JPH0377998B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pitch extraction device for extracting pitch periods of speech, and more particularly to a pitch extraction device capable of real-time pitch extraction necessary for variable length frame type vocoders and the like.

It is known that the voiced part of a speech waveform has a periodic repeating waveform, and the change characteristic of the period (pitch period) is an important parameter in speech analysis, synthesis, recognition, etc. For example, in a speech analysis and synthesis system, the pitch extraction result extracted by the analysis section has a large effect on the quality of synthesized speech synthesized by the synthesis section.

As a method for extracting the pitch period of an audio waveform, there are conventionally known methods that use various analysis parameters, such as a method of calculating and extracting an autocorrelation coefficient for each frame having a time length approximately equal to the pitch period.

Pitch extraction methods based on autocorrelation coefficients are widely used because the autocorrelation coefficients can be obtained by processing in the time domain and the influence of the phase between the analyzed waveform and the frame is relatively small. However, the pitch extraction method based on the autocorrelation coefficient uses an integer multiple of the pitch period, or N1/N2 of the pitch period, as described later.
This method has the disadvantage that double the period is often mistakenly detected as the pitch period. (However, N1, N2
is an integer and N1<N2). The waveforms to be analyzed in which the above-mentioned defects occur can be classified into so-called voiced sound stationary parts and voiced sound transient parts such as the beginning of words, based on their waveform shapes.

One of the reasons why the above-mentioned defects occur in voiced sound stationary parts is that the waveform to be analyzed has extremely strong stationarity. This is because the so-called voiced stationary part is, for example, several hundred
If we observe a relatively long period of time such as mSEC, it is recognized that the pitch period and the waveform element each change gradually. However, if we observe various segments of the stationary voiced sound over a relatively short period of time, such as the time length of the waveform extracted every frame period (for example, 30 mSEC), we find that the waveform is almost completely stationary. , that is, it often shows periodicity. For example, the autocorrelation coefficient waveform of a sine wave becomes a cosine wave with the same period as the sine wave, and as is well known, the autocorrelation coefficient waveform of a waveform that has stationarity, that is, periodicity, has periodicity. has. Therefore, for each frame period, for example
When a waveform cut out over a time length of about 30 mSEC exhibits almost perfect stationarity, that is, periodicity, the autocorrelation coefficient waveform exhibits almost perfect periodicity. Therefore, for example, as shown in FIG. 1, the maximum value 101 of the autocorrelation coefficient in the pitch period is almost equal to the maximum value 102 in the double pitch period,
This is because the local maximum value 102 in the double pitch period often becomes larger than the local maximum value 101 in the pitch period due to calculation accuracy, slight disturbances, and the like.

Another reason why the above-mentioned defects occur in the voiced stationary part is that in the speaker of the analyzed waveform, for example, the bandwidth of the first formant is narrow, and the center frequency of the first formant is 2 times the pitch frequency (the reciprocal of the pitch period).
In the case of an integral multiple of the pitch frequency, for example, the second harmonic of the pitch frequency resonates with the first formant, and the frequency component twice the pitch frequency is extremely emphasized, making the fundamental frequency of the analyzed waveform appear as if it were the pitch frequency. This is due to the fact that the amount is doubled. When the apparent period of the analyzed waveform in which the frequency component twice the pitch frequency is extremely emphasized, that is, the reciprocal of the apparent fundamental frequency becomes 1/2 of the original pitch period, the self-correlation of the analyzed waveform The number waveform exhibits periodicity with a period slightly more than half the original pitch period. Therefore, as shown in FIG. 2, for example, the maximum value 201 of the autocorrelation coefficient that appears due to resonance between formant and pitch is almost equal to the maximum value 202 in the original pitch period, which causes false detection of the pitch period. .

The reason why the above-mentioned defects occur in the voiced sound transition part is that the voiced sound transition part has large changes in pitch period and voice waveform shape, and is relatively irregular. The autocorrelation coefficient waveform of the analyzed waveform, which has a large change in the pitch period and the shape of the audio waveform and is relatively irregular, often shows rough periodicity due to the pitch period, but The maximum values of the autocorrelation coefficients at periods that are integral multiples of the pitch period are relatively uneven, and the maximum value of the autocorrelation coefficient at the pitch period is often smaller than the maximum value at periods that are an integral multiple of the pitch period, so-called integral multiple pitch periods. Many errors occur.

