TW454173B - Semi-automatic human voice dubbing method - Google Patents

Abstract

A semi-automatic human voice dubbing method in the field of human dubbing or timbre conversion field, simply converting original timbre into various different timbres. In human voice dubbing job of drama work (such as movie) usually many dubbers are required. Hence, the invention hope to go through the processing of timbre conversion with computer to achieve the function with one dubber capable of dubbing different timbres to save man power and fulfill the goal of personal dubbing studio. The reason why the human voice dubbing system in the invention is called ""semi-automatic"" is because the expression of happiness, anger, sadness and bliss still needs the voice signal of original persons who input. The achievement of the invention is to bring out a semi-automatic human voice dubbing method, using method of independent control base frequency, vocal track length, voice source signal and inner proportion of vocal track to attain goal of human voice timbre alternation. In addition, the invention also has implemented an on-line instant operation system to actually acquire quite abundant timbre variation after practical listening and measuring verification.

Description

PA880503.TWP-3/33

454173 V. Description of the invention ([Technical Field] The present invention relates to a method of semi-automatic human voice dubbing, and in particular to a method of simulating and controlling the characteristics of some vocal organs through a computer, thereby achieving the purpose of changing the tone color. The invention of the silk is a real-time semi-automatic vocal dubbing system. [Previous technology] Previous research on tone conversion, such as voice conversion, has focused on the tone conversion between a specific object and another-specific object. Relationship, that is, the tone of a certain person is converted into a tone of a certain B after processing. These methods must first record the common sentences spoken by the two people, and then obtain the two people according to these sentences. The corresponding relationship of tone colors can be converted into the speech of B according to the corresponding relationship. These conversion methods can be roughly divided into two categories. The first type is the parametric method. The effect of this method depends on the parameters. Whether it is possible to refine the features of the descriptor 15 4 H3]; the other is a non-parametric (n〇n_ parametric) method to various This program requires enough training sentences to obtain better results. [4,5]. In addition, Baudoin and Stylianou used four methods in the frequency domain. To compare the effects of speech conversion. 20 How to convert one tone into many different ones has not been found in the research results directly related to others, but the previous study of speech conversion has some reference value. Because Xi The known technology is still lacking, and cannot achieve the function of dubbing. The inventor of this case is eager to improve and innovate, and 1 mile-Fan painstaking research ► Install 丨 | (Please read the precautions on the back before E page) Order · ---- Line. Printed by the Consumer Cooperative of Intellectual Property Bureau of the Ministry of Economic Affairs 4δ4173

Description of the invention (Later, the successful development of this [invention purpose] scientific voice vocal dubbing method. PA880503.TWP-4/33 5 The purpose of the cat * month is to provide a semi-automatic voice vocal cube> storage method to come Change the sound source signal and the internal proportional voice of the sound channel. How to input the voice of adult male and male? How to input the voice of adult male and female. Stomach female, child or employee with different sounds. The second purpose is to borrow some of the characteristics of the sound generator's house / refined by the electric version to draw and control a ^ note to achieve the purpose of changing the tone. It has been made by Jiang Zhuang-a real-time online attention, ㈣p, The system 'and indeed through the actual listening test can get quite rich tone conversion. The application of human voice conversion, in many drama works (such as movies), ear dubbing work, usually requires a lot of dubbers, so it can be thrown off the computer to Zuo Li ^ We look forward to you ^ 3 color conversion processing to achieve the function of only -bit dubbing, month b with many different tone colors, so that the drama speaks and has The book's m-inch keynote for vocal dubbing. The labor cost of speaking and having a person can be saved, and the ideal of a dubbing studio can be realized. The vocal dubbing system we studied is called ^ semi-automatic. The expression of joy, anger, sorrow, and music must still be controlled by the original turn-in of the sound signal. Our system is only responsible for the 20 turns of the tone. A fully automatic vocal dubbing system is defined. No input is required. No. And the mood of my anger clothing music can be controlled by parameters. Such a text_t0_speech system is obviously based on the current technology. 'It is quite difficult to make a fully automatic vocal dubbing system. 10 15

