JPH0519792A - Pitch extractor - Google Patents

Abstract

(57) [Summary] [Object] The present invention eliminates the influence of the zero cross due to the overtone component when extracting the pitch from a voice signal such as a human voice or a musical sound, and eliminates the zero cross interval of the input waveform corresponding to one pitch period. The purpose is to enable the correct measurement of. [Configuration] The CPU first measures, for example, the interval between the zero-cross points at both ends of the waveform portion a having a positive polarity, and thereafter, the zero-cross points at both ends of the respective waveform portions b, c, and d having the same polarity. Each interval of is measured sequentially. Then, when the waveform portion d whose distance is almost equal to the distance of the waveform portion a and is closest in time is found, the distance between the start points (zero cross points) of the waveform portions a and d is determined. Extract as the pitch period of the input waveform. After that, the above series of processing is repeated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pitch extraction device for extracting the pitch (pitch) of its waveform from an input voice signal.

[0002]

2. Description of the Related Art Conventionally, there has been a demand for extracting the pitch frequency or pitch period (hereinafter, simply referred to as "pitch") from a voice signal having a repetitive waveform such as a singing voice of a person or a musical sound collected from a microphone. Is strong. If the pitch can be extracted, the extracted pitch information can be used for various processes, for example, a musical tone corresponding to the pitch of the pitch can be sounded from an electronic musical instrument or the like.

Conventionally, various techniques have been proposed as a technique for extracting a pitch from a voice signal, but among them, the method of directly extracting the pitch from a voice signal in a time domain requires a small amount of calculation and facilitates real-time processing. Is feasible.

As a first conventional example of the time domain pitch extraction method, for example, the time between adjacent zero-cross points (time positions where the amplitude becomes 0) of the input waveform is obtained as the pitch period of the waveform. There is.

As a second conventional example, there is also a system in which an input waveform is passed through a low pass filter to remove harmonic components and then a zero cross interval is extracted. Furthermore, the third
As a conventional example, a method of calculating the autocorrelation value of the input waveform and extracting the pitch period from the deviation amount of the autocorrelation at which the value has a peak is also proposed.

[0006]

However, since the human voice contains a lot of overtones and the zero-cross points appear several times during one pitch period of the input waveform, the zero-cross point simply occurs as in the first conventional example. The method of extracting the interval has a problem that a correct cycle cannot be obtained and a pitch extraction error is large.

Further, in the second conventional example, when the cutoff frequency of the low-pass filter is lowered, the pitch range in which the pitch can be extracted is limited to the low range, and conversely, when the cutoff frequency is raised, overtone components remain. There is a contradictory problem that the pitch extraction error increases.

Further, the third conventional example requires a very large amount of calculation, though it is a processing in the time domain, and is not suitable for the purpose of extracting the pitch of the input waveform in real time. have.

An object of the present invention is to eliminate the influence of the zero cross due to the above-mentioned overtone component and to enable the correct measurement of the zero cross interval of the input waveform corresponding to one pitch period.

[0010]

According to the present invention, first, an input waveform is zero-crossed from a negative or positive first polarity to the other second polarity, and then from the second polarity to the first polarity. It has the 1st space | interval measuring means which measures the space | interval to the zero-cross point which zero-crosses to the polarity of 1st as 1st space | interval.

Next, after the measurement by the first interval measuring means is completed, the input waveform is zero-crossed from the first polarity to the second polarity and then from the zero-cross point to the second polarity to the first polarity. It has the 2nd space | interval measuring means to measure sequentially the space | interval of each area | region to a zero cross point as a 2nd space | interval.

Further, a third interval for sequentially measuring the interval from the zero-cross point at which the measurement by the first interval measuring means is started.
It has a distance measuring means. Each of the above-mentioned first, second, and third interval measuring means is composed of, for example, a counter.

Interval storage means for storing the interval measured by the third interval measuring means at each time when the second measuring means starts measuring each second interval while sequentially updating it as cycle information. Have.

In addition, it is determined whether or not the first distance measured by the first distance measuring means and the respective second distances sequentially measured by the second distance measuring means match within a predetermined allowable error range. It has a determination means for sequentially determining.

