JPH0448239B2 - - Google Patents

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a pitch extraction processing method, in particular, a method for dividing a voiced section of speech into a plurality of frames, and calculating an autocorrelation coefficient etc. obtained for each of the divided frames. When calculating the pitch period using the above method, the pitch period is calculated in a form corresponding to the path where the cumulative value of the calculated autocorrelation coefficient is maximum while maintaining the continuity of the plurality of divided frames as a whole. This invention relates to a pitch extraction processing method configured to calculate easily and with high accuracy.

[Problems to be solved by conventional technology and invention]

Conventionally, when calculating the pitch period (period of the fundamental frequency) from the residual waveform of a voiced section of speech, the residual waveform is divided into frames at predetermined time intervals, and the self-phase calculated for each frame is calculated. The pitch period was calculated independently for each frame from the position where the largest value of the relationship coefficient was obtained. In this way, in this calculation method, the pitch period is calculated independently for each frame, so the pitch period is twice or three times the original pitch period.
There were cases where the pitch frequency was detected at a period such as double..., and there were cases where the pitch period was calculated in a form corresponding to the position where the value of the correlation coefficient became large for some reason. , there was a problem that it became difficult to calculate the pitch period with high accuracy.

[Means for solving problems]

In order to solve the above-mentioned problems, the present invention divides a voiced section of speech into a plurality of frames, and calculates the pitch period using the autocorrelation coefficient obtained for each of the divided frames. By calculating the pitch period in a form that corresponds to the path that maximizes the cumulative value of the calculated autocorrelation coefficient while maintaining continuity in the form of integrating multiple divided frames, a highly accurate and simple method can be used. The pitch period is calculated depending on the configuration. Therefore, the pitch extraction processing method of the present invention extracts the voiced section of speech and extracts the pitch period using the correlation regarding the audio signal of the extracted voiced section. a correlation extraction unit that divides the audio into a plurality of frames and extracts the temporal correlation of audio for each divided frame; and a cumulative value calculation unit that sequentially calculates the cumulative value of the correlation for all the divided frames on a path where the correlation is within a predetermined range between adjacent frames; and the cumulative value calculation unit The present invention is characterized by comprising a pitch extraction unit that determines a route with a maximum cumulative value calculated by , and sequentially extracts pitch periods corresponding to the maximum route for each frame.

〔Example〕

An embodiment of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a configuration diagram of one embodiment of the present invention, FIGS. 2 to 4 are operation explanatory diagrams explaining the operation of the configuration of one embodiment of the present invention shown in FIG. 1, and FIG. 5 is an illustration of another embodiment of the present invention. An explanatory diagram illustrating the configuration of the embodiment is shown.

In the figure, 1 is a spectrum analyzer (LPC), 2 is an inverse filter (INV), 3 is a low-pass filter (LPF), 4 is a residual correlation calculator (Φ), and 5 is a cumulative value calculation unit (g _ij ) , 6 is a pitch extraction unit (P _i ), 7 is a voiced/unvoiced determiner (VUV), 8 is a main control unit (CPU), 9 and 10 are buffer memories (BF),
11 represents a switch.

FIG. 1 shows a block diagram of an embodiment for calculating the right end indicated pitch period from the left end indicated audio signal.

