JPH09198095A - Pitch detecting device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately find a pitch cycle fast and accurately even when a speech waveform is a complicated waveform containing an overtone component or when a fine noise is superposed on the speech waveform. SOLUTION: A binarization part 8 which has a masking zone nearby the zero level converts a digital speech signal into a binary signal. Then, a timer 9 measures the inversion intervals of the binary signal to find successive zero- crossing intervals of the digital speech signal and stores them in a RAM 10. A pitch arithmetic part 11 assumes that the pitch cycle is the sum of 2n pieces of zero-crossing interval data as to n=1-4, calculates a reproduction rate as the degree of matching in each pitch cycle of each zero-crossing interval data constituting one pitch cycle, and employs the assumption by which the highest reproduction rate is obtained to find the pitch cycle.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pitch detecting device for detecting a pitch period or pitch frequency of a voice waveform.

[0002]

2. Description of the Related Art A pitch period (or pitch frequency) is one of the parameters that characterize a speech waveform, and a technique for detecting the pitch period of this speech waveform is generally used in speech analysis / synthesis systems, speech coding systems and the like. It is used. Recently, some karaoke systems also detect the pitch period of the voice of the singer, and are used for scoring the singing.

Conventionally, there have been the following methods for detecting the pitch period of voice. (1) Zero cross method Assuming that the voice waveform is very close to a sine wave, the voice waveform crosses the zero level line from the negative direction to the positive direction, then from the positive direction to the negative direction, and again from the negative direction to the positive direction. Since a monotonous change of crossing the zero level line is repeated, the pitch period is given by the time interval across the zero level line in the same direction. According to this idea, the zero-cross method is a method of simply measuring two zero-cross intervals and setting them as pitch periods. Further, as an idea similar to this, there is also a method of measuring the interval of the timing at which the instantaneous value of the voice waveform reaches the maximum value or the minimum value and sets it as the pitch cycle.

(2) Autocorrelation method In this autocorrelation method, time-series samples x (1), x (2), ... Obtained by sampling the speech waveform at a constant sampling period are used. The pitch period is obtained by calculating the autocorrelation function R (r). R (r) = 1 / N · Σ {x (n) / x (n + r)} (In the above equation, Σ is an operator that finds the sum within {} in the range of n = 1 to N · r. That is, the r is variously changed, the autocorrelation function R (r) is obtained for each r, and the pitch period of the speech waveform is calculated from r when R (r) is maximum (that is, the autocorrelation is maximum). calculate.

[0005]

By the way, the above-mentioned zero-cross method can detect the pitch period at a relatively low cost and at a high speed, but on the other hand, since human voice contains many harmonic components, it is accurate. There is a problem that it is not possible to detect different pitch periods. Further, the above-described autocorrelation method can detect the pitch period to some extent accurately, but the amount of calculation is enormous and the detection time is long. In addition, the cost is increased.

The present invention overcomes the above-mentioned problems, basically calculates the pitch period of the voice waveform by measuring the zero-crossing interval, and prevents the adverse effects caused by adopting this method. Therefore, it is an object of the present invention to provide a pitch detecting device having a low cost structure and capable of detecting a pitch period accurately and at high speed.

[0007]

According to the present invention, a speech waveform is binarized with a zero level as a reference and a binary signal is output.
Quantizing means, zero-crossing interval measuring means for measuring continuous zero-crossing intervals t ₁ , t ₂ , ... Of the voice waveform based on the binary signal, and variously changing n (n is an integer of 1 or more) , For each n, the sum of 2n zero-crossing intervals T = (t ₁ +
It is assumed that t ₂ + ... T _2n ) is a pitch period, and adjacent m periods (m is 2) based on the zero crossing intervals t ₁ , t ₂ ,.
The degree of coincidence of the voice waveforms during each pitch period of the above integer) is calculated, and the degree of coincidence of the voice waveforms is the highest n
And a pitch calculating means for determining a pitch cycle by selecting a pitch period, wherein the binarizing means uses a masking band within a predetermined range from the zero level, and the speech waveform A binary signal that is inverted only when it changes across the masking band, and maintains the binary signal just before entering the masking band if the speech waveform is within the masking band. The detection device is the main point. The invention according to claim 2 is characterized in that the width of the masking band is controlled according to the amplitude of the voice waveform.

[0008]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments will be described to make the present invention easier to understand. Such an embodiment shows one aspect of the present invention, and does not limit the present invention, and can be arbitrarily changed within the scope of the present invention.

A. Configuration of Embodiment FIG. 1 is a block diagram showing the configuration of an embodiment in which the present invention is applied to a karaoke system. The present embodiment relates to a part of the karaoke system that is used for scoring a song by a singer. In FIG. 1, reference numeral 1 is a CD (compact disc) on which a digital music signal is recorded. The digital music signals recorded on the CD 1 are sequentially reproduced in synchronization with the clock having the sampling frequency fs = 44.1 kHz. 2 is a vocal extraction unit, CD1
A signal corresponding to a vocal sound (hereinafter referred to as a digital model signal) is extracted from the digital music signal reproduced from. As an example, a digital model signal can be obtained by a process of extracting a signal in a frequency band including a voice band of a digital music signal reproduced from the CD 1 with a bandpass filter. If a medium recording only vocal sounds is available, the digital music signal reproduced from the medium may be used as it is as a digital model signal. 3 is a microphone, CD
The singing voice of the singer singing along with the reproduction of 1 is sampled and output as an analog audio signal. 4 is an A / D converter,
An analog audio signal from the microphone 1 is sampled at the same sampling frequency fs = 44.1 kHz as in the case of reproducing the CD1.
It is sampled in synchronization with the z clock and converted into a digital audio signal.