Then, the period information stored in the time storage means is output as pitch information of the input waveform at the time when the determination means detects a match, and then each processing is performed in order from the measurement processing by the first interval measurement means. It has a repetition control means for sequentially extracting pitch information from the input waveform by repeating.

In the above-mentioned structure of the invention, at the time when the determining means does not detect the coincidence, the interval measured by the third interval measuring means exceeds the interval corresponding to the preset pitch period of the lowest note. Whether or not it is determined, and when it is determined that the time has been exceeded, it is possible to have a limiting unit that restarts each process from the measurement process by the first interval measuring unit.

[0017]

The voice signal is formed as a waveform in which one or more peaks of a waveform repeat every pitch cycle, and the peaks of each waveform within each pitch cycle have a similar shape between successive pitch cycles. Further, the intervals between the zero cross points at both ends of the peaks of the waveforms having the similar pitches adjacent to each other are substantially equal to each other.

In the present invention, focusing on this point, the first
The distance measuring means measures the distance between the zero-cross points at the ends of one peak of the waveform, and the second distance measuring means measures the distance between the zero-cross points at the opposite ends of the peak successively appearing thereafter.

Then, the determining means sequentially determines whether or not the above-mentioned first interval and each second interval match within a predetermined allowable error range, and the first interval at the time of match is determined. An interval from the corresponding mountain to the head of the mountain corresponding to the second interval is extracted as pitch information. This pitch information can be obtained as cycle information which is each measurement interval of the third measurement means stored in the interval storage means.

In this way, the pitch of the input waveform can be stably extracted regardless of the influence of the overtone component.

[0021]

Embodiments of the present invention will be described below with reference to the drawings. <Structure of Embodiment> FIG. 1 is an overall structure diagram of an embodiment of the present invention.

In the figure, for example, an analog voice signal such as a singing voice sung by one person or a single note of an analog musical instrument collected by a microphone (not shown) is passed through a capacitor 1 for removing a DC component. A / D
It is input to the converter 2 and converted into a digital signal.

The CPU (central processing unit) 3 is a RAM
(Random Access Memory) 4 as work memory, RO
By executing the program stored in M (Read Only Memory) 9, information about the pitch (pitch) of the input voice signal is extracted.

Next, the sound source 5 is controlled by the CPU 3
A musical tone signal having a note number based on the above pitch information and a tone color selected by a performer using a tone color selection switch (not shown) (for example, a tone color of a wind instrument such as a trumpet that is completely different from the above singing voice). To generate. After that, the generated musical tone signal is converted into an analog musical tone signal by the D / A converter 6, and is emitted as a musical tone from the speaker 8 via the amplifier 7.

In addition to the above program, the ROM 9 has various data used for pitch extraction and a table showing the correspondence between the cycle of the obtained input signal and the note number, as will be described later. .

With the above-described circuit configuration, pitch data (hereinafter, simply referred to as pitch when abbreviated) is extracted from the input audio signal according to the basic operation principle described below. <Basic Operation Principle of the Embodiment> Normally, a voice that has a clear pitch feeling and is played as a single tone, such as a human singing voice, is repeated every pitch period (for example), as shown in FIG. It is composed of substantially similar waveform elements (for example, a and d, b and e, c and f in the figure).

Such a voice has many zero-cross points within one pitch period, whose waveform crosses the reference level of zero amplitude, due to the overtone components contained in the voice.
Therefore, in order to obtain the pitch period from a waveform having such periodicity, it is necessary to determine which zero-cross point is used as the pitch period. It has the following characteristics.

That is, the time periods (eg, and) defined by the zero cross points (eg, Z1 to Z2 and Z7 to Z8) of adjacent waveforms of substantially the same pitch period (eg, a and d above) are substantially equal.

Focusing on this point, the present invention finds two sections (for example, and) having almost the same time between the adjacent zero-cross points, and thereby the starting points (for example, Z1 and Z7) of the two sections are found. ) And extract the time interval between
Extract the pitch period.

The operation of this embodiment, which operates according to the above-described basic operation principle, will be described below. <Outline of Embodiment> In FIG. 2, for example, the time in the section between the zero-cross points Z1 and Z2 of the waveform a is counted, and the count value COUNT 1 is held in the first counter. Next, the time periods of the waveforms b, c, d, ... Having the same polarity as the waveform a are sequentially counted,
The count value COUNT 2 is held in the second counter while being sequentially updated until a value approximately equal to COUNT 1 is obtained.