A spectrum analyzer (LPC) 1 in the figure is for extracting spectral information from the illustrated audio signal by a known method, for example for extracting spectral parameters by a linear prediction method. The calculated spectrum information is passed through an inverse filter (INV)
2, the inverse filter 2 generates a residual waveform signal from the spectral parameters and the sampled audio signal. The generated residual waveform signal is supplied to a residual correlation calculator (Φ) 4 after removing high-frequency components by a low-pass filter (LPF) 3. The residual correlation calculator 4 calculates an autocorrelation coefficient for each predetermined frame period from the supplied residual waveform signal, and notifies the voiced/unvoiced determiner (VUV) 7 of the calculated autocorrelation coefficient. The voiced/unvoiced determiner 7 that received the notification of the calculated autocorrelation coefficient is the spectrum analyzer 1.
Based on the spectral parameters notified from the spectral parameters and the maximum value of the residual autocorrelation coefficient notified from the residual correlation calculator 4, it is determined whether the current frame is voiced or not.
It is determined whether there is no voice or not, and the main control unit (CPU) 8 is notified. When the main control unit 8 receives a notification that the discrimination result is voiced, it turns on a switch 11 provided between the residual correlation calculator 4 and the buffer memory 9, so that the residual correlation calculator 4 The output residual autocorrelation coefficient values are sequentially stored in buffer memory (BF).
9 to memorize it. On the other hand, when the main control unit 8 is notified that the discrimination result is silent, it turns off the switch 11 and stores the value of the residual autocorrelation coefficient output from the residual correlation calculator 4 in the buffer memory. 9 will not be stored.

Furthermore, the main control unit 8 issues a command to the cumulative value calculation unit (g _ij ) 5 to calculate equations (2), (4), (6), etc., which will be described later, to calculate the so-called cumulative value “g _ij ”. . The calculated cumulative value “g _ij ” is sequentially stored in buffer memory (BF).
10. The main control section 8 then controls the pitch extraction section (P _i ) 6 to calculate the pitch period by calculating equation (5), etc., which will be described later, based on the contents stored in the buffer memory 10. . By adopting the above configuration, the pitch period for each frame is comprehensively calculated based on the residual autocorrelation coefficient calculated for each frame divided into multiple frames, so the pitch period can be calculated with extremely high accuracy. It can be calculated. The configuration and operation of this embodiment will be further explained in detail below with reference to FIGS. 2 to 4.

2A shows the residual waveform signal generated by the inverse filter 2 shown in FIG. 1, and FIG. 2B shows each residual waveform signal generated by the residual correlation calculator 4 shown in FIG. Indicates the correlation coefficient.

The first to Mth frames in FIG. 2A are obtained by dividing the illustrated residual waveform signal corresponding to the voiced section determined by the voiced/unvoiced determiner 7 into a plurality of frames. Each frame is divided so as to include a plurality of pitch periods as shown in the figure, and is also divided into parts that overlap, if necessary.
When extracting the pitch period corresponding to the divided first frame to Mth frame, as will be described later, to execute a process of calculating the pitch period in the form of integrating the first frame to Mth frame. , it becomes possible to extract the pitch period with high accuracy.

For example, "φ _ij " in FIG. 2B indicates the residual autocorrelation coefficient specified by the value of the subscript. The second term "j" of the subscript indicates the value of the "time delay" shown at the left end of FIG. 2B, and the first term "i" indicates the i-th frame shown in FIG. In detail, the time delay “0…N ₁ …j
...N ₂ "indicates the time delay corresponding to an integer multiple of the sampling interval at which the residual waveform signal is sampled, and "1, 2...i...M" indicates the first frame in FIG. , second frame...i-th frame...corresponding to M-th frame, respectively. In the figure, "φ ₁₀ , φ ₂₀ ... φ _i0 ... φ _M0 " indicate zero-order residual autocorrelation coefficients. In the figure, the "pitch existence range" indicates the range in which the pitch period (period of the fundamental frequency) of the voice is expected to exist, and the pitch period described later is extracted within this range. FIG. 2B shows, for example, the residual autocorrelation coefficients φ _iN1 ...φ _ij ...φ _iN2 corresponding to the time delays N1 . _. . , j .

The residual autocorrelation coefficient shown in FIG. 2B is given by the following equation.

φ _ij =1/N _N-1-j 〓 ⁿ⁼⁰ e _(o)・e _(o+j) ...(1) Here, e _(o) represents the residual signal, and e _{(o+j )} indicates a residual signal with a time delay "j" for determining the residual autocorrelation coefficient, and N indicates the frame length. In this way, the residual autocorrelation coefficients shown in FIG. 2B can be obtained.