Reference numeral 5 denotes a DC removing unit which applies DC to the digital audio signal and digital model signal which are sequentially supplied.
The removal processing is performed, and the digital audio signal and the digital model signal from which the components in the low frequency band that can be regarded as DC, for example, the band of 0 Hz to 50 Hz are removed are output. 6
Is an LPF (low-pass filter), which removes a component having a frequency of, for example, 500 kHz or more from each of the digital audio signal and the digital model signal output by the DC removing unit 5, and outputs the result. These DC removing section 5 and L
The PF 6 selects and outputs only the component within the band of 50 to 500 Hz for each of the digital voice signal and the digital model signal.

Reference numeral 7 is a 4 × oversampling unit,
Digital voice signal and digital model signal that have passed through the LPF 6 (both sampling frequency fs = 44.1
(kHz) is interpolated, converted into a signal having a quadruple sampling frequency, and output.

FIG. 2 shows the quadruple oversampling unit 7
Among them, a circuit configuration necessary for processing one of a digital voice signal and a digital model signal (hereinafter referred to as an input digital signal) is illustrated. In this figure, the latch 71 receives and holds an input digital signal by being supplied with a clock corresponding to the sampling frequency fs. The delay devices 72, 72, ... Are cascade-connected to the latter stage of the latch 71 as shown. These delay devices 72 are supplied with a clock having a frequency four times as high as the sampling frequency fs, whereby the latch 7
The input signal held at 1 is sequentially shifted, and the input signal is sequentially delayed by one clock cycle, and each delayed signal is output. 73, 73, ... Are multipliers, and 74, 74 ,.
Interpolation calculation is performed in which a predetermined interpolation coefficient string is convoluted with each output signal of 72, .... With the above configuration, the interpolation operation is executed in synchronization with the clock having a frequency four times the sampling frequency fs, and the interpolated digital signal is sequentially output from the adder 74 at the final stage.

The quadruple oversampling unit 7 is a means provided to improve the accuracy in obtaining the pitch period. That is, in the present embodiment, the pitch period of each digital signal is obtained by measuring the time interval between the zero cross points of each of the digital voice signal and the digital model signal. Therefore, in order to improve the accuracy of measuring the pitch period, it is necessary to improve the accuracy of detecting the position of the zero cross point on the time axis. Therefore, by inserting the 4-fold oversampling unit 7, the time density of each sample of the digital voice signal and the digital sample signal is quadrupled to improve the detection accuracy of the position of each zero cross point. In this example, oversampling is performed by curve interpolation, but in view of the cost problem, linear interpolation that provides some accuracy may be used.

Reference numeral 8 denotes a binarizing unit which binarizes the levels of the digital voice signal and the digital model signal output from the quadruple oversampling unit 7. This binarization basically determines whether the input digital signal is positive or negative based on the zero level, and outputs "1" when the input digital signal is positive and outputs "0" when the input digital signal is negative. is there. That is, the binarization unit 8 is means for outputting a binary signal in which "0" / "1" is inverted every time the input digital signal crosses the zero level. However, in the present embodiment, even when the binarization is performed, a range of ± Δ around the zero level is set as a masking band, and even if the input digital signal has a minute vibration within the ± Δ masking band, it is still required. The binary signal is not inverted due to a slight vibration.

FIG. 3 exemplifies a circuit configuration necessary for processing one of the digital voice signal and the digital sample signal (hereinafter, referred to as an input digital signal) in the binarizing unit 8. In this figure, reference numeral 81 is an absolute value detecting section for detecting the absolute value of the input digital signal. 82
Is a comparison unit, which compares the absolute value of the input digital signal detected by the absolute value detection unit 81 with a predetermined value Δ, and if the absolute value exceeds Δ, exceeds “1”; Outputs "0". Reference numeral 83 denotes a sample and hold unit, which outputs the input digital signal as it is while the comparison unit 82 outputs “1” (sampling state), and the comparison unit 82 outputs “0” during the comparison unit. The input digital signal immediately before the output signal of 82 changes from "1" to "0" is held and output (hold state). Reference numeral 84 is a comparison unit, which determines whether the output signal of the sample-hold unit 83 is positive or negative with reference to the zero level.
If it is negative, a binary signal of "0" is output.

According to the above configuration, when the input digital signal is outside the range of ± Δ, it is output as it is through the sample hold unit 83. Further, when the input digital signal enters the masking band of zero level ± Δ, the value of the input digital signal immediately before that falls in the sample hold unit 83.
The binary signal output from the comparator 84 is not inverted during the period in which the holding operation is performed. Therefore, when the input digital signal changes across the masking band of zero level ± Δ, the binary signal is inverted when the crossing of the masking band is completed. On the other hand, when the input digital signal enters the masking band of zero level ± Δ and moves up and down in the masking band without crossing it, even if the input digital signal crosses the zero level, the sample hold unit Since the output signal value of 83 does not cross the zero level, the inversion of the binary signal does not occur.

In FIG. 3, the circuit preceding the comparison unit 84 may be replaced with the circuit shown in FIG. This figure 4
In 85, reference numerals 85 and 86 respectively compare the input digital signal with the reference level, and output "1" when the input digital signal is higher than the reference level and "0" when it is lower than the reference level. The reference level + Δ is given to the comparison unit 85, and the reference level −Δ is given to the comparison unit 86. 87 is a latch for holding an input digital signal, and 88 is a selector for selecting and outputting the input digital signal or the output signal of the latch 87. Reference numeral 89 is a control unit, and comparison units 85 and 86
Latch 87 and selector 89 based on each output signal of
Control. That is, it is as follows.