un installed ~ ^-I ----. I 丁 ______,,-^ i ----: ----- line ---- (Please read the precautions on the back before ^ this education), Γ -, j. u I-«-I. H--^^ 454 1 7 3 Α7 Β7 PA880503.TWP-5/33 V. Description of the invention (10 [Technical content] This method of semi-automatic human voice dubbing has the above advantages. A method of tone conversion was proposed 'to independently control the fundamental frequency, the length of the channel (vGcaltrack), _off, and the goal of changing the vocal color of the channel compared with _Fangsilai', and this method has been implemented into an online The real-time operation of the human voice dubbing system has been verified by actual listening. It can indeed obtain quite rich tone color conversion. The overall architecture applied to the present invention includes: One day ^ " The basic week L uses the voice signal of the rookie (Frame) is cut as a unit, and then the position of the vertices of each base period in the frame is also determined. This part uses our newly proposed method. A tone and channel length adjustment, this part is to modify our previously proposed TIPW (Time Proportioned Interpolation of Pitch Waveform) syllable signal synthesis method [7] (Patent Bulletin No. 309588), so as to be able to meet the immediate requirements According to the set pitch and channel length parameters of the voice signal, please read the precautions before t t 15 to adjust the source signal printed by the Consumer Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs, this department. It is based on LPC (lin.ear prediction coding) analysis to obtain the sound source signal, and then adjusts the sound source signal according to the previous research results [8,9]. The internal channel adjustment of one channel is constructed by LPC analysis. Based on the vocal tract model 20, the length ratio of the front and rear portions of the vocal tract (pharyngeal cavity, oral cavity) is changed to simulate the difference in the internal vocal tract ratio of different people. ί [Schematic description] Please refer to the following detailed description of the embodiments of the present invention and its drawings, to further understand the technical content of the present invention and its purpose and efficacy; the relevant national standards (CNS ) A4 specification (210 × 297 mm) 4 54 1 7 3 A7 B7 V. Description of the invention PA880503.TWP-6/33 The drawings of the embodiment are: Figure 1 is a schematic diagram of the processing flow of the semi-automatic human voice dubbing method of the present invention. Figure 2 is a flowchart of base position determination for semi-automatic voice dubbing method. Figure 2 (A) is a schematic diagram of buffer setting for base period. Figure 2 (B) is a schematic diagram of buffer setting for base period. Figure 4 It is a schematic diagram of the distribution of the array elements X⑴ ~ χ [κ] of the waveform vertices; Figure 5 is a schematic diagram of the selection of the base cycle with errors; Figure 6 is a schematic diagram of the correction range of the base cycle; Figure 7 is a schematic diagram of the modified result of the base cycle; Sound source waveform diagram when T1 is increased; Figure 9 is a schematic diagram of the period setting method for obtaining the remaining signal of LPC analysis; Figure 10 is a flowchart for adjusting the sound source signal; Figure 11 is the correspondence between the grid filter and the channel; I5胄 Twelve is the comparison chart of the internal ratio adjustment of the channel / a / sound wear area; Figure 13 (a) is the internal adjustment of the channel ratio (original ~ tone 'spectrum analysis chart; ° 10 (free ^ one ^ read the back of the Pay attention to poverty, ^ this summer} equipment. -M Frequency spectrum (u == 1.4) of the Intellectual Property Bureau of the Ministry of Economic Affairs (printed by the Consumer Cooperative). 20 Figure 13 (c) is an analysis of the internal ratio adjustment of the channel (u == 〇6); · Table 1 is a table of the correct rate selected for the base position; Table 2 is the three vowels at 11¾: 1.4 and 0.6 Comparison table. Table 3 is a list of 10 parameter settings for tone conversion. The spectrum of the paper is again applicable to China's national standard (CNS) 4 54 1 7 3 A7. 5. Description of the invention (Table 4 is the clarity and Table of hearing results of naturalness; Table 5 is the result of discrimination evaluation. [Representative symbols of main parts] 1 Voice signal input 3 Tone and channel length adjustment 5 Channel internal ratio adjustment [Example] (Please read first Note on the back again ^^: This summer) 2 Real-time base cycle detection 4 Sound source signal adjustment 6 Voice signal output The processing flow is shown in Figure 1. Here, each processing block is briefly introduced, and the detailed evaluation method is explained in the following sections. 1. Basic frame 10-Voice signal input 1; One basic week detection 2 The input voice signal is cut in units of time frames (frames) and then the time is obtained in real time. The positions of the peak peaks (pitch peaks) of each base cycle; the Consumer Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs prints a tone and Channel length adjustment 3, this part is to modify the TIPW (Time Proportioned ... Interpolation of Pitch Waveform) syllable signal synthesis method [7] (Patent Bulletin No. 309588) proposed by the applicant of this case Under real-time requirements, adjust the voice signal according to the set pitch and channel length parameters;) A sound source signal adjustment 4 'This part is to obtain the sound source signal through LPC (linear prediction 20 coding) analysis 'Then adjust the sound source signal according to previous research results [8,9]; -7- This paper size applies Chinese National Standard (CNS) A4 specification (210 X 297 mm φ) 454173 Α7 Β7 V. Description of the invention (PA880503 .TWP-8/33 Internal channel ratio adjustment5, based on LPC analysis, changing the front and back parts of the channel (pharyngeal cavity, 0 to simulate the difference in the internal ratio of the channels of different people. The length ratio of the channel model of the cavity structure, a voice signal Output 6 Real-time base-cycle position detection 10 15. Printed by the consumer property cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs 20 base-cycle (length or position) The calculation of many base-cycle (length or position) is an important part in many voice signal processing applications. For tone conversion processing, the base period is an important parameter, and the accuracy of the position of the base period is important to the efficiency of the system. 1 The method proposed for base period can be roughly divided into two. The big class is of which the class is based on spectrum or co-relation analysis to obtain the base period length, for example, using the auto-correlation coefficient [ι] of the remaining signal of Up, or using the cross-correlation coefficient (l0ss C〇- re_n) [U] to find the base period length; In addition, a method is to find the base period directly on the waveform, such as _ wave peak to find the position of the base period [13, 14]. These two types of methods, generally speaking, the base cycle length obtained by the spectrum and correlation analysis method is higher than 4 degrees, but it takes more time, and the method on the time domain waveform is relatively less accurate. , But it requires less processing time. ^ 2.1 Although the method of detecting the position of the base period is “but some voice signal processing is required, only the length of the base period is required, and the position of the base period is not required.” It is not necessary to perform instant processing. Requirement 'Besides finding the base position and the accuracy of the finding:

