JP2012118417A - Feature waveform extraction system and feature waveform extraction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately extract the middle out of music.SOLUTION: Frequency analysis is performed on waveform information of music and frequency variation values in a time axis direction are determined. A variation time having a plurality of high-order frequency variation values, in the determined frequency variation values, is extracted and waveform information of a fixed phrase continued from that variation time is extracted. A correlation coefficient of extracted phrase waveforms is then calculated to determine a repetition frequency and an average sound pressure in that phrase waveform is determined. A variation time of a phrase waveform having a plurality of or more repetition frequencies and a maximum average sound pressure is defined as a starting time of the middle, and waveform information until the next variation time is extracted as a feature waveform. Thereafter, processing for fixing a tempo value of the middle or rearrangement according to the tempo value is performed, positions of beats in the middles are matched and the middles are connected.

Description

  The present invention relates to a feature waveform extraction system capable of extracting rust from music.

  In general, the chorus part of a song is transposed from the previous phrase, or the sound pressure changes greatly over a long interlude or long notes / rests, or changes to a high frequency range. There are many. For this reason, conventionally, various methods have been proposed in which such rust characteristics can be used to extract rust information from music.

  For example, the following Patent Document 1 discloses a technique for extracting the position where the sound pressure is maximum from the music as rust, and the following Patent Document 2 has a sound pressure that is greater than a predetermined time. A technique is disclosed in which, when a tendency to continue and the frequency spectrum shifts to a high frequency region continues for a predetermined time or more, the portion is determined to be rusted.

  Furthermore, Patent Document 3 below discloses a technique in which a rust section extracted using sound pressure or the like is used as a rust when a section of rust repeatedly appears in music.

JP 2007-080304 A JP 2008-159252 A JP 2008-262043 A

  However, the rust extraction method as described above has the following problems.

  That is, in the method of extracting a rust area using only sound pressure as in Patent Document 1, if there is an interlude section or a long rest section, the section is excluded from rust candidates as “not rust”. For example, if an instrument with high sound pressure, such as a bass or beat, is played intermittently, the sound pressure will increase throughout the song, so you can determine exactly which part is chorus. It becomes difficult to judge.

  Similarly, in the method in which a large sound pressure continues for a predetermined time or more and a portion containing a lot of high frequency components is rusted as in Patent Document 2, the sound pressure such as bass or beat is high. When the instrument is played intermittently, it becomes difficult to extract the portion that becomes rust. In addition, even if narrowing is performed based on the high frequency component from the extracted section, if the wrong rust candidate is extracted based on the sound pressure, the wrong rust is extracted as a result. Will end up.

  Furthermore, even if the repetition frequency is calculated as in the above-mentioned Patent Document 3, in the music in which the A melody and the B melody are repeatedly played, only the chorus does not protrude and is repeatedly played. , It will be difficult to extract the rust accurately.

  Therefore, the present invention has been made paying attention to the above problems, and an object of the present invention is to provide a system that can extract a rust portion from music with high accuracy.

  That is, in order to solve the above-described problem, the present invention provides an input receiving unit that receives an input of music waveform information and a frequency analysis of the waveform information received by the input receiving unit to obtain a frequency fluctuation value in the time axis direction. Calculation means, fluctuation time extraction means for extracting a fluctuation time having a frequency fluctuation value greater than or equal to a predetermined value among the frequency fluctuation values obtained by the fluctuation calculation means, and waveform information in a predetermined section continuous from the fluctuation time Section frequency extracting means for extracting the frequency, frequency calculating means for determining the repetition frequency of the extracted section waveforms, sound pressure calculating means for determining the sound pressure in the section waveform, and the repetition frequency is a plurality of times or more. The feature is that the fluctuation time of the section waveform where the sound pressure is maximum is extracted as the feature start time, and the waveform information up to a predetermined end time is extracted as the feature waveform It is obtained by so and a shape extraction unit.

  As described above, if the position where the frequency variation is large is listed as a candidate for the start time of the chorus first, then the chorus candidate is extracted based on the sound pressure and the repetition frequency within the predetermined section from the start time. It is possible to extract rust accurately by giving priority to the part where the musical idea changes greatly (that is, the part that transposes, the part where a long note, a long note or a rest exists). That is, if a rust candidate is extracted first based on sound pressure, it will be difficult to extract a rust with little change in sound pressure, but there will always be a change in the thought at the part that switches to the rust. It becomes possible to accurately extract rust based on the above.