In this case, the time from the zero-cross point Z1 to the respective zero-cross points Z3, Z5, Z7 of the waveforms b, c, d, and, are counted by the third counter. Since the count value COUNT 3 has a possibility of becoming the pitch period of this input voice signal, it is sequentially updated and stored in the variable PITCOUNTER in the RAM 4 until the count value of the pitch period is obtained.

When a section having a time substantially equal to the time of the section is found, the counter value corresponding to the time between the zero cross points Z1 and Z7 is set to the memory variable PIT COUN.
It is read from TER and is determined as the pitch period. The note number corresponding to the pitch cycle is RO
It is calculated from the table in M9. <Operation of this Embodiment> Next, the operation of obtaining the pitch of the input waveform according to the above-described operation principle will be described with reference to the operation flowcharts of FIGS. These operation flowcharts are realized as operations in which the CPU 3 of FIG. 1 executes the control program stored in the ROM 9.

In FIG. 3, first, the RAM 4 is initially set. That is, a status flag (STATUS FLAG, see FIG. 2) indicating which status of the signal waveform corresponds to the time when the audio signal is read, a variable ENV indicating the size of the waveform envelope of the input audio signal, and a pitch. The variable PIT that holds the note number corresponding to the pitch obtained as a result of extraction, the variable ON NOW that holds the note number that is currently sounding, the count values of the three counters of the first, second, and third Variable to hold
All contents such as COUNT 1, COUNT 2, COUNT 3 and variable ADMEM that holds the previous A / D converted value are cleared (step S301). A PIT or ONNOW value of 0 means that there is no pitch-extracted sound or there is no sound currently being sounded.

Next, after waiting for the end of the A / D conversion for one sampling of the input audio signal (S302), the next step
In S303, the A / D converted data is set in the register reg1 in the CPU3.

After that, the processing is branched into three according to the status flag (step S304). The first is Figure 2.
Since the status flag of is 0, the step
Proceed to S305. In the status flag = 0 state, the variable AD in which the previous A / D conversion value is written in order to detect the zero-cross point Z1 from the negative side to the positive side of the input waveform.
It is determined whether the value of MEM is negative and the value of the register reg1 in which the current A / D conversion value is stored is positive.

In this case, the determination in step S305 is NO, and the process proceeds to step S307. Then, the value of reg1 in which the A / D conversion value is set this time is transferred to the variable ADMEM (step
S307), envelope extraction (step S308), musical note on / note off processing (pronounce / silence processing, step
Then, the process proceeds to S309) and waits for the end of A / D conversion again (step S302).

The processing relating to the envelope extraction (step S308) and the note-on / note-off (step S309) of musical tones is shown in the respective operation flowcharts of FIGS.

In FIG. 5, first, the absolute value of the current A / D conversion value (contents of reg1) is calculated (step S501),
After that, it is passed through a low-pass filter (by executing a digital filtering operation) to shape the input waveform (step S502). And the value obtained as a result is the variable EN
It is stored in V (step S503).

In the note-on / note-off processing of FIG. 6,
First, it is determined whether or not the size of the variable ENV exceeds a predetermined threshold value (THRESHOLD) and the value of the variable PIT holding the note number is larger than 0 (step S604).

If the determination is NO, the variable ON NOW is cleared in order to indicate that the sound source 5 of FIG. S609) and returns to the main routine (FIG. 3).

If the determination in step S604 is YES, it is determined in step S605 whether or not there is a sound currently being sounded. If the determination is YES, the process proceeds to step S608.
If the determination in step S605 is no, each value of the variable ON NOW indicating the note number being sounded and the extracted variable PIT indicating the note number is compared (step S606). As a result of the comparison, if they match, it is not necessary to re-pronounce, so the process returns to the main routine (Fig. 3). If they do not match, the note number to be pronounced has changed, and the currently sounding note is muted. To the sound source 5 (step S607), and then the process proceeds to step S608.