FIG. 3 shows said time delay "j". The time delay "j" is "1," which is an integer multiple of the sampling period of the audio that is the source of the illustrated residual signal e _(o) .
The pitch period, which is the repetition period within the frame, is calculated by finding “j” that has a value corresponding to one of “2, 3, 4, …” and that maximizes the formula (1) above. becomes possible.

FIG. 4 shows an operation explanatory diagram for explaining the operation for calculating the cumulative value "g _ij " and the pitch period "p _i " according to the present invention. An equation for calculating the pitch period "p _i " will be described below, and then a procedure for calculating the pitch period using this equation will be sequentially explained.

First, the pitch period "p _i " is calculated using the following formula.

The cumulative value "g _ij " is given by the following formula.

() When N ₁ <= j <= N ₂ g _ij = max [w _j・g _i-1 , _k ] + φ _ij ...(2) However, k=N ₁ , N ₁ +1,...N ₂ − 1, N ₂ g _pj =0 w _jk is the weight related to the continuity limit as described later.

() In the case of j＜N ₁ orj＞N ₂ , g _ij =−∞ ...(3) In addition, () V _i =K ...(4) is defined. Here, K is "k" which gives max [w _jk ·g _i-1 , _k ].

Therefore, the pitch period P _i corresponding to i=M, M-1, . . . 1 is given by the following formula.

P _i-1 = V _i ...(5) However, P _M is j that gives the maximum g _Mj .

Further, the weight w _jk mentioned above is given by the following formula. Note that n ₁ and n ₂ below are values that give a time range.

() If j−n ₁ ＜=k＜=j+n ₂ then w _jk =1 …(6) () If k＜j−n ₁ ork＞j+n ₂ then w _jk =−∞ …(7) Next , how the pitch period P _i is calculated using the above-mentioned formula will be explained with reference to FIG. 4.

First, by using equation (1) and using the residual signal waveform stored in the buffer memory 9 shown in FIG.
All the residual autocorrelation coefficients within the pitch existence range shown in the figure are sequentially obtained, and the buffer memory 10 shown in FIG.
Store in.

Second, each residual autocorrelation coefficient shown in FIG. 4 stored in the buffer memory 10 is sequentially read out, and the equation
The cumulative value "g _ij " is calculated using (2), (4), and (6). That is, () When i=1 (for the first frame) g _1j = max[w _ik・g ₀ , _k ]+φ _1j =φ _1j Here, g ₀ and _k in the formula are 0 (the formula (2) proviso). This indicates that the time delay that provides the maximum value of the residual autocorrelation coefficient in the first frame shown in FIG. 4 is assumed to be j.

() When i=2, g _2j = max [w _jk・g ₁ , _k ] + φ _2j = φ _1k + φ _2j where w _jk indicates the weight, and is the maximum residual value in the first frame in Fig. 4. Upward (direction where the value of j decreases) centered on the indicated φ _1j , which is the differential autocorrelation coefficient.
If the maximum residual autocorrelation coefficient in the second frame is located within the range n ₁ shown in the figure and n ₂ shown downward (in the direction in which the value of j increases), for example, φ _{2 (} shown in the figure) If _j-3) is the largest residual autocorrelation coefficient, it is multiplied by a weight of "1" as shown in equation (6) to obtain the above-mentioned value.

On the other hand, if it is outside the range, for example, if φ _2(j+3) shown in the diagram is the largest residual autofunction, then the formula
Multiply the weight “−∞” as shown in (7) to obtain the cumulative value
g _2j =−∞, which is excluded from the calculation of the pitch period, and the one shown in the figure, for example, which has the largest residual autocorrelation coefficient within the above range is selected.
In this way, by excluding the range shown by equation (7) around the predetermined residual autocorrelation coefficient obtained in the previous frame, the pitch period can be calculated with high precision. is configured to extract.