A. When the output signals of the comparators 85 and 86 are both "1" or "0", the input digital signal is outside the masking band of zero level ± Δ. In this case, the control unit 89 uses the latch 8
7 is set in the sample state, and the selector 88 outputs the input digital signal. b. When the output signal of the comparison unit 85 is “0” and the output signal of the comparison unit 86 is “1” This is the case where the input digital signal is inside the masking band of zero level ± Δ. In this case, the control unit 89 puts the latch 87 in the hold state when the input digital signal enters the masking band, and causes the selector 88 to output the output signal of the latch 87.

In FIG. 1, 9 is a time interval at which inversion of each output signal of the binarization unit 8 corresponding to a digital voice signal and a digital model signal occurs, that is, a time at which a zero cross point of each of these digital signals occurs. A timer for timing the interval, and 10 is a RAM for storing the timing result of the timer 9.

FIG. 5 is a block diagram showing the timer 9 and the RAM 10 together with their control systems. It should be noted that this figure shows only a portion necessary for processing corresponding to one of the digital voice signal and the digital model signal. FIG.
In the figure, 91 is a delay device, and 92 is an exclusive OR circuit. These configure a differentiating circuit 90 that differentiates the binary signal output from the binarizing unit 8, and outputs a pulse each time the inversion of the binary signal occurs. The timer 9 is reset each time an output pulse from the differentiating circuit 90 is applied, and counts clocks having a constant frequency of 4 fs until the next reset after this reset.

The count value of the timer 9 is given to the latch 93 as input data. The latch 93 fetches and holds the count value of the timer 9 immediately before reset by being supplied with the output pulse from the differentiating circuit 90. The count value held in the latch 93 is the number of clocks having the frequency 4fs output from the time when the inversion of the previous binary signal was detected to the time when the current inversion was detected, and therefore, the zero cross point Can be said to represent the time interval at which Therefore, hereinafter, the data held in the latch 93 will be referred to as zero-cross interval data.

The write control section 94 sequentially reads the zero-cross interval data in the latch 93 every time the output pulse from the differentiating circuit 90 is given, and the zero-cross interval data within a certain range is equal to or more than a predetermined value (timer 9 When the count value of is large), a limit is provided and written in the RAM 10, and when it is less than a predetermined value (the count value of the timer 9 is small), a limit is provided and the RAM 10 is discarded without being written. Thus, only the zero-crossing interval data within a certain range is stored in the RAM 10.
The reason why the data is written in is to prevent the wrong pitch period from being calculated due to the use of the zero-cross interval data, which is not valid as the time interval of the zero-cross points of the audio signal, in the calculation.

The pitch calculator 11 in FIG. 1 is a RAM
By referring to the zero-cross interval data stored in 10, the pitch period of each of the digital voice signal and the digital model signal is calculated.

If the digital audio signal or the like is a sine wave, the zero level line is crossed at the start point and the end point of the sine wave for one period, and the zero level is only once at the midpoint between these zero cross points. Cross the line. Therefore, the pitch period can be obtained by adding two consecutive zero-cross interval data.

However, since a digital voice signal representing a human voice waveform contains many overtone components, a waveform for one pitch period has three or more zeros between the start point and the end point of the pitch period. In some cases, a cross point is included, and in such a case, the correct pitch period cannot be obtained even if two consecutive zero-cross interval data are added.

Therefore, in the present embodiment, it is assumed that one pitch period has a length corresponding to the sum of 2n zero-cross interval data for each of a plurality of types of integers n. Then, the pitch period is calculated under each assumption, and 1
The degree to which the generation timing of each zero cross point within the pitch cycle matches between the pitch cycles is determined. The details of the detection of the degree of coincidence of the generation timings of the zero cross points will be described later. Then, the pitch cycle with the highest degree of coincidence is selected as the true pitch cycle. This is based on the property of an audio signal that a large waveform change does not occur within a short time.

Next, in FIG. 1, reference numeral 12 is a level detection unit for detecting the level of each of the digital voice signal output by the A / D converter 4 and the digital model signal output by the vocal extraction unit 2. , Outputs a signal representing each level.

Reference numeral 13 denotes a scoring unit, which detects the deviation of the pitch cycle of each of the digital voice signal and the digital model signal obtained by the pitch calculating unit 11 and the level detecting unit 12.
Comprehensively evaluate the deviation of both signal levels obtained by
Score the song of the singer. The scoring result is displayed on the display unit 14.

B. Operation of Embodiment Hereinafter, the operation of this embodiment will be described. When the song is selected by the singer, digital music signals are sequentially reproduced from the CD 1 corresponding to the song. And the vocal extraction unit 2
As a result, a digital model signal is extracted from the digital music signal and output to the DC removing section 5 and the level detecting section 12. On the other hand, when the CD 1 is reproduced, the singer starts singing, and the singing voice is collected by the microphone 3 and output as an analog audio signal. This analog audio signal is converted into a digital audio signal through the A / D converter 4, and is output to the DC removing section 5 and the level detecting section 12.

The digital voice signal and the digital model signal are sequentially passed through the DC removing section 5 and the LPF 6, whereby unnecessary frequency band signals are removed and a waveform consisting of only components within the frequency band of human voice is represented. The digital signals are output to the 4 × oversampling unit 7, respectively.

Then, the digital voice signal and the digital model signal are processed by the 4 × oversampling unit 7.
Each is interpolated on the time axis, converted into a signal having a quadruple sampling frequency, output, and converted into a binary signal by the binarization unit 8.