'(Please read the precautions on the back page first.). _ Order and Thread 4 64 1 7 3 A7 B7 PA880503.TWP-9/33 5. Description of the invention () (Please read the precautions on the back before ^^ this (Page) At the same time, the need for real-time processing must be taken into account, so the direct detection method on the time domain waveform is used. Figure 2 is a flowchart of the base cycle detection method developed by the present invention. The functions of each block in the figure will be described below. 5 2.1.1 Buffer and time frame length The length of the buffer (buffer) is set to 500 * deduction / 11,025 sampling points, so it can contain sound signals above 22Hz, and the buffer method is as follows As shown in Fig. 3, the signal waveform in the buffer area in Fig. 3 (A) can be obtained after detecting the base position, and the base position shown in Fig. 3 (B) can be obtained. As shown in Figure 3 (B). In addition, the signal samples in one buffer must be divided into time frames for processing. The length of the time frame is set to half of the buffer, and 50% of the portion is the same as the next time frame. Alas, so a buffer can be divided into three time frames, this division is to increase the base detection The accuracy of the measurement _. 15 2.1.2 Energy and zero crossing rate Printed by the Intellectual Property Bureau of the Ministry of Economic Affairs, printed by the Consumer Cooperative, where energy is defined as the sum of the squared amplitudes of signal samples within a time frame, usually a periodic signal Its energy will be much larger than the energy of the mute or noise signal. However, the threshold of the energy of a time frame should be set to 20. How much is reasonable? According to experience, with 250 samples of 16 bits in a time frame, the energy threshold is It can be set to 256,000, so when the energy of a time frame is less than 256, POO, the periodic flag of the time frame is set to zero, otherwise it is set to 1. The zero crossing rate means that the signal passes in a unit time. The number of times on the time axis, the periodic signal usually has a zero crossover rate, and the noise signal is low. The paper size is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm). System 4 1 7 3 A7 B7 V. Description of the invention (PA880503.TWP-10/33 10 15 20 The invention uses a time frame as a unit to calculate the zero-crossing rate, with η · — (at a sampling rate of 11025 Hz, For a time frame with a time frame length of 25 sample points), the threshold is set to 60 '. So when the zero-crossing rate of a time frame is greater than 仞, it is considered that there is no periodic signal in the time frame, and the time frame The periodic flag of the frame is set to zero, otherwise it is set to 1. ， 75 Λ The result of the combined energy and the zero crossing rate. Only when the periodic flag of the frame is equal to zero in all three frames is the buffer considered as not It has a periodic signal. '2.1.3 Completion of the base position The following describes the processing steps for selecting the base position. Step (1): Set the periodic flag according to the energy and the zero crossing rate, and then determine whether the periodic flag is All are zero, if all are zero (that is, there is no periodic signal), it returns zero cycles directly. Step (2): Combine the 15 maximum amplitude values taken out from each of the three time frames, sort them in time order ’, and then store them in the array υ [_ι] ~ γμ. Step (3): Calculate the amplitude threshold and store it in the variable Clip. Let ’s set the procedure of Clip as: 'The three time frames of the buffer each take one amplitude maximum value, set it to maxl, max2, max3, and then judge the previous one. Whether the weighing area has a periodic signal. If not, let Qip = (maxl + max2 + max3) * 0.2. If there is a periodic signal, let clip = mm (maxl, max2, max3) * 0.6 ° Step (4) : Because Y [l] ~ Y [45] are not every point is at the peak position, but form a group of points near the peak and the peak, because -10- spring paper bearing scale applies Chinese national standard ( CNS) Α4 specification (210 × 2.97 mm) (please read the precautions on the back Γ) and write this page). Equipment_ • I 73 Α7 ------- Β7 PA880503.TWP-11/33 V. Invention Explanation () This sequence takes out those in Y [l] ~ Υ [45] whose amplitude is greater than Clip and are frontier, and stores them in the array X [l] ~ X [K]. Figure 4 shows an example of the distribution of the array elements X [l] ~ χ [κ]. Step (5): Set the threshold of the upper and lower cycle lengths: 5 If the previous buffer has a periodic signal, then set the lower cycle length threshold = the average value of the previous buffer's period multiplied by 0.75, which is (avejpitch) * 0.75 Upper cycle length threshold = Period average of the previous punching area multiplied by 10 1_75, which is 15 --- ^-; --- Ί ---: equipment --- (Please read the precautions on the back first Page) Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 20 (avejitch) * 1.75 (1) Otherwise, the cycle length threshold is set to 35 * samp 丨 ing_rate / 1 i, 〇25 Upper cycle length threshold = 200 * sampling_rate / 11,025 (2) That is, when the first periodic signal starts to appear, the present invention sets its i frequency to be in the range of 55Hz to 315Hz, and when the buffer area currently being analyzed is not the starting point of the continuous periodic signal, it follows Adjust the average cycle length so far. 'Step (6) Find all the positions of the periodic vertices in the buffer this time from the array X⑴ ~ x [κ]. The method is to use the starting point of the buffer as the reference point and set the distance in x [l] ~ χ [κ] Deng] between the upper and lower cycle length thresholds] Find out, take out a point with little maximum purity as the point of the cycle; and then use the boundary point as a reference point to find out 叩] ~ 乂 MIN 中The next two batches are between the upper and lower cycle lengths. The ones between the thresholds are used to pick out the vibration of the big X [ι] as the next cycle boundary point, so continue to find Shimoda

line

• 1 I- I 4 54 1 7 3

PA880503.TWP-12/33, invention description (printed by the Consumers' Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs, the weekly position, and the selection of the position of the second week or more: the correct position of the base week, but this # 去 、 # 瓦 约 了You can get a standard of 80/0, and this method also reports the accuracy of the correct base cycle position, or a longer cycle ^^ is smaller than the Clip value, which will cause the quality of the tone to change when the tone is changed later. : Two periods, points, so the method of correcting the selection result of the base period was developed. The lack of two period lengths will not change drastically. The point is that the average period length and In accordance with the position of the base cycle obtained by another. + Corrected by the amount of change in amplitude. Based on multiple experimental observations, the present invention defines two: T2 is similar 'if they meet the following conditions 4 length Tl and ABS (TrT2) < ΜAX (Τι, Τ2) χι1 / 5. Based on this condition, it can be detected whether or not (33) is shown in Figure 5. Figure 5 is a buffer after selecting the base position ^ = 5 similar Pi represents the amplitude at the Si point, and then, Perform steps 2 listed below to correct it. V Step ⑻: Remove Ding and two periodic long sounds from the periodic sequence, 20 and T2, T2 meet the formula (3), and meet the amplitude change amount The control of μλχ (Ρι, P2) / μχν (Ρι, P2) < 2: That is, keep Tι and Ding 2 without modification, and then save the average long-term sound 10 15 as avejpitch = (Tj + T2) / 2 -12- This paper size is applicable to China National Standard (CMS) A4 specification (210X 29 " 7mm Γ

II equipment — ^ ^ order; service _ * 、.:; (Like the precautions on the back ^ L ^ this page) __ K (5) 454173 A7 PA880503.TWP-13/33 B7 V. Description of the invention () Steps (b): Compare T2 and T3, because it does not meet the formula (3), so start to modify T3, that is, modify the position of S4. As shown in Figure 6, the possible position range of S4 is defined between a and b, and the positions of points a and b are given by formula (6) a = S3 + ave_pitch * 90%, and '5 b = S3 + ave_pitch * 110 % (6) to set, and then find a peak with the maximum amplitude value from the [a, b] interval, set the found maximum peak point to the new S4, and calculate the new T3 value, and the average period length is Change to ave_pitch = (ave_pitch + T3) / 2 ◦) 10 Step (c): Continue to check Ti and Ti + 1 and take the processing method of step ⑻ or (b) until the entire buffer is processed. After the processing is completed, the detection result of the base position shown in Fig. 7 is obtained. 2.2 Evaluation of base-period position detection 15. After adding the process of base-period position correction, an evaluation experiment is performed. The results are printed as shown in Table 1. The correct selection rate of the base position was increased to more than 95%. In addition, from Table 1, it can be seen that those that have a greater impact on the accuracy rate are those base cycles that have not been selected. According to observations, these base cycles that have not been selected are all concentrated when the sound is switched from the silent zone to the periodic one. 20 The previous one to two cycles, and the last one to two cycles that have entered the mute interval periodically, can be summarized because of the two exchange intervals