In step S608, a sounding instruction corresponding to the variable PIT indicating the note number is given, and the value of the variable PIT is substituted into the variable ON NOW, and the variable PIT is currently set.
It is indicated that a musical tone having a pitch corresponding to is being sounded (step S608). After that, the process returns to the main routine (Fig. 3).

The above processing is repeated until the zero cross point from the negative side to the positive side (Z1 in FIG. 2) is reached, and eventually the determination in step S305 in FIG. 3 becomes YES, and the status
The flag STATUS FLAG is set to 1 (step S306).

After that, the same operation as described above is performed, and the process returns to step S304 again. Where STATUS FLAG = 1
Since it has become the state of, the process moves to step S310. In this state, in order to count the time exemplified in the section of FIG. 2, the variable COUNT 1 (hereinafter, referred to as the first counter) as the first counter.
(Counter of the counter COUNT 1) is also incremented to count the time instantiated in the segment.
The increment of the variable COUNT 3 (hereinafter, referred to as the third counter COUNT 3) as the counter of each is started.

First, in step S310, the value of the first counter COUNT 1 is incremented by 1 (step S31
0). Next, the zero-cross point from the positive side to the negative side of the input waveform, that is, the zero-cross point that is the end point of the section in FIG.
Whether the value of the variable ADMEM in which the previous A / D conversion value was written to detect Z2 is positive and the value of register reg 1 in which the current A / D conversion value is written is negative It is determined whether or not (step S311).

If the determination is NO, the third counter COUN
The value of T 3 is incremented by 1 (step S313). After that,
Steps S311 → S313 → S307 → S308 → S309 → S302 ~ S304 ~
The loop processing of S311 ... Is repeated, and the first counter COUNT 1 and the third counter CO
The value of UNT3 is increased by 1.

When the determination in step S311 is YES and the zero-cross point Z2 from the positive side to the negative side is reached, the status flag is set to 2 (step S312). After that, the routine proceeds to step S313, where COUNT 3 is incremented by 1, and then the routine proceeds to steps S307 → S308 → S309 and returns to step S302.

After the loop processing as described above is performed, FIG.
The counter value corresponding to the time of the section is the first counter
Held in COUNT 1. Next, since the status flag is set to 2 in step S312, this time step S3
From 02 to S304, the process proceeds from step S401 to S402 in FIG.

In step S402, in order to detect the zero-cross point (Z3 in FIG. 4) from the negative side to the positive side, the same processing as in step S305 in FIG. 3, that is, the value of the variable ADMEM (the previous A / It is determined whether the D conversion value) is negative and the value of register reg1 (current A / D conversion value) is positive.

While the determination is NO, S402 → S313
(Fig. 3) → S307 to S309 → S302 to S304 → S401 (Fig. 4) → S4
The loop process of 02 ... Is repeated, and the count value of the third counter COUNT 3 is incremented by 1 for each repetition (S313).

When the determination in step S402 is YES and the zero-cross point Z3 from the negative side to the positive side is reached, the status flag is set to 3, and the value of COUNT 3 at this time (corresponding to the time in FIG. 2) ) Is stored in the variable PITCOUNTER (step S403). The value of this PITCOUNTER is held as a candidate for the cycle corresponding to the pitch to be extracted, that is, as a possibility of becoming the count value of the section in FIG.

Then, steps S403 to S313 (FIG. 3) to S3.
09-Go to S302-S304-S401 (Fig. 4). This time STATUS F
Since the state of LAG = 3 is reached, the process proceeds from step S401 to step S404, the count value of the second counter COUNT 2 is incremented by 1, and the counting of the section of FIG. 2 is started from this point.

Next, the zero-cross point from the positive side to the negative side of the input waveform, that is, the zero-cross point Z4 corresponding to the end point of the section of FIG. 2 is detected, so that the value of the variable ADEM (the previous A / D conversion value) is positive and the value of register reg1 (current
It is determined whether the A / D conversion value) is negative (step
S405).

While the determination is NO, steps S404 → S4
05 → S313 (Fig. 3) ~ S309 → S302 ~ S304 → S401 (Fig. 4) →
The loop processing of S404 is repeated, and the value of the second counter COUNT 2 and the value of the third counter COUNT 3 are incremented by 1 for each iteration.