FIG. 6 shows the 4 × oversampling unit 7
The operation of FIG. In FIG. 6A, the horizontal straight line is a zero level line. In addition, the plot of ○ is shown along the sinusoidal signal waveform, but the latter plot is a digital audio signal (digital sample signal).
Represents the individual original samples that make up the above, and the former represents the original signal waveform that is the parent of these original samples. Further, three plots marked with X are inserted between the plots marked with ◯ representing each original sample, and these plots each represent the interpolated sample obtained by the 4 × oversampling unit 7. .

FIG. 6B shows a binary signal obtained when only the original sample (marked with a circle) is given to the binarizing unit 8 without performing the 4 times oversampling, and FIG. ) Is the original sample (○)
And a binary signal obtained when the interpolation sample (x mark) is given to the binarization unit 8. For convenience of description, these figures show an example in which the digital audio signal (digital model signal) does not include vibration of a level smaller than the masking band of the binarization unit 8.

Here, the digital audio signal or the like is sampled at a constant sampling period regardless of the signal waveform. Therefore, when the digital audio signal or the like repeats the same waveform, as shown in FIG. 6A, which timing the instantaneous value is sampled differs depending on each waveform. For this reason, if the sampling period is rough, as shown in FIG. 6B, when the pitch period is switched, a binary signal having a different waveform may be obtained although the waveform is the same. However, in the case where binarization is performed after performing 4 times oversampling of a digital audio signal or the like as in the present embodiment,
As shown in FIG. 6C, a binary signal that inverts at a timing close to the original zero-cross point is obtained, and the problem shown in FIG. 6B is prevented.

FIGS. 7A to 7D illustrate the operation of the binarizing unit 8. First, in FIG. 7A, a sinusoidal signal waveform represents a digital audio signal (digital model signal) output from the quadruple oversampling unit 7, and a horizontal line represents a zero level line. FIG. 7 (b)
Shows the operation of the sample hold unit 83 in FIG. As shown in this figure, the sample hold unit 8
When the digital audio signal (digital model signal) as an input signal is outside the zero level ± Δ masking band, 3 is in a sampled state (denoted as “S” in the figure) and zero level ± Δ masking. If it is inside the band, it is in the hold state (indicated as "H" in the figure).
Is done. As a result of such control of the sample hold unit 83, the signal waveform input to the comparison unit 84 is as shown in FIG.
7C, the binary signal obtained from the comparison unit 84 is shown in FIG. 7D. In this way, when the digital voice signal (digital model signal) changes across the masking band of zero level ± Δ, the binary signal is inverted at the time when the crossing of the masking band is completed. Even if the digital audio signal (digital model signal) includes a minute vibration part with an amplitude of ± Δ or less, if the digital audio signal (digital model signal) is within the zero level ± Δ masking band. Since the sample hold unit 83 performs the previous value holding operation, the binary signal is not inverted in the vibrating portion.

In the present embodiment, the pitch period is calculated using the zero-crossing interval, so that if too many zero-crossing intervals are detected in the input digital signal waveform corresponding to one pitch period, the pitch period The calculation load will increase. However, in the present embodiment, since the binary signal is generated by the binarizing unit 8 having the masking band as described above, in the input digital signal, a slight movement near the zero level which is not important for the calculation of the pitch period is generated. A binary signal that is ignored and does not include more "0" / "1" inversion points than necessary is obtained, and it becomes possible to detect an appropriate number of zero-cross intervals for pitch period calculation.

As described above, a binary signal is generated based on each of the digital voice signal and the digital model signal. Then, for each binary signal, the time interval at which "1" / "0" inversion occurs is sequentially counted by the timer 9, and the zero-cross interval data as the timing result is latched 93 shown in FIG.
Are held in sequence. In this way, the zero-cross interval data sequentially held in the latch 93 is sequentially written in the RAM 10 under the control of the write controller 94. That is, the write control unit 94 uses the inversion of the binary signal to differentiate the differential circuit 90.
In response to the output of the pulse from, the write control routine whose flow is shown in FIG. 8 is executed. First, the write controller 94
Receives the zero-cross interval data t from the latch 93 (step S1) and determines whether or not the zero-cross interval data t is the lower limit value "8" or more. If the result of this determination is "NO"
In the case of, the routine is terminated without writing the zero-cross interval data t. The determination result of step S2 is "YE
In the case of "S", the process proceeds to step S3, and it is determined whether the zero-cross interval data t is larger than the upper limit "8192".
If the result of this determination is "NO", the zero-cross interval data t
Is written in the RAM 10 (step S4), and the routine ends. On the other hand, when the result of the determination in step S3 is "YES", "8
"192" is written in the RAM 10 (step S5), and the routine is ended. By the above control, "8" to "819"
Since only the zero-cross interval data within the range of "2" is written to the RAM 10, the zero-cross interval data that is not valid as the time interval of the zero-cross points of the audio signal is used for the calculation, and the incorrect pitch period is calculated. Can be prevented.

In this way, the zero-cross interval data stored in the RAM 10 is referred to by the pitch calculator 11 to find the pitch period of each of the digital voice signal and the digital model signal. Here, with reference to FIG. 9, an outline thereof will be described by taking a process of calculating a pitch period of a digital audio signal as an example. Assuming that the digital audio signal as illustrated in FIG. 9A is given to the binarization unit 8, the zero-cross interval data t ₁ , t ₂ , ...
It is accumulated within 0. The pitch calculator 11 makes the following four assumptions regarding the relationship between the zero-cross interval data t ₁ , t ₂ , ... And the pitch period of the digital audio signal, and examines the validity of each of them. To find the pitch period.