I The period length varies greatly, and the energy is less than the energy threshold we set. Therefore, it is judged as an aperiodic signal, and the base period selection error in Table 1 is mainly caused by the blasting sound, such as / 六 / and /勹 /. / / -13-This paper size applies to China National Standard (CNS) A4 (210X 297 mm) 414173 A7 B7 PA880503.TWP-14/33 V. Description of the invention Select the correct number of weeks. Select the wrong number of base weeks. Select the wrong number of base weeks. Total correct number of base weeks. Experiment 1 122 0 6 128 95.3% Experiment 2 283 3 11 297 95.3% Experiment 3 1279 8 60 1347 95.0% (Please read the precautions on the back before reading this page)

3. Tone and channel length adjustment J Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economy The tone and channel length adjustments here refer to the two parts of pitch adjustment and 5-degree channel length adjustment. Generally speaking, the impact The biggest factor in timbre is the pitch (fundamental frequency) and the resonance frequency (formant frequency) of the channel. The resonance frequency is closely related to the length of the channel. When the channel becomes shorter, the resonance frequency will increase, and vice versa It will decrease, that is, it shows an inverse relationship. Therefore, the purpose of controlling the length of the channel 10 can be changed by changing the level of the resonance frequency. In the past, a simple way to control the pitch and resonance frequency was to increase or decrease the playback speed of the recorder. This approach can indeed simultaneously increase and decrease the pitch and the resonance frequency in the opposite direction (low, However, the duration of the pronunciation will also decrease or increase, so this is not a good method. Therefore, the present invention adopts the TIPW syllable signal synthesis method proposed by the applicant of the previous 15 (Patent Bulletin No. 309588), and the TIPW can be independent Controls the duration, pitch and resonance frequency, so it can solve the aforementioned problems and meet the requirements for controlling pitch and channel length separately. The processing steps of the TIPW method are briefly described below in detail. For the situation, please refer to the original article [7]. -14-: This paper size applies the Chinese National Standard (CNS) A4 specification (210X297 mm) # 54173 A7 B7 PA880503.TWP-15/33 Printed by the Intellectual Property Bureau of the Ministry of Economic Affairs, Consumer Consumption Cooperative Fifth, the description of the invention () 3.1 Basic frequency adjustment First introduce the adjustment of the fundamental frequency (pitch). The physical meaning of the fundamental frequency refers to the frequency of the vocal cord vibration. The fundamental frequency of children is often high, followed by adult females, and the fundamental frequency of adult males is low. We can achieve the initial tone conversion by adjusting the fundamental frequency. The following is the processing steps of fundamental frequency adjustment: (STEP 1) Adjust the ratio of the base cycle and the length of the base cycle in the input signal to find the length of the base cycle that is currently being synthesized. (STEP 2) Find two corresponding 10s in the input signal based on the inter-temple position of the synthesis cycle. , And then calculate the weighting of the two original base period waveforms according to the linear time ratio. (STEP 3) Each of the two original base period waveforms is multiplied by its own weight. (STEP 4) Based on the length of the synthesis period and the length of the original period. Determine the length of the cosine window, then multiply the original base period waveform 15 by the cosine windows of the two halves, align them on the left and right borders of the composite base period, and superimpose them. (STEP5) Combine the two processed original bases. The cycle waveforms are added. 3.2 Channel Length Adjustment The adjustment of 20 channel length is achieved by adjusting the resonance frequency of the channel, and the adjustment of the resonance frequency is achieved through resampling. If you want to increase the resonance frequency ^ as a whole by 1.25 times, it is equivalent to setting a new sampling point (but the sampling rate is not changed) every 1.25 sampling points on the original signal waveform, so the new sampling The point may fall between the two old -15- (Please read the precautions on the back before ^ r ^ 'page)-Binding and binding This paper size is applicable to China National Standard (CNS) A4 size (210X297 mm) 454173 A7

^ --- * Μ, --- Please read the notes on the back before r'J page) 10 Find A, B, and C, and then bring x into / (x) with the value of xx〇 Obtain new sample values, and then continue to obtain new periodic waveforms. After the sound, replace the post-post waveform with the original base-period waveform used in pitch adjustment (STEP 3) to obtain the adjusted pitch and Newly synthesized periodic waveform for channel length adjustment. 15 4. Sound source signal 镅 cake 1 丨 _ Cancel 20 The sound source signal refers to the airflow signal generated by the vocal cord vibration. The frequency of the sound source signal (that is, the frequency of the vocal cord switch) is the fundamental frequency mentioned above. Regarding the adjustment of the fundamental frequency, the previous section has been discussed, and the adjustment of the sound source signal referred to in this section refers to adjusting the waveform of the sound source signal. As shown in Figure 28, the period of a sound source signal is expanded from the vocal cord to The biggest moment &,-16-This paper size applies to China National Standard (CNS) A4 (210 X 297 Gong Chuding 1 7 3

V. Description of the invention (5 10 IS :: Stretching and shortening such as 2 to maintain it-t = According to previous research results _, such changes can change: When Tl is increased, the sound can have a thin effect, and when τ2 is increased, it has a shimm_ effect. Considering the problem that can be implemented in practice, the sound source signal waveform is not obtained through the remaining signal of the integral knife, but is used instead. The approximation method is as described in 4 婿 i. The sound source signal obtained by doing an LPC analysis is 1 = Ding Shang-Yu Lili, ~ 7. When there is a break, it is not like the 所示 Γ / peak shown in Figure 8. Then, how does the program distinguish which peak is the one you want? 1 i I Note l: i Page printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs. Before explaining how the moment a of the sound source signal is determined, the present invention first Explain how the remaining signal is obtained, that is, first obtain a period of speech signal 'analyze it by LPC to establish an all-pole (an #) model of the channel 1 and then pass the analysis voice signal through the inverse all-pole model. The remaining red number can be obtained [11]. The Yu signal and the sound source signal we know that when the amplitude of the remaining signal reaches the maximum, it is usually the position of point a of the sound source signal in Figure Eight. Looking at the correspondence between the Yu signal and the original signal, we can find that When the original speech signal reaches the end of the base period (p touch shout), the remaining signal amplitude will also reach the maximum, so it is inferred that when the original speech signal is located at the end of the base period, it is usually the maximum amplitude bit f of the sound source signal 15-period speech signal-refers to b to