  Moreover, in such an invention, the predetermined area which continues from the fluctuation | variation time is extracted as a fixed area.

  In this way, if only the start time is determined based on the information of a certain section extracted from the fluctuation time, the end time is erroneously compared with the case where the end time of the chorus is estimated in advance and the repetition frequency is calculated. It is possible to prevent rust misjudgment based on estimation.

  Furthermore, when determining the end time, it is set as the next change time of the extracted start time.

  In this way, since the end time is determined only by the frequency fluctuation, it is possible to accurately extract the end time even when the sound pressure moves to the next melody without extremely changing. become able to.

  Further, a connection means for matching the tempo values of the extracted feature waveforms with the feature waveforms of other music pieces is also provided.

  In this way, when creating a chorus medley by connecting choruses, it is possible to smoothly transition to another chorus by keeping the tempo value of each chorus constant.

  Further, it is possible to provide connection means for rearranging the characteristic waveforms thus extracted in order of tempo values.

  Even in this case, it is possible to smoothly transition to another rust.

  The extracted feature waveforms can be linked to the feature waveforms of other music pieces by matching the positions of the beats.

  In this way, when creating a chorus medley by connecting choruses, it is possible to smoothly transition to other choruses by matching the beats of the choruses.

  In the present invention, the position where the frequency fluctuation is large is listed as a candidate for the start time of the chorus first, and then the chorus candidate is extracted based on the sound pressure and the repetition frequency within the predetermined section from the start time. It is possible to extract rust accurately by giving priority to the part where the musical idea changes greatly (that is, the part that transposes, the part where a long note, a long note or a rest exists). That is, if a rust candidate is extracted first based on sound pressure, it will be difficult to extract a rust with little change in sound pressure, but there will always be a change in the thought at the part that switches to the rust. It becomes possible to accurately extract rust based on the above.

Functional block of music feature waveform extraction system in one embodiment of the present invention The figure which shows the waveform extracted by each function implementation means in the form The figure which shows the process in the case of estimating the tempo value in the same form The figure which shows the relationship between the zero cross and tempo value in the same form The figure which shows the process which connects the rust in the form Flow chart for extracting rust in the same form Flow chart for connecting rust in the same form

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The feature waveform extraction system 100 in the present embodiment can extract a rust portion from one piece of music, and is a system in which a personal computer, a plurality of computers, and related devices are connected. Consists of. The functions shown in FIG. 1 are realized by cooperating these CPU, programs stored in storage devices such as ROM and RAM, disk drives, and the like.

  Specifically, the feature waveform extraction system 100 roughly divides the function of receiving music data and extracts the waveform of the portion that becomes a chorus, and creates a chorus medley by connecting the extracted chorus. It has the function to do.

  In such a configuration, the feature waveform extraction system 100 that extracts a rust feature waveform includes an analysis unit 2 that performs frequency analysis on the waveform information received by the input reception unit 1, and a frequency in the time axis direction from the analyzed frequency information. Fluctuation calculation means 3 for obtaining fluctuation values, fluctuation time extraction means 4 for extracting fluctuation times having a frequency fluctuation value of a predetermined value or more or a predetermined number or more (for example, the top 50), and a predetermined interval continuous from the fluctuation time Section waveform extraction means 5 for extracting the waveform information, frequency calculation means 6 for determining the repetition frequency of the extracted section waveforms, and sound pressure calculation means 7 for determining the sound pressure in the section waveform. The means 8 extracts the fluctuation time of the section waveform where the repetition frequency of the section waveform is a plurality of times and the sound pressure is the maximum as the feature start time, and And so as to extract a section up dynamic time the rust waveform information as the characteristic waveform. Hereafter, each function realization means in this Embodiment is demonstrated in detail.

  First, the input receiving means 1 receives the input of the music selected by the user. When receiving the input of the music, in addition to receiving the music data from the CD inserted in the device, the music data downloaded via the Internet or the like is received. There are various formats of music data, but in this embodiment, music data is received in the WAVE format. When receiving music data in this format, if music data in another format is received, it is converted to WAVE format by format conversion and accepted. The waveform information of the received music is shown in FIG. FIG. 2A shows waveform information that is an amplitude waveform of the received music, where the horizontal axis indicates the time axis and the vertical axis indicates the sound pressure.

