られる。FluxΓ₂[x"]は、元の波形情報f [x] を短時間周波数解析することで得られ、時刻x"、周波数y における元の波形情報の周波数解析結果をFx" [y] とし数３で求められる。このときの短時間周波数解析はSFFTを用いることとし、用いたサンプル数Y は8192とした。また、この数ごとにシフトして解析することとした。

The average value Γ _ave2 and standard deviation Γ _2SD of Flux are obtained from Flux Γ ₂ [x ”] at time x ″ (x ″ = 0, 1, 2,... X ″). FluxΓ ₂ [x "] is obtained by short-time frequency analysis of the original waveform information f [x]. The frequency analysis result of the original waveform information at time x" and frequency y is Fx "[y]. In this case, SFFT is used for the short-time frequency analysis, and the number of samples Y used is 8192. The analysis is shifted for each number.

Centroid's mean value Γ _ave3 and standard deviation Γ _3SD are obtained from CentroidΓ ₃ [x "] at time x" (x "= 0, 1, 2, ... X"). CentroidΓ ₃ [x "] is The original waveform information f [x] is obtained by performing frequency analysis for a short time, and the frequency analysis result of the original waveform information at time x "and frequency y is Fx" [y]. In the short-time frequency analysis, SFFT was used in the same manner as FluxΓ ₂ [x "], and the number of samples Y used was 8192. In addition, the analysis was shifted for each number.

ZeroCrossingsΓ ₄ (hereinafter referred to as zero cross) is the number of times that the original amplitude waveform information f [x] changes from positive to negative and from negative to positive (that is, the number of times crossing the time axis), as shown in FIG. It is obtained by counting this. Here, when comparing the correlation shown in Table 1 above, ZeroCrossings shows a high correlation with the tempo value. For this reason, ZeroCrossings calculated from the music is used as an index of the speed of the music, and the zero cross calculation means 6 calculates a value for this speed index. At this time, the zero cross value is an index for selecting a powerful tempo value. The tempo value and the zero cross are subjected to regression analysis, and the tempo value obtained from the zero cross is stored as a Z tempo value as an index of the final music speed.

  The correlation storage means 7 stores the correlation between the zero cross value and the Z tempo value. In general, as shown in FIG. 6, the relationship between the zero cross value and the Z tempo value increases as the zero cross value increases. Therefore, these correspondences are stored in the correlation storage means 7.

  The estimation means 8 adds up the power spectra at the peak position where the frequency is half-half from the frequencies rearranged in order of the power spectrum by the half-half processing means 5 and extracts candidates for the top two tempo values. Then, a Z tempo value is calculated from the zero cross value by a regression equation, and a tempo value candidate close to the Z tempo value among the two enumerated tempo value candidates is output as a finally estimated tempo value.

  Next, the flow of processing for estimating the tempo value of a song using each function realizing means configured as described above will be described with reference to the flowchart of FIG.

  First, when estimating the tempo value from the music, WAVE format music data is received (step S1), and the smoothed down-sampling is performed by squaring the amplitude waveform and storing the average value for each smoothed section. Thus, downsampling is performed while expressing the calculation of the amplitude envelope (step S2). The downsampled amplitude envelope is then normalized (step S3).

  Next, the smoothed down-sampled waveform information is subjected to frequency analysis for a short time (step S4), and as shown in FIG. (Step S5). Then, the result of the short-time frequency analysis after multiplication of the power spectrum is added in the time axis direction for each frequency (FIG. 3B), and a tempo value having a predetermined number of power spectra from the maximum power spectrum is selected as a candidate. Rearranged and enumerated, and finally, the upper half predetermined number (for example, two) candidates are extracted by adding the power spectra of the half-half relationship (FIGS. 3C and 3D) (step S6).

  In parallel with this, the zero cross value is calculated from the amplitude waveform received in step S1 (step S11), and the Z tempo value is extracted with reference to the information stored in the correlation storage means 7 (step S12). . Then, a tempo value close to the extracted Z tempo value is selected (step S13), and is finally output as an estimated tempo value (step S14).

  As described above, according to the above-described embodiment, the reception unit 2 that receives the input of music, the envelope calculation unit 3 that calculates the amplitude envelope of the waveform from the time-amplitude waveform in the received music, and the amplitude envelope as a frequency. Frequency analysis means 4 for analyzing and extracting a power spectrum, double-half processing means 5 for multiplying the power spectrum of that frequency by the power spectrum of the frequency that is half-fold for all frequencies throughout the music, and time in the music Obtained by the zero-cross calculating means 6 for calculating the zero-cross value that crosses the time axis of the amplitude waveform, the correlation storage means 7 for storing the correlation between the zero-cross value and the tempo value in advance, and the half-half processing means 5 In addition, a plurality of tempo value candidates are extracted in order from the maximum power spectrum in the power spectrum. The estimation unit 8 for estimating the tempo value close to the Z tempo value from the tempo value candidates is provided by taking into account the correlation from the zero cross calculated by the cross calculation unit 6, thereby reducing the amount of calculation. In this state, the tempo value can be estimated with high accuracy.