Assumption 1 The pitch period of the digital audio signal has a length T ₁ corresponding to the sum of _two zero-cross interval data t ₁ and t ₂ . That is, the times T ₁₁ , T ₁₂ , ... Shown in FIG. 9B1 are the pitch periods of the digital audio signal. Assumption 2 The pitch period of the digital audio signal has a length T ₂ corresponding to the sum of _four zero-cross interval data t _{1 to} t ₄ . That is, the times T ₂₁ , T ₂₂ , ... Shown in FIG. 9B2 are pitch periods of the digital audio signal. Assumption 3 The pitch period of the digital audio signal has a length T ₃ corresponding to the sum of the _six zero-cross interval data t _{1 to} t ₆ . That is, the times T ₃₁ , T ₃₂ , ... Shown in FIG. 9B3 are pitch periods of the digital audio signal. Assumption 4 The pitch period of the digital audio signal has a length T ₄ corresponding to the sum of _eight zero-cross interval data t _{1 to} t ₈ . That is, the times T ₄₁ , T ₄₂ , ... Shown in FIG. 9B4 are pitch periods of the digital audio signal.

The examination of the validity of each of the above assumptions and the calculation of the pitch period based on the examination result are executed according to the flow shown in FIG. First, the pitch calculator 11 calculates the recall ratio CR1 of the waveform of the digital audio signal under the assumption 1 above (step S101). This recall ratio is a numerical value indicating how much the respective digital audio signal waveforms corresponding to the respective pitch periods match when the above-mentioned assumptions are followed. In the present embodiment, the zero-cross interval data t ₁ , It is calculated based on t ₂ , ...

Now, with reference to the flowchart of FIG. 11, the procedure of the calculation for obtaining the recall ratio CR1 performed in step S101 will be described. First, step S
In step 201, "0" and "1" are set as initial values for the counter CNT and the control variable i.

Next, in step S202, the control variable i
Is incremented by “2” to set i = “3”. Next, in step S203, it is determined whether the condition of 0.9t ₁ −t _i <0 is satisfied, that is, whether the zero-cross interval data t ₃ is larger than 90% of the zero-cross interval data t _1. To do. If the result of this determination is "YES", the counter CNT is incremented by "1" (step S204),
The process proceeds to step S205, and if "NO", the step S205
The process proceeds to step S205 without passing through 204. Then proceeds to step _{S205, -1.1t 1 + t i <} 0 satisfies the condition or not made, i.e., the zero cross interval data t ₃
It is equal to or zero less than 110% of the cross interval data t _1. If the result of this determination is "YES", the counter CNT is incremented by "1" (step S
206), the process proceeds to step S207, and if “NO”, the process proceeds to step S207 without passing through step S206.

Next, in step S207, it is determined whether or not the control variable i has become "7". If the result of this determination is "NO", then the process returns to step S202. Thereafter, step S202~S207 over twice is performed, zero for each of the cross interval data t ₅ and t ₇ the determination in steps S203 and S205 are performed, the zero-cross interval data the zero cross interval data t Counter CN when greater than 90% of ₁ or less than 110%
T is incremented (steps S204, S
206).

When i = “7”, step S
The determination result of 207 is "YES" and step S20.
Going to step 8, the control variable i is set to "2".

Next, in step S209, the control variable i is increased by "2" to set i = "4". Next, in step S210, it is determined whether or not the condition of 0.9t ₂ −t _i <0 is satisfied, that is, whether or not the zero cross interval data t ₄ is larger than 90% of the zero cross interval data t _2. To do. If the result of this determination is "YES", the counter CNT is incremented by "1" (step S21).
1) The process proceeds to step S212, and if “NO”, the process proceeds to step S212 without passing through step S211.
Next, when proceeding to step S212, -1.1t ₂ + t _i <0
It is determined whether the following condition is satisfied, that is, whether the zero-cross interval data t ₄ is smaller than 110% of the zero-cross interval data t ₂ . If the result of this determination is "YES"
In the case of, the counter CNT is incremented by "1" (step S213), and the process proceeds to step S214. In the case of "NO", the process proceeds to step S21 without going through step S213.
Proceed to 4.

Next, in step S214, it is determined whether or not the control variable i has become "8". If the result of this determination is "NO", then the procedure returns to step S209. Thereafter, step S209~S214 over twice is performed, zero for each of the cross interval data t ₆ and t ₈ determines in step S210 and S212 are performed, the zero-cross interval data the zero cross interval data t Counter CN when greater than 90% of ₂ or less than 110%
T is incremented (steps S211, S
213).

When i = “8”, step S
The determination result of 214 becomes "YES" and step S21.
5, the value of the counter CNT is normalized by the number of judgments regarding the zero-cross interval data, and the result is defined as the recall ratio CR1. In the case of this flow, since the determination is performed 12 times, CNT / 12 is set as the recall ratio CR1.

Here, the assumption that the length of the pitch period is the sum T _{1 of the} two zero-cross interval data is correct, and the waveform of the digital audio signal does not change even if the pitch period is switched four times. At, t ₁ = t ₃ = t ₅ = t ₇ and t ₂ = t ₄ = t ₆ = t ₈ . Therefore, in this case, the recall rate CR1 obtained by the above processing is 100%.
Also, even if there is some error in each zero-cross interval data,
When t ₃ , t ₅ and t ₇ are within the range of t ₁ ± 10% and t ₄ , t ₆ and t ₈ are within the range of t ₂ ± 10% The recall rate CR1 is 100%. On the other hand, if the above assumption is incorrect, the pitch period is switched, which causes a large difference between mutually corresponding zero-cross interval data. Therefore, it is easy to make a negative judgment in step S203 and the like,
The recall rate CR1 decreases as the number of such negative determinations increases.