  The analysis means 2 performs frequency analysis by fast Fourier transform (FFT) while shifting the received music piece for each predetermined analysis frame. Here, the fast Fourier transform is performed with 512 sampling points and 512 shift points (corresponding to about 11.6 msec). When this frequency analysis is performed, information is obtained in which the horizontal axis represents the frequency, the vertical axis represents the time of the analysis frame, and the height axis represents the spectrum.

  The fluctuation calculation means 3 calculates the magnitude of frequency fluctuation for each analysis frame based on the waveform information subjected to frequency analysis. When calculating this frequency variation, first, the flux function is extracted by calculating the sum (absolute amount) of the absolute value of the difference in each frequency bin of the adjacent analysis frame. As a result, the difference in frequency spectrum between adjacent analysis frames is extracted. As shown in FIG. 2B, the extracted flux function is represented as a time axis on the horizontal axis and the sum of absolute values of differences in each frequency bin (scalar amount) on the vertical axis. In the waveform in FIG. 2B, a signal with a large vertical axis indicates a portion where the frequency fluctuation of the adjacent analysis frame is large, and indicates a portion where the frequency spectrum has greatly changed from the time of the adjacent analysis frame. . However, there is a possibility that an erroneous determination may be made when the frequency fluctuates greatly due to noise only by the sum of the frequency fluctuations of adjacent analysis frames. Therefore, the extracted information is smoothed.

  In the smoothing process, the amplitude envelope of the time function of the flux is obtained using a moving average filter. As this moving average filter, in this embodiment, a moving average filter of 60 points (corresponding to about 696.6 msec) is used. Then, the flux function shown in FIG. 2B is replaced with a smoothed function as shown in FIG.

  Next, processing is performed so that the signal smoothed in this way can be determined as a tendency of a larger frequency fluctuation. In this process, a regression line is obtained using the least square method, and a differential value of the regression line (that is, the slope value in FIG. 2C) is obtained (FIG. 2D). There are two types of differential values, plus and minus. Here, in order to determine only the magnitude of frequency fluctuation, the differential value is extracted as an absolute value. Then, as shown in FIG. 2E, a signal in which all the values in FIG. 2D are positive values can be obtained. In FIG. 2C and FIG. 2D, the horizontal axis indicates the time axis, and the vertical axis indicates the value of the slope of the frequency variation of the smoothed flux function. If such processing is performed, even if noise is included or the frequency of adjacent analysis frames is changing rapidly, even if there is little frequency variation as a whole, is it rusted? It can be judged.

  The fluctuation time extraction means 4 extracts a predetermined number of times having a large inclination from the extracted signal information. The part marked with a circle in FIG. 2 (f) shows the signal at the extracted time. Here, a predetermined number of signals (for example, 150 signals) are extracted from the top, but for this number, a threshold is set in advance for the slope value, and the time of the signal having a slope exceeding the threshold is extracted. You may do it. The time extracted in this way is stored in the storage unit as a fluctuation time that is likely to correspond to the start time of the chorus.

  Next, the section waveform extraction means 5 extracts the waveform information of the music for a predetermined time from the fluctuation time as a reference. When extracting this waveform information, WAVE waveform information is extracted. The extracted waveform information is shown in FIG. In FIG. 2G, the horizontal axis indicates the time axis, and the vertical axis indicates the sound pressure, which is part of the signal information of FIG. Here, when extracting the waveform information of the music of a predetermined time from the fluctuation time, the time width to be extracted is set as a predetermined length of time, and the signal of the predetermined time width from each fluctuation time. To extract. Here, a section waveform of 2048 points (corresponding to about 23.7 sec) is extracted from the fluctuation time. At this point in time, if the fluctuation times are close, the extracted section waveforms may overlap.