Similarly, the cumulative values g 3j , g _4j shown in equation (2) are calculated using the weights given by equation (6), _which similarly yield significant values.
g _ij ...g _Mj are calculated in sequence and stored in the buffer memory 10 in FIG. Then, as shown in equation (5), the pitch period P _i for each frame is calculated. For example, from Fig. 4, P ₁ = j, P ₂ = j-3...P _i =
Calculate in the form of j-2... As explained above, the residual waveform signal is divided into multiple frames using a method similar to the DP method (dynamic programming method), and then the residual self-correlation of the multiple divided frames is calculated. The system uses a configuration that extracts the pitch period using the residual autocorrelation coefficient in a manner that corresponds to the path that maximizes the cumulative value g _ij of the number, so it is possible to extract the pitch period with high accuracy. It becomes possible.

5A shows a long voiced section, FIG. A case is shown in which the pitch period is calculated by dividing a voiced section into a plurality of short L frames.

The so-called
When extracting the pitch period using the DP method (dynamic programming), the pitch period of each frame cannot be finally determined until after the continuous voiced section of the input voice ends. Therefore, since the delay in extracting the pitch period depends on the length of consecutive voiced sections, it is not possible to quickly extract the pitch period when performing real-time analysis, for example. Therefore, in this embodiment, a long voiced section is divided into a fixed number of frames (blocks), and the pitch period is extracted for each divided section, so that the pitch period can be detected within a predetermined time. There is. At this time, the M frame is
The cumulative value described above in a continuous form from the beginning of the frame.
It is configured to use g _ij to improve the accuracy of pitch extraction.

〔Effect of the invention〕

As explained above, according to the present invention, when a voiced section of speech is divided into a plurality of frames and a pitch period is calculated using the autocorrelation coefficient obtained for each of the divided frames, The pitch period of each frame is calculated by determining the cumulative value of the autocorrelation coefficients calculated above while maintaining the continuity of the plurality of frames as one, and corresponding to the path where the cumulative value is the maximum. Therefore, the pitch period can be extracted with high accuracy using a simple configuration.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram of one embodiment of the present invention, FIGS. 2 to 4 are operation explanatory diagrams explaining the operation of the configuration of one embodiment of the present invention shown in FIG. 1, and FIG. 5 is an illustration of another embodiment of the present invention. An explanatory diagram illustrating the configuration of the embodiment is shown. In the figure, 1 is a spectrum analyzer (LPC), 2 is an inverse filter (INV), 3 is a low-pass filter (LPF), 4 is a residual correlation calculator (Φ), and 5 is a cumulative value calculation unit (g _ij ) , 6 is a pitch extraction unit (P _i ), 7 is a voiced/unvoiced determiner (VUV), 8 is a main control unit (CPU), 9 and 10 are buffer memories (BF),
11 represents a switch.

Claims (1)

[Claims] 1. A pitch extraction processing method in which a voiced section of speech is extracted and a pitch period is extracted using a correlation regarding the audio signal of the extracted voiced section, comprising: dividing the extracted voiced section into a plurality of frames. a correlation extraction unit that extracts the temporal correlation of the audio in each divided frame; and a correlation extraction unit that extracts the temporal correlation of the audio in each divided frame; a cumulative value calculation unit that sequentially calculates the cumulative value of the correlation for all the divided frames on a path that is within a predetermined range between adjacent frames; 1. A pitch extraction processing method, comprising: a pitch extraction unit that determines a route with a maximum accumulated value, and sequentially extracts pitch periods corresponding to the maximum route for each frame. 2 The extracted voiced section is divided into sections of a predetermined length, and the cumulative value is the value from the beginning of the voiced section, and pitch extraction processing is performed for each of the divided sections. Claim 1
Pitch extraction processing method described in section.