If the determination in step S405 is YES, it means that the zero-cross point Z4 has been detected, and the status flag is changed to 4 here (step S406). After that, the process proceeds to step S313 of FIG. 3 and after COUNT 3 is incremented by 1, the process proceeds to steps S307 → S308 → S309 to S401 (FIG. 4).

Since STATUS FLAG = 4 this time, the process proceeds from step S401 to step S407. In this state, the divisions are sequentially compared with the divisions of FIG. 2, and when a division having a time substantially equal to that of the division is found, a processing operation of setting the division time as the pitch period of the input waveform is performed. .

First, the first counter obtained so far
It is determined whether or not the respective values of COUNT 1 and the second counter COUNT 2 are substantially equal, that is, whether or not the absolute value of the difference between the respective values of COUNT 1 and COUNT 2 is within the allowable error range. (Step S407). This tolerance range is
It depends on the sampling frequency and pitch frequency of the input signal, but is determined empirically.

In this case, since COUNT 1 and COUNT 2 respectively hold the count values for the section of FIG. 2, the determination in step S407 is NO and the process proceeds to step S408. Here, if noise or the like is mixed in the input voice signal, this may cause the determination of step S407 to remain abnormal for a long time, and the musical tone pronunciation instruction may not be given.

In order to prevent the occurrence of such an abnormal state,
The value of a third counter COUNT 3 which counts the time which should be the pitch period of the voice signal is constantly monitored and has a predetermined limit value which has a value which is much greater than the pitch period of the lowest pitch of the voice signal. It is determined whether or not the limit is exceeded (step S408).

This limit value is also the same as the above-mentioned allowable error range.
Determined empirically. For example, the sampling frequency is 20KH
If z and the minimum extracted pitch (pitch) are 110.5 Hz (note name: equivalent to A1), then 20000 ÷ 110.5 = 180.995, so the limit value should be around 180.

If the determination in step S408 is YES, that is, if it is in an abnormal state, the process proceeds to step S314 in FIG. 3, and the sound source 5 in FIG. 2 is instructed to mute all the sounds being sounded. , The variable ON NOW is cleared. After that, returning to step S301, the above-described pitch extraction processing operation is executed again.

If the determination in step S408 is no, the process proceeds to step S409, the value of the second counter COUNT 2 is cleared, and the status flag is returned to 2. Then, again, the process for obtaining the next zero cross is started.
At this time, the value held in the variable PIT COUNTER becomes invalid, but the third counter COUNT 3 continues to count up, and the first counter COUNT 1 also holds the count value in the section in FIG. ing.

As described above, the status flag 2
After returning to the state of No.2, as in the previous time, the status flag goes to No.3, No.4, and the second counter COUNT
2 holds the count value of the section in FIG. 2, and the variable PIT COUNTER holds the count value (from FIG. 2) from the start of the section in the figure to the zero-cross point Z5 of the section.

However, since the value of COUNT 1 (corresponding to the classification time) and the value of COUNT 2 (corresponding to the classification time) do not match this time, the determination in step S407 is NO. Therefore, the process proceeds from step S408 to step S409, and the status flag is restarted from the state of 2.

Then, when the status flag becomes the state of 4 again next, COUNT 2 holds the count value of the section of FIG. 2, while PIT COUNTER holds the count value of the time of FIG. is doing. Then, in step S407, YES is determined for the first time, and the step
Proceed to S410.

The correct value of the pitch period of the input waveform at this time is the third counter CO held in PIT COUNTER.
It becomes the value of UNT 3 (Fig. 2). Then, the note number corresponding to this value is read from the table in the ROM 9 of FIG. 1 and is transferred to the variable PIT (step S410).

Next, the status flag and all counters are cleared and reset to the initial state (step S411). After that, the process proceeds to the note-on / off process of steps S307 to S309 of FIG. 3, and in this process, a new tone generation instruction of the musical tone having the note number corresponding to the pitch (pitch) of the input voice signal is given. .