  The frequency calculation means 6 calculates the similarity frequency of the extracted section waveform. When calculating this similarity frequency, the correlation coefficient between each section waveform is calculated. By calculating this correlation coefficient, a large correlation coefficient value can be obtained when the similarity of each section waveform is high, while a small correlation coefficient is obtained when the similarity of each section waveform is low. Can be obtained. Then, the combinations of the section waveforms are sorted in descending order of the correlation coefficient. At this time, among the combinations of sorted interval waveforms, as described above, there are those in which the change times are close and most of the intervals overlap. Are excluded. Here, it is assumed that combinations whose respective variation times are within 15 seconds are excluded, and two variation times that constitute the combination are excluded. Then, combinations having a correlation coefficient of a predetermined value or more or a predetermined number or more (specifically, combinations having a repetition frequency of 2 or more) are stored as rust candidates.

  The sound pressure calculation means 7 calculates the average sound pressure of the section waveform extracted as the rust candidate. In this calculation, the sound pressure is calculated using the extracted WAVE waveform information.

  The feature waveform extraction means 8 extracts a section waveform having the maximum average sound pressure from the combinations having the correlation coefficient equal to or larger than a predetermined value for the section waveform calculated in this way. Since the section waveform obtained in this way is a waveform sandwiched between two fluctuation times, they are set as the start time and end time of the chorus, and the chorus (WAVE file) is extracted.

  In this way, the frequency fluctuation is calculated, a predetermined number of fluctuation times having the largest frequency fluctuation value are extracted, and rust is extracted in consideration of the repetition frequency and sound pressure of the waveform data waveform in a certain section. By doing so, it becomes possible to accurately extract rust by paying attention to the part where the musical composition changes.

  Next, how to use the rust extracted in this way will be described. In this embodiment, the choruses extracted in this way are connected to each other so that a series of music pieces can be created as a chorus medley. The structure of the connecting means 9 for connecting the rust will be described.

  The extracted chorus has different tempo values and beats. For this reason, if rust is connected at random, rust cannot be changed naturally. Therefore, processing is performed in which the tempo values, beats, and the like of the chorus are matched and connected.

  First, when the tempo values are matched, the tempo value of the music is estimated and the time stretch process is performed so as to be a constant tempo value.

  In estimating the tempo value of the music, various methods can be adopted. In this embodiment, the following processing is performed.

  That is, the amplitude envelope is calculated from the received amplitude waveform of the music, and the frequency envelope is extracted by frequency analysis of the amplitude envelope (FIG. 3A). Then, the result of the short-time frequency analysis after multiplication of the frequency spectrum is added in the time axis direction for each frequency (FIG. 3B), and a tempo value having a predetermined number of frequency spectra from the maximum frequency 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 frequency spectrum of the half-half relationship (FIGS. 3C and 3D). Considering the frequency spectrum in the double-half relationship is, for example, when there is a strong beat at the first beat and the third beat in a 120 bpm song (ie when playing 4/4 time), one measure Because there is a peak value at the 1st and 3rd beats, there is a possibility that it is judged as 60bpm, which is half of 120bpm, depending on how the peak value is grasped. If there is a beat of another instrument between the beats to be played, it may be judged as 240 bpm. Therefore, in order to consider each case, the half-fold process is performed.

  Next, the zero cross value at which the amplitude waveform of the music crosses the time axis is calculated (the number of circles in the upper diagram of FIG. 4). At this time, the correlation between the zero cross value and the tempo value is stored in advance, and the tempo value is extracted from the previously calculated zero cross in consideration of the correlation. Extract from the replaced tempo value. In general, the zero cross value and the tempo value of the amplitude waveform have a relationship as shown in the lower diagram of FIG. 4, 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, the correct tempo value can be extracted by using such a correlation. Then, assuming that the chorus is played at the estimated tempo value, the chorus tempo value is determined, and the tempo values of all the music pieces are time stretched to a constant speed. When making this tempo value constant, the entire music may be time stretched, or only chorus may be time stretched. When time-stretching the entire piece of music, when extracting the beat time later, it is possible to extract the strong sound pressure part corresponding to the beat from the whole data, and to extract the beat time accurately. There is a merit that you can. On the other hand, when time-stretching only rust, there is an advantage that the overall processing time can be shortened.

  Next, a method for estimating the beat position will be described. There are various methods for estimating the position of the beat. Here, the following method is used.