When the calculation of the reproduction ratio CR1 is completed in this way, the process returns to the flow of FIG. 10 and proceeds to step S102 to calculate the reproduction ratio CR2 of the waveform of the digital audio signal under the assumption 2. That is, it is assumed that the pitch period has a length T ₂ corresponding to the sum of four zero-cross interval data. Then, the first-th zero crossing interval data t ₁ ~t ₄ corresponding to the pitch period, respectively as a reference, the second, third and fourth zero corresponding to the pitch period of the cross interval data t ₅ ~ t _8, t ₉ each ~t ₁₂ and t ₁₃ ~t ₁₅ determines whether or not the match in the reference and the predetermined error range. Then, the number of times that a positive determination result is obtained is counted and normalized by the total number of determinations to obtain the recall rate CR2.

The assumption that the length of the pitch cycle is the sum of four zero-cross interval data is correct, and in the ideal state in which the waveform of the digital audio signal does not change even if the pitch cycle is switched four times, t ₁ = T ₅ = t ₉ = t ₁₃ t ₂ = t ₆ = t ₁₀ = t ₁₄ t ₃ = t ₇ = t ₁₁ = t ₁₅ t ₄ = t ₈ = t ₁₂ = t ₁₆ All the conditions are satisfied and the recall is CR2 is 100%. Even if there is some error in each zero-crossing interval data, if it is within ± 10%, recall ratio CR
2 is 100%. When the pitch cycle is switched, if zero-cross interval data that largely deviates from the reference (that is, the zero-cross interval data corresponding to the first pitch cycle) is generated, the recall ratio CR2 is set according to the number.
Will decrease.

Next, in step S103, the above assumption 3
Reproduction rate C of digital audio signal waveform
Calculate R3. That is, it is assumed that the pitch period has a length T ₃ corresponding to the sum of 6 zero-cross interval data. Then, the zero-cross interval data t _{1 to} t ₆ corresponding to the first pitch period are used as references, and the zero-cross interval data t ₇ to corresponding to the second, third, and fourth pitch periods are compared. t _{1 2,} t ₁₃ ~t ₁₈ and t ₁₉ ~t ₂₄
It is determined whether or not each of them matches the reference within a predetermined error range. Then, the number of times that a positive determination result is obtained is counted and normalized by the total number of determinations, and the recall ratio C
Find R3.

This recall rate CR3 is t ₁ = t ₇ = t ₁₃ = t ₁₉ t ₂ = t ₈ = t ₁₄ = t ₂₀ t ₃ = t ₉ = t ₁₅ = t ₂₁ t ₄ = t ₁₀ = t ₁₆ = T ₂₂ t ₅ = t ₁₁ = t ₁₇ = t ₂₃ t ₆ = t ₁₂ = t ₁₈ = t ₂₄ If all the conditions are met or each zero crossing interval data has some error, it is within ± 10%. If the error is within, the recall CR3 is 100%. Further, when the zero-cross interval data having a large error is generated, the recall ratio CR3 is lowered according to the number of the zero-cross interval data.

Next, in S104, the recall ratio CR3 of the waveform of the digital audio signal under the assumption 4 is calculated. That is, it is assumed that the pitch period has a length T ₄ corresponding to the sum of eight zero-cross interval data.
Then, with the zero-cross interval data t _{1 to} t ₈ corresponding to the first pitch cycle as a reference, the zero-cross interval data t ₉ to t ₉ corresponding to the second and third pitch cycles, respectively.
It is determined whether or not each of t ₁₆ and t _{17 to} t ₂₄ matches the reference within a predetermined error range. Then, the number of times that a positive determination result is obtained is counted and normalized by the total number of determinations to obtain the recall ratio CR4.

[0054] In the process up to the step S101~S103 is has been processed the pitch period of the 4 pieces of, T ₄₁ through T ₄₃ in the pitch period corresponding to three in step S104 (FIG. 9 (b4) ) Is the processing target. This is for the following reason. That is,
In step S104, it is assumed that the pitch period is a long time corresponding to eight zero-cross interval data. Therefore, if four pitch periods are to be processed in step S104, even if the assumption 4 is correct, the digital audio signal waveform will be stable for a very long time of four pitch periods. Otherwise, the recall rate CR4 will decrease. However, the waveform of the digital audio signal can maintain the same waveform for a certain short time, but the waveform changes after a certain amount of time. Therefore, when four pitch periods are to be processed, even if Assumption 4 is correct,
There is a high possibility that an unreasonably low recall ratio CR4 will be calculated due to the influence of the temporal change of the waveform of the digital audio signal. Therefore, in step S104, three pitch periods are processed as described above.

In step S104, the recall ratio CR4
Is t ₁ = t ₉ = t ₁₇ t ₂ = t ₁₀ = t ₁₈ t ₃ = t ₁₁ = t ₁₉ t ₄ = t ₁₂ = t ₂₀ t ₅ = t ₁₃ = t ₂₁ t ₆ = t ₁₄ = t ₂₂ Reproduced if all the conditions of t ₇ = t ₁₅ = t ₂₃ t ₈ = t ₁₆ = t ₂₄ are satisfied, or if there is some error in each zero-cross interval data, but within ± 10% of the error. The rate CR4 is 100%. Further, when the zero-cross interval data having a large error is generated, the recall rate CR4 is reduced according to the number of the zero-cross interval data.

Next, in step S105, assumptions 1 to 1 are made based on the recall rates CR1 to CR4 obtained as described above.
It is judged which of the four is appropriate. The detailed flow of this determination is shown in FIG. First, step S301
Then, it is determined which of the recall rates CR1 to CR4 is the largest. Then, when the recall ratio CR1 is the maximum, it is determined whether or not this CR1 is larger than a predetermined reference value ref (step S302), and the determination result is "Y.
In the case of "ES", the assumption 1 is followed, that is, the pitch period is obtained from the length T ₁ of the two zero-cross interval data. The same applies when the other recall rates CR2 to CR4 are maximum, and it is determined whether CR2 or the like is larger than a predetermined reference value ref (steps S303 to S30).
5) If the result of this judgment is "YES", the length T ₂ of four zero-cross interval data and the six zero-cross interval data according to the assumption on which each recall is calculated. The pitch period is calculated from the length T ₃ or the length T ₄ of eight zero-cross interval data. If the recall is the same, the priority is CR1>CR2> CR3.
> CR4 (CR1 is the highest priority).