Note that in the voiced sound transient part, the pitch frequency generally changes largely, and the effect on the speech waveform due to the resonance between the harmonics of the pitch frequency and the formant frequency is small compared to the effect in the voiced sound steady part, and The frequency of occurrence of so-called formant pitch errors is lower than the frequency of occurrence in voiced sound stationary parts.

The impact of pitch detection errors on the quality of synthesized speech in a speech analysis and synthesis system is auditory:
The error in the voiced sound stationary part is large, and the effect of the error in the voiced sound transient part is relatively small.

In the past, various methods have been attempted to reduce or eliminate pitch detection errors, particularly in voiced sound stationary parts, which have a large effect on the quality of synthesized speech. However, since conventional methods treat the steady part of voiced sounds and the transient part such as the beginning of a word, if a pitch detection error happens to occur at the beginning of a word, for example, the pitch detection error will affect future pitch detection characteristics. It has the disadvantage of having negative effects.

Conventional methods, for example, take advantage of the fact that the pitch period of audio changes relatively slowly and limit the difference in pitch period between adjacent frames to a predetermined range to extract the pitch period. There is a known method for preventing pitch cycle detection errors by performing the following steps. However, with this method of limiting the search range, for example, as shown in FIG. Once the so-called double pitch period is detected, it becomes difficult to detect the correct basic pitch period again.However, when transitioning from a silent part such as the beginning of a word to a voiced part, or from an unvoiced part. When transitioning to a voiced part, there is a high risk of erroneously detecting the double pitch period.

In order to alleviate the above-mentioned drawback, when determining the pitch search range from pitch cycles detected in the past few frames, it is difficult to determine the pitch search range at the beginning of a word.

An object of the present invention is to provide a pitch extraction device that performs pitch extraction based on autocorrelation coefficients, etc., which prevents errors in pitch period detection and more reliably detects correct pitches. be.

The pitch extraction device of the present invention uses a pitch determination interval of a certain length of time, uses the most temporally latest pitch period in the pitch determination interval adjacent in the temporal past as the only starting point, and further measures the continuity of the pitch periods. A means for evaluating by dynamic programming that has an inclination limit corresponding to the rate of change of the pitch period for the purpose of constraining the rate of change of the pitch period, and a plurality of pitch period sequences at the end of the pitch determination interval to determine the optimal pitch period sequence. It consists of means for making a unique decision from among candidates.

Next, embodiments of the present invention will be described with reference to the drawings. FIG. 4 is a block diagram showing an embodiment of the present invention, and the area surrounded by a dashed line 4013 shows the configuration range of the present invention.

A speech waveform to be analyzed is supplied to an A/D converter 4002 via a waveform input terminal 4001 . The A/D converter 4002 band-limits the audio wave to, for example, 3.4 KHz, samples it at 8 KHz, and linearly quantizes each sample with 12 bits. A/D converter 400
2 converts the quantized audio signal into a window processor 400
Output to 3. The window processor 4003 accumulates the quantized audio signal and outputs 30 mSEC (240 samples in this embodiment) at once to an autocorrelation coefficient calculator 4004 and a voiced/unvoiced discriminator 4005. Note that the output repetition period from the window processor 4003 matches the frame period in the pitch extraction process, and its value is, for example,
It is 10mSEC. The autocorrelation coefficient calculator 4004 calculates autocorrelation coefficient sequences τ _MIN , τMIN+ from the input 240 samples of audio signal by calculating the following equation (1).
₁ , τMIN+ ₂ ,..., _τMAX are calculated.

However, x ₅ (i) is the quantized audio sample, and j is the reference number of audio samples (120 in this embodiment).
The calculated autocorrelation coefficient example is the maximum value searcher 400
A local maximum value searcher 4006 supplied to a transmission line 4007 searches for the local maximum value of the coefficient sequence, and further outputs the local maximum value and the delay time corresponding to the local maximum value as a search result to a transmission line 4007.
is output to the DP processor 4008 via the DP processor 4008. The DP processor 4008 uses as its only starting point the most temporally newest pitch period in the temporally adjacent pitch determination interval supplied via a transmission path 4009 (described later), and uses the transmission path 4007 from the maximum value searcher 4006.
The optimum path of the maximum value and the delay time corresponding to the maximum value supplied via the pitch determination interval (in this example, 20 frames:
200mSEC). DP processor 400
8 further outputs the selected path (that is, 20 pitch cycle data in this embodiment) to the pitch output terminal 4010.
Output via. Further, the DP processor 4008 outputs the temporally newest pitch cycle data constituting the path, in other words, the pitch period data at the end of the path, to the pitch memory 4011 via the transmission line 4009. The pitch memory 4011 temporarily stores the pitch cycle data and returns it to the DP processor 4008 via the transmission path 4009 as pitch start point data in a pitch determination interval adjacent in the temporal future.