  That is, in this embodiment, the previously estimated tempo value is used, and the portion with the highest sound pressure is extracted from the entire music. When extracting the portion with the highest sound pressure, the portion with the highest sound pressure is extracted from the amplitude waveform of the music whose tempo value is unified by time stretching. Alternatively, the amplitude envelope may be calculated from the amplitude waveform of the time stretched music, and the portion with the highest sound pressure may be extracted based on the amplitude envelope. Then, the time corresponding to the tempo value is determined as the beat position with reference to the portion with the highest sound pressure extracted as described above.

  As shown in FIG. 5, the connecting means 9 shifts the beat to the next chorus while adjusting the beat position and fading the volume, as shown in FIG.

  Next, a rust extraction method for music in the characteristic waveform extraction system 100 configured as described above will be described with reference to the flowchart of FIG.

<Rust extraction process>
When extracting rust from a piece of music, the input of the music is accepted as a WAVE file (step S1). Then, with respect to the amplitude waveform of the received music, the 512 analysis frames are shifted by 512 points while performing fast Fourier transform to perform frequency analysis (step S2), and the frequency information represented by the time of the frequency / spectrum / analysis frame is obtained. obtain.

  Next, the magnitude of the frequency variation along the time axis direction is calculated based on the frequency-analyzed information. In calculating the frequency fluctuation, first, the flux function is calculated by obtaining the sum of the absolute values of the differences in the frequency bins of the adjacent analysis frames (step S3), and this is smoothed. In the smoothing process, a moving average filter of 60 points is obtained to obtain the amplitude envelope of the flux function (step S4), and the least square method is used in order to grasp the partial inclination tendency of the amplitude envelope. Then, the regression line is obtained (step S5), and the absolute value of the differential value of the regression line is obtained (step S6).

  Then, the time of the top 150 signals is extracted with respect to the absolute value of the differential value, and the time is set as the start time of the rust candidate (step S7).

  Further, a section waveform of a certain section is extracted from the start time of each chorus (step S8), two arbitrary section waveforms are extracted to obtain a correlation coefficient (step S9), and a section waveform of a predetermined value or more is obtained. After removing combinations whose waveform fluctuation time is 15 seconds or less, a predetermined number (for example, 50) of combinations of waveforms having a high correlation coefficient are left and the others are deleted. (Step S10).

  Finally, among the remaining interval waveforms, the variation time of the interval waveform with the highest sound pressure is set as the start time (step S11), and the waveform information from the start time to the next variation time is extracted as rust (step S12). ).

<Rust concatenation>
Next, a flowchart in the case of creating a chorus medley musical piece by connecting choruses extracted in this way will be described with reference to FIG.

  First, the tempo value of the music is estimated in parallel with the process of extracting the music chorus. When this tempo value is estimated, based on the frequency-analyzed information (step T1), the frequency spectrum having the double-half relationship of all frequencies is multiplied by the frequency spectrum (step T2), The results of the short-time frequency analysis are added in the time axis direction for each frequency, and tempo values having a predetermined number of frequency spectra are rearranged as candidates from the maximum frequency spectrum and listed. Then, the upper half predetermined number (for example, two) candidates are extracted by adding the frequency spectra of the half-half relationship (step T3).

  At the same time, the zero cross value at which the amplitude waveform of the music crosses the time axis is calculated (step T4), and the tempo value is extracted in consideration of the correlation between the zero cross value stored in advance and the tempo value. (Step T5). Then, the tempo value closest to the tempo value is extracted from the tempo values rearranged first (step T6), and is estimated as the tempo value of the music including the chorus.

  After estimating the tempo value in this way, this time, a time stretch process is performed to a constant tempo value (step T7).

  Next, a portion having a strong sound pressure corresponding to a beat is extracted from the time stretched music, and the time corresponding to the tempo value is estimated from the time when the sound pressure is strong (step T8). Then, rust that has been time-stretched at random is extracted, and each rust is connected so that the respective beat times coincide (step T9).

  As described above, according to the above-described embodiment, positions with large frequency fluctuations are listed as rust start time candidates first, and then rust candidates based on sound pressure and repetition frequency within a predetermined section from the start time. Therefore, it is possible to extract the rust accurately by giving priority to the portion where the musical composition greatly changes. That is, if a rust candidate is extracted first based on sound pressure, it will be difficult to extract a rust with little change in sound pressure, but there will always be a change in the thought at the part that switches to the rust. It becomes possible to accurately extract rust based on the above.