Order line 454173 A7 B7 V. Description of the invention (PA880503.TWP Signal between W ·. The definition of w C is b = (S1 + S2) / 2 and c = (S2 + S3) / 2 and 'This invention makes The time point a in Figure 8 is equal to S2, which is the base-period boundary point in the original voice message break. It turns out that the Lpc is used to analyze a period of the speech signal. 4.2 Sound source signal adjustment Fang Zhan ^ Inference from Section 4.1 We set the values of 1 ^ and D2 in Figure 8 according to the parameters a, b, and c in Figure 9. That is, let T \ = a-b, T2 (ίο) 10 Since point a is at the boundary point of the cycle, Therefore, Ding in Figure 8 is usually not equal to Ding 2 ', but in order to maintain the same length as the original cycle. Therefore, the adjustment of Ding 1 and Nai has a joint relationship. Therefore, if L < r2, let Γι, = Γι > < and, ----- j --------- install --- .... (Please read the phonetic on the back? Matters page) i. 15 τ2 '= τ-τ ',' Then let T ^ K2-R) τλ '= τ-τ2' (11). Line _ Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economy 20 where T = Τι + Τ2, and R is the adjustment ratio , The value range of R is 0 < R < 2 ι ', after Ding 2', then use the method of resampling in Section 3.2 to change the sample points of the rising and falling intervals of the sound source signal ', and the entire sound source signal adjustment process is shown in Figure 10. In terms of hearing, when the R signal is adjusted to a value greater than 10 (such as 1.5), the sound signal sounds like a sound made when the throat is dry (inflamed), and when When R is much smaller than 1 · 〇 (such as 0.5), the sound of the sound has a flickering feeling. -18 This paper size is applicable to China National Standard (CNS) A4 (210 X 297 mm) 454173

V. Description of the invention (A7 B7 PAOOOCOD.TWn g. Internal ratio of the sound channel. The human's sound channel is composed of the pharyngeal cavity, oral cavity, nasal cavity, etc. after the vocal cord. The color of 5 is not the same. Here we consider that even if two people with the same channel length 'their sounds still sound different, the reason for this phenomenon, we think that the internal ratio of the channel (The ratio of the length of the pharyngeal cavity and the oral cavity) is one of the factors that have a significant influence, so this section explains one way we achieve the internal adjustment of the vocal tract. 10 According to previous studies, we know what is obtained by LPC analysis. The reflection coefficient, which represents the meaning is the cross-sectional area ratio [15] of each adjacent section (sectiOn) inside the channel, and a lattice filter established by the reflection coefficient, as shown in Figure 11, It can be regarded as a model of the channel. Since the reflection coefficient obtained by Lpc analysis reaches ^^, it represents the cross-sectional area ratio of η adjacent sections between the vocal fold and 15 lips, so The present invention contemplates using this as the basis for adjusting the internal ratio of the audio channel, and the detailed situation is as follows. __Method for adjusting the internal ratio of the audio channel Calculate the relative cross-sectional area of each section of the channel based on the reflection coefficient, and then use interpolation to adjust the cross-sectional area of each section of the channel, and then reverse the changed cross-sectional area into a new set of reflections. The coefficient is used to establish a new channel model after the channel internal scale is adjusted. The detailed processing steps are as follows: (STEP a) Perform Lpc analysis on each voice signal period to obtain the reflection. -19- This paper scale applies to Zhongguanjiabiao Wei ( CNS) A4 specification (210 X 297 Gong Chu Γ 5 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 4 Π 73 Α7 Β7 PA880503.TWP-20/33 V. Description of the invention (10 coefficients and residual signals, set LPC analysis order L is 24; (STEPb) Bring Kjl-Kl into the formula (丨 2) to find the channel cross-sectional area Areai ~ AreaL;

Area Y + l 1-. .U. X Are%, make Area 100 (12) (STEPc) adjust the proportion u according to the required channel internally, interpolate ~ jreay into jrea

Area, ~ Arear is interpolated into Area: ~ Area [-+ 1 +1 (STEP d) Areai, ~ Area !; bring into formula (13) to find the adjusted reflection coefficient Kf ~ Kι;

Area]-Area'M Area, + Areal (13) (Please read the notes on the back first to write this page)

M (STEP e) replaces the reflection coefficient K in FIG. 11 with K ′, and substitutes the remaining signal obtained by (STEP a) into the grid filter to obtain the speech signal after changing the internal ratio of the channel. 15 5.2 Verification and comparison 5.2.1 Verification on cross-sectional area A verification method for the previously mentioned internal adjustment method of the channel is to 'take the LPC of the original speech signal to analyze the cross-sectional area of the channel, and adjust the 20-channel cross-section of the sound. The cross-sectional area of the LPC analysis of the synthesized signal after the area is compared to determine -20- This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) 4 54 ί73ι V. Description of the invention () 7 Whether the method can achieve the effect of cost. For example, Figure Day:-: The second is to adjust the ratio of the / a / tone signal, = 0.6, and three of the two. : The cross-sectional area of the channel obtained by the LPC analysis of the Cheng and Chi signals. The method used in Yu Ting's method is to directly change the length ratio of the two front and back sections of the channel. In theory, it should only affect the length ratio of the channel's internal length. However, after analyzing the synthesized speech signal, I found that in addition to the channel, In addition to the change in the internal length ratio, the cross-sectional area will also be affected. It is speculated that the remaining signal obtained by LPC analysis is not just a clean sound source signal. It also contains the channel information. Here, when the remaining channel signals including the channel information 10 are adjusted, the cross-sectional area will be changed. Similar results were found for other phonemes such as / w and / u /. 5.2.2 Comparison on the spectrum 15 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 20 About the comparison on the spectrum, only the analysis of the / ay tone on the spectrum, and the rest of the / 1 / and / u / The spectrum also has similar results. Figure 13 is the spectrum obtained after LPC analysis. Figures 13 (a), (b), and 为 are the wall adjustment ratios u when they are set to 1.0, 1.4, and 0.6, respectively. The analysis results obtained. Observing Figure XIII, it can be found that when 11 = 1.4, the amplitude of F1 appears to increase, and the number of hits decreases, and at the frequency of standing, both ^ and us have not changed, so it can be related to this phenomenon. It is explained that the energy of the low frequency part increases; in addition, when u == 〇6, the amplitude of F1's amplitude decreases and the trend tends to increase. At the frequency position, F1 has a tendency to move toward F2, but they have not changed. Therefore, this phenomenon can be interpreted as an increase in the energy of the south frequency part. _ -21- This paper size applies to China National Standard (CNS) A4 (210 X 297 public love) 4ί4173 A7 B7 V. Description of the invention (.PA88050-i.TWP-? '33 10 more · 2 · 3 Comparison on the comparison of the tone '. The experiment is based on the three vowels of / a /, / i /, / u /. After the actual audition, the results are recorded as shown in Table 2, that is, the adjustment ratio u = 0.6 When the three vowels are uniformly duller (the pitch feels higher), and when the adjustment ratio u == 1.4 is set, the three vowels are also uniformly thicker (the pitch is lowered) This phenomenon can be explained by the spectral changes observed in Section 5.2.2; and, the internal channel ratio adjustment method here can show consistency for the tone change of various phonemes, without the tone change situation of the same. . N ___n n 1-: .- nin 1 nnn L 1 nn I nn (Please read the phonetic on the back? Matters to write on this page). Table—Two · & Strive 11 Hot 1 zi by iV CtHb 仏, 丨4 · ^ u = 0.6 u = 1.4 / a / is thicker than the original sound / 1 / is thicker than the original sound / 1 / The original sound is thick and thicker than the original sound. The systemic evaluation of the semi-automatic vocal dubbing method of the present invention is constructed and developed on a 15% operating system. Since the 4 is 4 and the computer is 1 No. Lil · 20 Α The next day, there is no question B. The rotation of the sound of the search bar, the property has invented two processes for recording and playback; the two-practice approach to software writing, Figure 1 〖P 忪 is similar to the other steps, which are based on the signal cycle as a single step to process independently and independently; so Κ6-2 / 333MHz. ⑽ 人 电 知 CPU 疋 AMd -22-This paper is suitable for national standards