On the other hand, if the maximum one of the recall rates CR1 to CR4 is less than or equal to the reference value ref, step S3.
The determination result is “NO” regardless of which of the steps from 02 to S305. In this case, it is not possible to conclude which of the assumptions 1 to 4 is valid, and the judgment result is not applicable.

When the above judgment is completed, the process returns to the flow shown in FIG. 10 and proceeds to the step corresponding to the judgment result. That is, when it is determined that the pitch period is to be obtained from the length T ₁ of the two zero-cross interval data, step S1
In step 06, four pitch periods each consisting of two zero-cross interval data are obtained (T _{11 to} T _{14 in} FIG. 9B1).
Corresponding to the pitch period of the digital audio signal. If it is determined that the pitch cycle is to be calculated from the length T ₂ of the four zero-cross interval data, the process proceeds to step S107, and the pitch cycle is calculated for four cycles according to the result of this judgment (see FIG. 9 (b2)). (Corresponding to T _{21 to} T ₂₄ ), and the average value of them is used as the pitch period of the digital audio signal. In addition, the length T ₃ of the zero-cross interval data for 6 pieces
Proceeds to step S108 when it is determined that obtaining the pitch period, the pitch period four periods determined according to the determination result (equivalent to T ₃₁ through T ₃₄ in FIG. 9 (b3)), the
Let these average values be the pitch period of the digital audio signal. When it is determined that the pitch cycle is to be calculated from the length T ₄ of the eight zero-cross interval data, the process proceeds to step S109, and the pitch cycle is calculated for three cycles according to the judgment result (see FIG. 9 (b4)). (Equivalent to T _{41 to} T ₄₃ ), and the average value of them is used as the pitch period of the digital audio signal.

When the above processing is completed, step S10
Return to 1 and repeat the same process. In this way, the pitch period of the digital audio signal is continuously output. On the other hand, in the determination of FIG. 12, when the conclusion “not applicable” is obtained, the pitch period is not calculated, and a signal indicating that the pitch period is not calculated is output.
It returns to step S101. In the above description, the calculation process of the pitch period is described by taking the case of the digital audio signal as an example, but the pitch period is calculated by the same process for the digital model signal.

As described above, in this embodiment, assumptions 1 to 4 are assumed.
For all of the above, recall is selected, the assumption that yields the highest recall is selected, and pitch calculation based on this selected assumption is performed only when the recall is within the allowable range. In some cases, it is a prudent procedure that is not done. The reasons for choosing such a careful procedure are as follows.

A. As a procedure other than the above procedure, for example, each recall rate corresponding to assumptions 1 to 4 is sequentially calculated, and when the recall rate within the allowable range is obtained, the calculation is ended, and the recall rate is obtained. An alternative method is conceivable in which is selected to obtain the pitch period. However, depending on the voice waveform, for example, the recall rates corresponding to Assumption 1 and 3 are within the allowable range, and the recall rate corresponding to Assumption 3 is higher than that of Assumption 1. If this alternative would be followed in such a case, assumption 1 would be selected and the wrong pitch period would be determined. It is conceivable to set the allowable range to be narrow so that the assumptions can be selected correctly, but in that case, there may be a number of cases in which it is determined that “not applicable”.

B. Further, an alternative method is also conceivable in which all recall rates corresponding to assumptions 1 to 4 are calculated, the assumption that yields the maximum recall rate is unconditionally adopted, and the pitch period is obtained. However, there are cases in which the recall corresponding to any of the assumptions is uniformly low, and the recall corresponding to a particular hypothesis may be slightly superior to the others. There is no guarantee that an accurate pitch cycle can be obtained even if the pitch cycle is forcibly calculated by adopting. For example, when the waveform of the digital audio signal changes abruptly in the pitch period, the recall is likely to be low under any of the above assumptions.

C. Therefore, in the present embodiment, the pitch cycle is calculated according to the above procedure to prevent the output of an inappropriate pitch cycle.

The pitch periods of the digital voice signal and the digital model signal obtained as described above are sequentially reported to the scoring unit 13, and the deviation of the pitch period of both signals and the level of both signals obtained by the level detection unit 12 are reported. The song of the singer is scored by the comprehensive evaluation of the deviation and the scored result is displayed on the display unit 14.

C. Evaluation Results of the Device According to the Present Embodiment With respect to the pitch period detection device described above, various operating conditions of each part were set, and the pitch period detection time and the detection error were evaluated. 13 to 16 show the results. First, FIG. 13 shows a result of measuring a pitch cycle detection error in a practical range by using a circuit that performs linear interpolation as the quadruple oversampling unit 7, variously changing the oversampling frequency of the circuit. From this result, it is understood that the detection error in the practical range can be sufficiently reduced by performing the interpolation of about 4 times oversampling. Next, FIG. 14 shows the process until the pitch period is detected for each of the case where the pitch period is obtained by correlation between three periods (m = 3) and the case where it is obtained by correlation between four periods (m = 4). It shows the result of measuring the delay time for each input frequency. As shown by the results of this experiment, if m = 3 or 4, the detection delay can be kept within a range without any problem. FIG. 15 shows the relationship between the number of times of averaging and the pitch cycle extraction error. FIG.
Shows the result of an experiment on how accurately the pitch period can be extracted by comparing with the waveform of the past period (pitch period). This experimental result shows that there are many errors when comparing the past two cycles or so, and when the input waveforms of the past five cycles or more are compared, the waveform is too old and the pitch period is wrong, and, in the It shows that the comparison of input waveforms over four cycles is the most suitable for accurate pitch extraction.