That is, in the flow of FIG. 6 corresponding to step S309, if it is determined in step S605 that there is no currently sounding tone (YES) after the determination in step S604 is YES, as described above, In step S608, a sounding instruction is given according to the variable PIT indicating the note number. If it is determined in step S605 that there is a note currently being sounded (NO), as described above, in step S606, the variable ON NOW indicating the note number being sounded and the variable PIT indicating the extracted note number are set. Each value is compared. As a result of the comparison, if both match, it is not necessary to re-pronounce, so the process returns to the main routine (FIG. 3). If they do not match, the note number to be pronounced has changed, and the note currently being pronounced is changed. The sound source 5 is instructed to mute the sound (step S607), and then, in step S608, a new sounding instruction is issued in accordance with the variable PIT in which the new note number is stored.

After the above pronunciation instruction, step S307 in FIG.
~ S302 ~ S304, the same operation as before is repeated from the status of STATUS FLAG = 0. In the embodiment described above, the waveform portion on the positive side with respect to the reference level of the amplitude value 0 of the waveform (for example, the waveforms a, b, c, d ... In FIG. 2).
Regarding, the pitch was extracted by comparing the times of the respective sections, but the present invention is not limited to this, and the pitch extraction operation can be performed in exactly the same manner by using the waveform portion on the negative side.

Further, the above-mentioned pitch extraction operation is performed in parallel for each waveform on the positive side and the negative side with respect to the reference level, and the value of each third counter COUNT 3 allows a predetermined error. If they match, for example, two COUNs
If the average value of the values of T 3 is obtained as the pitch period of the input waveform, more accurate pitch extraction is possible.

Further, in the note-on / off processing of FIG. 6, when the pitch of the musical tone being sounded changes,
Although processing is performed such that the sound is muted and then regenerated, in this case, for example, if the envelope does not change so much, only the pitch is changed without muting the sound source 5 of FIG. It is also possible to give instructions.

[0072]

According to the present invention, by extracting the time points at which the intervals between zero cross points are almost equal and the peaks of the waveforms are determined to be similar to each other, the pitch of the input waveform can be determined regardless of the influence of the overtone component. It becomes possible to extract stably.

As a result, it is possible to reliably extract the pitch indicating the pitch from a voice waveform such as a singing voice or musical sound including overtones.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention.

FIG. 2 is a waveform diagram of an input voice signal.

FIG. 3 is an operation flowchart (No. 1) of the present embodiment.

FIG. 4 is an operation flowchart (No. 2) of the present embodiment.

FIG. 5 is an operation flowchart regarding envelope detection according to the present embodiment.

FIG. 6 is an operation flowchart regarding note on / off according to the present embodiment.

[Explanation of symbols]

1 capacitor 2 A / D converter 3 CPU 4 RAM 5 sound sources 6 D / A converter 7 amplifier 8 speakers 9 ROM

Claims (2)

[Claims] 1. An interval from a zero-cross point at which an input waveform zero-crosses from a negative or positive first polarity to the other second polarity and then to a zero-cross point at which the second polarity zero-crosses to the first polarity. Is measured as the first interval
After the measurement by the interval measuring means and the first interval measuring means, the input waveform is zero-crossed from the first polarity to the second polarity from a zero-cross point, and then from the second polarity to the second polarity. 1
The second interval measuring means for sequentially measuring the intervals of the respective sections up to the zero-cross point where the zero-crossing occurs to the polarity of the second interval and the interval from the zero-cross point at which the measurement by the first interval measuring means is started. Third interval measuring means for sequentially measuring, and the interval measured by the third interval measuring means at each time point when the second measuring means starts measuring each of the second intervals as cycle information. The interval storage means for storing while updating, the first interval measured by the first interval measurement means, and the second intervals sequentially measured by the second interval measurement means have a predetermined allowable error range. Determining means for sequentially determining whether or not there is a match, and outputting the period information stored in the time storage means at the time when the determining means detects the match as pitch information of the input waveform, And a repetition control unit that sequentially extracts pitch information from the input waveform by repeating each of the processes from the measurement process by the first interval measuring unit. 2. At the time when the judgment means does not detect the coincidence, whether or not the interval measured by the third interval measuring means exceeds an interval corresponding to a preset pitch period of the lowest note. If it is determined that it has been exceeded,
The pitch extracting device according to claim 1, further comprising a limiting unit that re-executes each process from the measurement process by the first interval measuring unit.