4 5 4 1 73 A7 B7 PA880503.TWP-23/33 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 5. Description of the invention () 6.1 Tone test The present invention adopts the method of audiometry to evaluate the tone, please ask one in advance. The male and a female record one sentence each and then archive them, then take these two sentences to the system of the present invention for tone conversion processing, each sentence is set to 5 different tone conversion processing with different parameters, and the more representative ten sentences are selected. Synthesized sentences of different tones are given to the listeners. There are 18 people participating in the test, and the intelligibility, naturalness and discrimination of the ten synthesized speeches are evaluated respectively. The so-called intelligibility here is compared with the original recorded sentences. The same clearness and no noise will be scored out of ten points, and the points that are worse than the original sentences will be deducted. The naturalness is 10 fingers. Does the synthesized sentence have a boy pretending In the case of a female voice or a female voice pretending to be a male speaker, the score will be a full tenth if not, otherwise the points will be deducted according to the degree of pretending. The relevant settings of the system of the present invention during the test were that the sampling frequency of the voice signal was 44,100Hz, and the test sentence was "Please send this basket of rabbits away." The other parameters are set as listed in Table 3, including pitch elevation, Channel length, sound source adjustment, channel internal ratio and other parameters. For the ten sentences played, the listener gave two scores of clarity and naturalness one by one according to their feelings. As a result, the score data listed in Table 4 were obtained. Looking at the male voice in Table 4, we can find that the intelligibility and naturalness of the fifth and ninth sentences are relatively lower than 20. The settings in Table 3 are all to reduce the pitch and lengthen the total channel length, that is, Change the direction of the sound towards the bass. Compared with the seventh sentence, I is considered to be full, so the effect is poor. Conversely, the sixth and eighth sentences are the most natural and clear. Three settings, in order to increase the pitch, the total length of the channel is reduced, that is, the direction of the soprano is changed -23-Please read the precautions for the back price again / The binding line is too paper, and it also applies the Chinese National Standard (CNS) Α4 Specifications (210X297 mm) 4 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 5 4 1 73 A7 B7 PA880503.TWP-24/33 V. Description of the invention () Change, because it was originally a male voice, so the effect of this change not bad. Table 3 The 10 parameters for tone conversion. Set the tone% channel length% R% U% tone% channel length% R% U% first sentence 125 100 100 140 sixth sentence 200 60 120 100 second sentence 140 70 100 100 seventh sentence 80 110 100 170 third sentence 400 50 100 100 eighth sentence 140 70 100 80 fourth sentence 60 60 100 100 ninth sentence 50 130 80 80 fifth sentence 50 150 100 100 tenth sentence 70 90 100 30 5 According to the female voice in Table 4, the third and fourth sentences are considered to have lower intelligibility and naturalness, while the fifth, seventh, and nine sentences are the higher three. According to Table 3, you can also find that the original It is better to adjust the female voice to the male voice (fifth, seventh, and nine sentences), but to adjust the voice that is originally female to the soprano voice (third and fourth sentences) is less effective. 10 From the average point of view, the difference between the male voice and the female voice is not large in terms of naturalness, but the female voice is obviously higher in clarity. After analyzing this phenomenon, we found that the two original recordings (of which One sentence was read by a boy and the other was read by a girl.) When you put it out, you can hear that the original male voice is worse than the female voice, so you have to close these 15 results. However, on the whole, both in terms of clarity and naturalness are considered to be of a medium-to-upper level, because the adjusted voice signal will more or less introduce noise (such as the detection of the base period, which has not reached 100%). Correct) and affect its intelligibility. In addition, the listener also knew in advance that the test sentence was -24- This paper size applies the Chinese National Standard (CNS) A4 specification (210X297 mm) m- I — ^ — ^ 1--I! — ^ Ϋ ID ^^^^ 1 fll ^ i In I— nm ls nn In fm ml, V5 service--(Please read the precautions on the back-Ef > c page), ：: .. '173 4 5 4 A7 B7 V. Description of the invention () The score converted by the male voice (or female voice), with a lower alpha ratio. PA880503.TWP-25/33 Therefore, in the evaluation of naturalness, we will ___________ the intelligibility results of natural clarity and naturalness ---- ___ female voice clarity naturalness clarity naturalness first sentence. 9.1 7.71 Six sentences 7.28 6.87 — second sentence 9.02 7.61 seventh sentence 9.31 8.15 third sentence 7.17 6.67. Eighth sentence 9.24 7.33 ~ fourth sentence 7.5 6.65 _ ninth sentence 8.85 8.36 fifth sentence 8.99 8.54 tenth sentence 7.5 7.29 _ average 8.396 7.518 Male voice ------------ First sentence clarity—7: 42 "Naturalness -------- 7.87 ~ Sixth sentence clarity—8.4 ~ Naturalness 7.77 Second Sentence 7.52 7.66 Seventh sentence 7.71 7.41 __Second sentence 7.37 ^ 7.24 Eighth sentence 8.51 Γ 8.51 Fourth sentence 6.53 6.7 Remaining nine sentences 6.52 6.06 Five. Sentence 6.32 5.75 Tenth sentence 7.05 7 13 Average 7.335 7.21 * ^^ ^^ = * === 81 (Please read the note on the back page first) • Binding and ordering Printed by the Consumers' Cooperative of the Zhihui Property Bureau of the Ministry of Economic Affairs-In addition, the tone of the synthesized sentence is used to identify whether the speakers are the same Evaluation of human discrimination, our evaluation method is: randomly select 10 points from male (or female) points 10 test sentences are selected from the completed sentence, and the same sentence can be repeatedly selected, and the war sentence is given to the listener 10 according to the selected order. 'Let him distinguish these 10 sentences from several people. Finally, Divide the number of people answering by the listener P by the number of people Q with really different timbre as the score of discrimination. For example, if there are 4 repetitions in 10 randomly selected sentences, then Q = -25- _ · _ ---..... The paper size applies the Chinese National Standard (CNS) M specification y ^ 〇X2, 97 public holidays)-line