  Hereinafter, evaluation of tempo estimation in the present embodiment will be described. The evaluation of the tempo estimation in the above embodiment is performed based on the evaluation of the existing tempo estimation method performed by Xiao (Non-Patent Document 1 above). Whether or not the estimated tempo value is correct is determined based on Xiao's tempo estimation evaluation method when Equation 5 is satisfied.

  In Equation 5, the ratio between the estimated tempo value and the correct tempo value is obtained, and the correct answer is obtained when the error is 0.04 or less. The evaluation index is a correct answer rate that is a ratio of correct music in the total number of music groups to be evaluated. The group of songs used for the evaluation was a group of 100 songs with no bias in the tempo values used to investigate the index of music speed.

  Table 2 shows the evaluation results of the tempo estimation in the above embodiment.

  In Table 2, the correct rate of the tempo value for each candidate in the proposed tempo estimation is set to “correct rate 1”, and the correct rate of the tempo value of the double-half relationship that is a double or half tempo value for each candidate is set to “correct rate 2 " From Table 2, it can be confirmed that the proposed method has an accuracy rate of 83%, the accuracy rate is 10% due to the difference in candidates, and the accuracy rate is 11% due to the difference in the double-half relationship. Therefore, a correct answer rate of 93% was obtained when the candidate error was correct, and 94% was obtained when the double-half error was correct.

  Here, compared with the existing tempo estimation methods shown in Table 1, the highest accuracy rate of the existing method is 61%, whereas the above method has a high accuracy rate of 83%. I understand. When referring to the fact that the music group used for the evaluation is different at this time, the music group used for the evaluation has no bias in the tempo value as shown in FIG. 4, and the difference in the correct answer rate due to the difference is considered to be slight. It is done. Therefore, it can be confirmed that a high accuracy rate was obtained compared with the existing tempo estimation method.

  In addition, this invention can be implemented in various aspects, without being limited to the said embodiment.

  For example, in the envelope calculation means 3 in the above embodiment, the amplitude envelope is squared and smoothed downsampling is used. However, the envelope information may be calculated as the amplitude information, or the smoothed down A method of increasing the number of sample points without sampling or reducing the sampling frequency as preprocessing may be used.

  Further, in the above-described embodiment, the half spectrum process is performed for all the frequencies subjected to the frequency analysis, and the power spectrum is multiplied. However, the half spectrum process is performed for frequencies having a power spectrum stronger than a predetermined threshold for all the music pieces. May be performed.

DESCRIPTION OF SYMBOLS 1 ... Tempo value estimation apparatus 2 ... Acceptance means 3 ... Envelope calculation means 4 ... Frequency analysis means 5 ... Half-half processing means 6 ... Zero cross calculation means 7 ... Correlation storage means 8 ... Estimation means

Claims (4)

In the tempo value estimation device for estimating the tempo of the music from the input music,
A receiving means for receiving input of music;
A frequency analysis means for extracting a power spectrum by performing frequency analysis on the amplitude waveform in the received music;
Double processing means for adding to the power spectrum of each frequency a value that takes into account the power spectrum of the frequency that is in a half-half relationship for all frequencies throughout the song,
Zero-cross calculating means for calculating a zero-cross value at which the amplitude waveform in the music received by the receiving means crosses the time axis;
Correlation storage means for storing the correlation between the zero cross value and the tempo value in advance;
Extracting a plurality of frequencies as tempo value candidates in order from the maximum power spectrum among the power spectrum obtained by the half-half processing means, taking into account the correlation from the zero cross calculated by the zero cross calculation means, and the tempo value An estimation means for estimating one tempo value from the candidates;
A tempo value estimation device characterized by comprising: An envelope calculating means for calculating an amplitude envelope of the waveform from an amplitude waveform in the music received by the receiving means is provided, and the calculated envelope waveform information is subjected to frequency analysis to extract the power spectrum of the entire music. The tempo value estimation apparatus according to 1. 3. The tempo value estimation according to claim 2, wherein when the envelope calculating unit calculates the amplitude envelope from the amplitude waveform, an average value of a waveform obtained by squaring the amplitude waveform is calculated for each predetermined smoothing section. apparatus. In the tempo value estimation method for estimating the tempo of the music from the input music,
Receiving the input of the music;
Extracting the power spectrum by performing frequency analysis on the amplitude waveform of the received music;
Adding to the power spectrum of each frequency a value that takes into account the power spectrum of the frequency that is half-fold for all frequencies throughout the song;
Calculating a zero cross value at which the amplitude waveform in the received music crosses the time axis;
A step in which the correlation between the zero cross value and the tempo value is stored in advance;
A plurality of frequencies are sequentially extracted as tempo value candidates from the maximum power spectrum among the power spectrum obtained by the halving process, and the correlation is taken into consideration from the calculated zero cross, and one frequency is selected from the tempo value candidates. Estimating a tempo value of
A tempo value estimation method characterized by comprising:

Images