  In addition, since a predetermined interval that is continuous from the fluctuation time is extracted as a fixed interval, the end time is estimated incorrectly compared to the case where the end time of rust is estimated in advance and the repetition frequency is calculated. It is possible to prevent rust misjudgment based on doing.

  Furthermore, when determining the end time, since it is set as the next fluctuation time of the start time, the end time is determined only by the frequency fluctuation, so that the sound pressure does not change drastically and the next melody is moved. Even in such a case, the end time can be accurately extracted.

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

  For example, in the above embodiment, when the frequency fluctuation value is obtained, the regression line by the smoothing process or the least square method is obtained, but the frequency fluctuation may be obtained from the sum of the differences between adjacent frequency bins. Good.

  Moreover, in the said embodiment, when calculating | requiring end time, the fluctuation | variation time next to start time is made into end time, However, Time when sound pressure becomes smaller than a threshold-value may be made into end time.

  Furthermore, in the above embodiment, the interval waveform of a certain interval is extracted from the variation time and the repetition frequency, sound pressure, etc. are calculated. However, this interval may be the interval until the next variation time. Good.

  Further, in the above embodiment, the application in the case of creating a chorus medley by connecting extracted choruses has been described as an example. However, it is cut off at a certain time (eg, 10 seconds) from the chorus start time. A digest version of rust may be created. In this case, the next chorus may be started with an interval of several seconds without connecting each chorus.

  In the above embodiment, time stretching is performed when creating a chorus medley. However, characteristic waveforms (rusts) are rearranged in order of tempo values without performing time stretching, and the rearranged order is used. You may make it perform connection. Even when the connections are performed in the order of the tempo values, smooth transition can be performed.

DESCRIPTION OF SYMBOLS 100 ... Feature waveform extraction system 1 ... Input reception means 2 ... Analysis means 3 ... Fluctuation calculation means 4 ... Fluctuation time extraction means 5 ... Section waveform extraction means 6 ... Frequency calculation Means 7 ... Sound pressure calculation means 8 ... Feature waveform extraction means 9 ... Connection means

Claims (7)

Input receiving means for receiving input of waveform information of the music;
Fluctuation calculating means for analyzing the frequency of the waveform information received by the input receiving means and obtaining a frequency fluctuation value in the time axis direction;
Of the frequency fluctuation values obtained by the fluctuation calculation means, fluctuation time extraction means for extracting a fluctuation time having a frequency fluctuation value greater than or equal to a predetermined value;
Section waveform extraction means for extracting waveform information in a predetermined section continuous from the fluctuation time;
A frequency calculating means for obtaining a repetition frequency between the extracted section waveforms;
A sound pressure calculating means for obtaining a sound pressure in the section waveform;
A feature waveform extraction means for extracting a fluctuation time of a section waveform in which the repetition frequency is a plurality of times or more and the sound pressure is maximum as a feature start time, and extracting waveform information up to a predetermined end time as a feature waveform;
A feature waveform extraction system for music, characterized by comprising: The characteristic waveform extraction system according to claim 1, wherein the predetermined section continuing from the fluctuation time is a fixed section. The characteristic waveform extraction system according to claim 1, wherein the end time is a variation time next to the extracted start time. The feature waveform extraction system according to claim 1, further comprising a connecting unit that matches the tempo value of the extracted feature waveform with a feature waveform of another music piece. The feature waveform extraction system according to claim 1, further comprising a connecting unit that rearranges and connects the extracted feature waveforms in order of tempo values. The feature waveform extraction system according to claim 1, further comprising connecting means for matching beat positions of the extracted feature waveforms with the feature waveforms of other music pieces. Receiving the input of the waveform information of the music;
A frequency analysis of the received waveform information to obtain a frequency fluctuation value in the time axis direction;
Extracting a fluctuation time having a frequency fluctuation value equal to or greater than a predetermined value from the obtained frequency fluctuation values;
Extracting waveform information in a predetermined interval continuous from the fluctuation time;
Obtaining a repetition frequency between the extracted section waveforms;
Obtaining a sound pressure in the interval waveform;
Extracting the fluctuation time of the section waveform in which the repetition frequency is a plurality of times or more and the sound pressure is maximum as a feature start time, and extracting waveform information up to a predetermined end time as a feature waveform;
A feature waveform extraction method for music, characterized by comprising:

Images