• -—- II--II 73 73 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs' 5 4 A7 ___P Xian Ding WP · Ka V. Invention Description (6), and the number of distinguished persons P is 4, 5, Four values, such as 3 and 7 (assuming there are 4 listeners). At this time, the discrimination is calculated as (4 +5+ 3+ 5) / (6 * 4) = 70.83%. Please note that the fourth listener answered There are 7 different tones, but Q = 6, so change the P of the fourth position to P1, that is, make 5 P 'two Q-, PQ). If P > Q (14) So, the higher the degree of discrimination, it means The more different the tones are, the results of our discrimination experiments are shown in Table 5. The values are higher than 85%, that is, the 10 converted tones will be considered to be spoken by 8.5 people on average. The different tones converted from girls 'voices have a higher discrimination of 10, however, we cannot yet make such a conclusion that the tone conversion of girls' tones will have higher discrimination because only one boy has been tested with The voice of a girl. • Table 5 Results of the discrimination evaluation The discrimination between male and female voices is 85.7% 92.9% 15 6.2 Immediate test The input method of this system is divided into two types: standard input and microphone input, and the sampling frequency is divided into 11,025Hz, 22,050Hz, and 44,100. There are three types including Hz, and four parameters can be adjusted at the same time. Therefore, the test 20 on the immediacy is performed when all four parameters are adjusted. The immediate definition is when continuous input of voice signals to the system for processing. The voice signal converted from the tone color output by La 0 Ba will not be intermittent. Consider losing -26- This paper method scale is applicable to Chinese National Standard (CNS) A4 specification (210X297mm) '' (Please read the precautions on the back before this page)-Install

, 1T line_ 54 1 73

V. Description of the Invention (2 :: Various combinations of sampling frequencies. The test result is that when the tone color is converted by ㈣, the instant color conversion is achieved under the three sampling rates described above. However, when the sampling rate is higher than In the case of helmets, +, and temples, and all parameters are adjusted by the microphone feed_Instantaneous wind input and the sampling rate is not adjusted in the old Ma Ma Lake's microphone [Features and:] A HZ can still achieve real-time Requirement :: Ming proposed a method of using the channel characteristics to convert a variety of timbres, / Ming method to make a real-time semi-automatic vocal dubbing = in the process, the present invention considers the number of vocal organs, Zeng_ 卜 Τ Vibration frequency, channel length, sound source signal adjustment, sound, ratio, etc., to change these characteristics to simulate the change of channel shape, hoping to get a variety of sounds. Because of the semi-automatic vocal dubbing method 'must be done in real time The processing of base period detection, and the accuracy of base position detection has a great impact on downstream tone conversion processing and voice quality, so it took a lot of d to invent it. Ji Zhouxian 丨 The accuracy rate of the “current” Ji Zhou measurement in the research of the problem can reach 95%. In addition, the speech synthesized by the produced system is used for the hearing test, and the results show that the method of the present invention can indeed convert a variety of The above detailed description is a specific description of one of the feasible embodiments of the present invention, but this embodiment is not intended to limit the scope of the patent of the present invention, and any equivalent implementation or change without departing from the technical spirit of the present invention , Should be included in the scope of patents in this case. In summary, this case is not only innovative in terms of technical ideas, but also can be compared to I____ -27- This paper size applies to the national cap standard (CNS04 specification (210 X 297 public love) ) -Ni · I n Γ ----- tri ————! Line- /! (Please read the notes on the back before filling out this page), 73 4 V. Description of the invention (A7 B7 P'AfiflUSO ^. TWl ·-The customary treatment method begins to promote the statutory progress of μ, +. Xi; = work: according to = fully in line with the new nature and the Bureau approved this invention patent application, in order to encourage the invention of the application, I implore you, please ≪ Please read the back first Note on this page) Order · Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs -28- This paper size applies to the Chinese national standard (CNS) A4 specification (210 X 297 mm) 454173 A7 B7 rMoououo. \ Vvr -^ y / jo V. Description of the invention () [References] [1] H. Kuwabara and Y. Sagisaka, " Acoustic Characteristics of Speaker Individuality: Control and Conversion ", Speech Communication, Vol. 16, pp. 165 -174, 1995. 5 [2] H. Mizuno and M. Abe, " Voice Conversion Algorithm Based on Piecewise Linear Conversion Rules of Formant Frequency and Spectrum Tilt ", Speech Communication, Vol. 16, pp. 153-164, 1995 .

[3] Y. Stylianou and O. Cappe, " A System for Voice Conversion Based on Probabilistic Classification and a Harmonic plus Noise 10 Model ", IEEE Int. Conf. On Acoustics, Speech, and Signal Processing, Vol. L, pp 281-284, 1998.

[4] M. Narendranath, H. A. Murthy, S. Rajendran, and B. Yegnanarayana, " Transformation of Formants for Voice Conversion Using Artificial Neural Networks ", Speech Communication, Vol. 16, 15 pp. 207-216, 1995.

[5] N. Iwahashi and Y. Sagisaka, " Speech Spectrum Conversion Based on Speaker Interpolation and Mulit-functional Representation with Weighting by Radial Basis Function Networks ", Speech Communication, Vol. 16, pp. 139-152, 1995. 20 [6] G. Baudoin and Y. Stylianou, " On the Transformation of the

Speech Spectrum for Voice Conversion ", Int. Conf. On Spoken Language Processing, Vol. 3, pp. 1405-1408, 1996.