D. Modification (1) In the above-described embodiment, the pitch period is set to the sum of 2n zero-cross interval data depending on how much the zero-cross interval data forming one pitch period match between the pitch periods. It was judged whether the assumptions made were valid. Instead of this method, for each n, a predetermined number of pitch periods is calculated by calculating the sum of 2n zero-cross interval data, and n having the smallest variation in these pitch periods is selected, You may make it select. That is, in FIGS. 9 (b1) to 9 (b4), T
_When the variation of _{11 to} T ₁₄ is the smallest, the average value of T _{11 to} T ₁₄ is the pitch period, and when the variation of the T _{21 to} T ₂₄ is the smallest, the average value of T _{21 to} T ₂₄ is the pitch period, ... That is, the pitch period is calculated. Further, the determination method based on the zero-cross interval data disclosed in the above embodiment and the determination method for determining the variation of the pitch period are used together, and the zero-interval interval data and the variation between the pitch periods of the pitch period are comprehensively evaluated, The pitch cycle may be selected.

(2) In the above embodiment, the binarization unit 8
The width Δ of the masking band was fixed. However, since the amplitude of the minute vertical movement of the voice waveform that occurs near the zero level depends on the amplitude of the entire voice waveform, it may be difficult to determine an appropriate Δ. Therefore, the width Δ of the masking band of the binarization unit 8 is controlled by a method such as detecting the amplitude of the digital voice signal or the digital model signal, multiplying the amplitude value by a predetermined coefficient, and setting the result as Δ. It is preferable to carry out. (3) In the above embodiment, the pitch period is obtained by digital processing. However, the zero crossing interval may be obtained by directly binarizing the analog voice waveform, and the pitch period may be obtained based on the result.

[0068]

As described above, according to the present invention,
The voice waveform is converted into a binary signal by a binarizing means having a masking band near the zero level, and continuous zero crossing intervals of the voice waveform are obtained based on the binary signal, and the pitch period is 2n for various n. Assuming that the sum of the zero-cross interval data for each piece, the degree of coincidence of the speech waveform between each pitch period is obtained, and the pitch period is obtained by adopting the assumption of the highest degree of coincidence. Even if there is a minute vibration in the vicinity of the zero level, there is an effect that the pitch period can be accurately obtained at high speed.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration example of a 4 × oversampling unit in the same embodiment.

FIG. 3 is a block diagram illustrating a configuration of a binarizing unit in the same embodiment.

FIG. 4 is a block diagram illustrating a configuration of a binarization unit in the same embodiment.

FIG. 5 is a block diagram showing a timer, a RAM, and their control system in the same embodiment.

FIG. 6 is a diagram showing an operation of a 4 × oversampling unit in the embodiment.

FIG. 7 is a diagram showing an operation of a binarizing unit in the same embodiment.

FIG. 8 is a diagram showing an operation of a write control unit in the same embodiment.

FIG. 9 is a diagram illustrating an outline of pitch period calculation processing according to the first embodiment.

FIG. 10 is a flowchart showing a pitch cycle calculation process in the same embodiment.

FIG. 11 is a flowchart showing pitch period calculation processing in the same embodiment.

FIG. 12 is a flowchart showing a pitch period calculation process in the same embodiment.

FIG. 13 is a diagram showing a performance evaluation result of the same embodiment.

FIG. 14 is a diagram showing a performance evaluation result of the same embodiment.

FIG. 15 is a diagram showing a result of performance evaluation of the same embodiment.

FIG. 16 is a diagram showing a performance evaluation result of the same embodiment.

[Explanation of symbols]

1 ... CD, 2 ... vocal extractor, 3 ... microphone, 4 ... A / D converter, 5 ... DC remover, 6 ... L
PF, 7 ... 4 times oversampling unit, 8 ... Binarizing unit, 9 ... Timer (zero crossing interval measuring means), 10 ...
RAM (zero cross interval measuring means), 11 pitch calculating section (pitch calculating means), 12 level detecting section, 13
…… Scoring section, 14 …… Display section.

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Yusuke Yamamoto 10-1 Nakazawa-machi, Hamamatsu-shi, Shizuoka Yamaha Stock Company

Claims (2)

[Claims] 1. A binarizing means for binarizing a voice waveform with a zero level as a reference and outputting a binary signal; and a continuous zero crossing interval t ₁ , t ₂ , of the voice waveform based on the binary signal. A zero crossing interval measuring means for measuring ... And variously changing n (n is an integer of 1 or more), and for each n, a total sum T of 2n zero crossing intervals T = (t ₁ + t ₂ + ...
Assuming that t _2n ) is a pitch period, the zero crossing interval t ₁ ,
Based on t ₂ , ..., the degree of coincidence of the voice waveforms between adjacent pitch periods of m periods (m is an integer of 2 or more) is calculated, and n having the highest degree of coincidence of the voice waveforms is calculated. A pitch detecting device comprising a pitch calculating means for obtaining a pitch cycle by selecting, wherein the binarizing means uses a masking band within a predetermined range from the zero level, and the voice waveform is Pitch detection characterized by inverting the binary signal only when it changes across the masking band and maintaining the binary signal just before entering the masking band if the speech waveform is within the masking band. apparatus. 2. The pitch detecting apparatus according to claim 1, wherein the width of the masking band is controlled according to the amplitude of the voice waveform.