[7] Y. Y. Gu and WL Shiu, " A Mandarin-syllable Signal Synthesis Method with Increased Flexibility in Duration, Tone and -29- This paper size applies to China National Standard (CNS) A4 (210 X 297 mm) ) Please read the precautions before printing the A7 PA < 5503l! L5.TWP-όΰ / οό 73 _ ^ _ B7 printed by the Intellectual Property Bureau's Consumer Cooperatives of the Ministry of Economy

Timbre Control ", Proceedings of the National Science Council, Republic of China, Part A: Physical Science and Engineering, Vol. 22, No. 3, pp. 385-395, 1998.

[8] D. G. Childers, " Glottal Source Modeling for Voice 5 Conversion ", Speech Communication, Vol. 16, pp. 127-138, 1995.

[9] P. H. Milenkovic, " Voice Source Model for Continuous Control of Pitch Period ", J. Acost. Soc. Am., Vol. 93, No. 2, pp. 1087-1096, 1993.

[10] L. R. Rabiner, et al., &Quot; A Comparative Performance Study of 10 Several Pitch Detection Algorithms ", IEEE trans. Acoust., Speech, and Signal Processing, pp. 399-418, Oct. 1976.

[11] J. D. Markel and A. H. Gray, Jr., Linear Prediction of Speech, New York: Springer-Verlag, 1976.

[12] Y. Medan, E. Yair, and D. Chazan, " Super Resolution Pitch 15 Determination of Speech Signals ", IEEE trans. Signal Processing, pp. 40-48, Jan. 1991.

[13] Gu Hongyan, Tan Baihua, "Instant Conversion System of Singing to Instrumental Sound", Proceedings of National Computer Conference (Zhongli), pp. 228-234, 1995. [14] JF Wang, et al. ,, " A Hierarchical Neural Network Model 20. Based on a C / V Segmentation Algorithm for Isolated Mandarin

Speech Recognition ", IEEE trans. Signal Processing, pp. 2141-2146, Sep. 1991.

[15] LR Rabiner and BH Juang, Fundamentals of Speech Recognition, Prentice Hall, 1993. -30- This paper size is applicable to China National Standard (CNS) A4 (210 X 297 mm) --------- ---- Equipment ----- Order ------ line e ft. * (Please read the notes on the back before w page)

Claims (1)

454173 Α8 Β8 C8 D8 Application scope of patent 1_ -Semi-automatic vocal dubbing method 'is a method of tone conversion, in order to change the fundamental frequency, channel length, sound source signal and internal ratio of the channel to achieve- The target of the tone conversion; the main processing sequence of this method includes:-base cycle ❹j, which cuts the input voice message—time division _) as a unit, and then instantly finds the position of each base cycle vertex 时 in the time frame ; The consumer cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs printed a tone and channel length adjustment. This part is to modify the TIPW (Time Proportioned Interpolation of Pitch 10 Waveform) syllable signal synthesis method previously proposed by the applicant of this case, so that Under the request, 'the voice signal is adjusted according to the set pitch and channel length parameters; a sound source signal adjustment' This part is to obtain the sound source signal through LPC (linear prediction coding) analysis and adjust the sound Source signal; 15 — Internal channel ratio adjustment, based on the channel model constructed by LPC analysis, changing the length-to-length ratio of the front and rear parts of the channel , The ratio of the difference of internal channels to simulate different people. 2. The semi-automatic vocal dubbing method described in item 1 of the scope of the patent application, wherein the steps of selecting the position of the apex of the base cycle are: 20 Step ⑴: Set the periodic flag according to the energy and the zero crossing rate, and then judge the periodic flag If the targets are all zero, if they are all zero (that is, there is no periodic t-signal), return zero zero periods directly; Step (2): Combine the 15 maximum amplitude values taken out from each of the three time frames. Sort and store them in the array Y [1] ~ Y [45]; -31 _ This paper size is applicable to China National Standard (CNS) A4 (210 X 297 mm) (Please read the precautions on the back first (More on this page) I or -'— τ I. • m XV, vs. II---1 ^ 1. I 1 Printed by the Consumer Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs 4 5 ^ A8 3 B8 C8 D8 -PA880503. TWP-32/33 VI. Patent Application Step (3): Calculate the amplitude threshold and store it in the variable Clip. We set the procedure of Clip as' the two time frames of the buffer each take a maximum amplitude value and set it to maxl , Max2, max3, then determine whether the previous buffer has a periodic signal, if not If yes, let Clip == 5 (maxl + max2 + max3) * 0.2. If there is a periodic signal, let Clip = min (maxl, max2, max3) * 0.6; Step (4): Since Y [l] ~ Y [45] is not that every point is at the position of the peak, but forms a group of points near the peak and the peak. Therefore, in Y [l] ~ Υ [45], the amplitude is greater than Clip and is frontal. ίο Valuers, store them in the array χ [ι] ~ χ [κ]; Step (5): Setting the threshold value of the length of the upper and lower cycles: If the previous buffer has a periodic signal, the next cycle is ordered Length threshold = Period average of the previous buffer multiplied by 0.75, which is 15 (ave_pitch) * 0.75 Upper cycle length threshold 2 Period average of the previous buffer multiplied by 1.75, which is (ave_pitch) ^ 1.75 Otherwise, set the cycle length threshold. = 35 * sampling_rate / 11,025 20 Upper cycle length gate weight = 200 * sampling_rate / 11,025 That is, when the first periodic signal starts to appear, set its fundamental frequency to be in the range of 55Hz to 315Hz, and when the buffer being analyzed is currently : When the area is not the starting point of a continuous period signal, the The length of the cycle is adjusted; -32- This paper size applies the Chinese National Standard (CNS) A4 specification (210X297 mm) Hi-........ •--!-—1---|! 1 1 — .1:-!; =----I-Yi aJ--I----i I h,-f V (Please read the note on the back first and then breathe. This page) _ ^^ A8 B8 C8 08_PA880503.TWP-33/33 6. Scope of patent application (please read the precautions on the back and then step (6): find all the cycles in the buffer this time from the array X [l] ~ X [K] The position of the vertex; the method is to use the starting point of the buffer as a reference point, find x [i] in the distance between X [l] ~ χ [Κ] between the upper and lower cycle length thresholds, and take a point with the largest amplitude from it Take it as the 5 boundary point of the cycle; then use this boundary point as a reference point to find the next batch of X [l] ~ X [K] X [i] between the upper and lower cycle length thresholds Then, pick X [i] with the largest amplitude as the boundary point of the next cycle, and continue to find it. Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs -33-This paper size applies to China National Standard (CNS) A4 (210X297 mm)