The above description describes the operation of the present invention when there is no unvoiced (including silence) frame within the pitch determination interval. When an unvoiced frame exists within the pitch determination interval, the operation of the present invention is as follows.

The voiced/unvoiced discriminator 4005 is the window processor 40
For the 30 mSEC worth of audio signals supplied from 03, voiced/unvoiced discrimination is performed using a linear discriminant or the like, for example, by a method using a vocal tract cross-sectional area ratio function as a discrimination parameter. The above-mentioned method is described in, for example, the published patent publication, 1984-151303 "Voiced/unvoiced discriminator"
This is the method described in . Voiced/unvoiced classifier 4
005 transmits the discrimination result to the transmission path 4 as a voiced/unvoiced signal.
012 to the DP processor 4008.
In the case of the voiced and unvoiced signal, a DP processor 4008
represents the absence of pitch by setting the pitch data of the corresponding frame to, for example, "0".
Further, when the voiced and unvoiced signal changes from voiced to unvoiced, the DP processor 4008 selects the optimal pitch period sequence from a plurality of pitch period sequence candidates in the final voiced frame. Further, when the voiced/unvoiced signal changes from unvoiced to voiced, the DP processor 4008 uses the first voiced frame as the starting point, and also uses each of the plurality of pitch cycle candidates at the starting point as starting points.

It is obvious that the present invention is not subject to any restrictions even if the silent section exists over a plurality of pitch determination sections through the above processing.

Further, the discrimination result of the voiced/unvoiced discriminator 4005 is not limited to the binary values of voiced and unvoiced, but can be easily output as a continuous quantity, making it possible to perform the following processing.
The temporal change characteristic of the pitch period has the following properties. That is, when the degree of voicing is high, for example, the energy of the voice is large, and the voiced sound stationary part where the rate of change of the onset is slow, the change characteristic is gradual. It is empirically known that when the degree of voicing is low, for example, the energy of the voice changes, and that the change characteristics are severe in voiced sound transition parts where the phoneme changes quickly. Further, as the discrimination parameter used in the voiced/unvoiced discriminator 4005, for example, the energy of the voice and the spectral envelope parameter corresponding to the phoneme are used. Therefore, the DP processor 40
In step 08, the setting value of the slope limit is adaptively changed depending on the degree of voicing of the voice, thereby enabling more stable pitch extraction.

Next, the operation of the DP processor 4008 will be explained in detail using waveform diagrams. Figure 5 shows DP processor 4008
FIG. 2 is a waveform diagram for explaining the operation of FIG. point 510
0 represents the temporally newest pitch period in the pitch determining interval temporally adjacent to the pitch determining interval 1. Points 5201, 5202, 5203
is the delay time corresponding to the maximum value of the autocorrelation coefficient in the first frame included in pitch determination section 1. Moreover, line segments 5101 and 5102 indicate slope restrictions. A dotted line 5211 connecting points 5100 and 5201 is outside the slope limit. Therefore point 520
Pitch column candidates 5001 and 5002 connected to 1 are excluded from the candidates. Similarly, points 5100 and 5203
A dotted line 5213 connecting the two points is also outside the slope limit.
Therefore, pitch periodic sequence candidate 500 connected to point 5203
5 is also excluded as a candidate. Point 5100 and point 5
202 is within the slope limit. Therefore, pitch periodic sequence candidate 5 connected to point 5202
003 and 5004 exist as candidates in the first frame. In pitch determination interval 1, pitch periodic sequence candidates 500 are used for all frames from the second frame to the 20th frame.
The points included in 3 and 5004 are within the slope limits from their respective previous frames. Therefore, the optimum pitch period sequence is determined in the 20th frame in conjunction with R _MAX determined by the following equation (2).

R _MAX = Max (ρ5003, ρ5004) ...(2) However, ρ5003 and ρ5004 are pitch period candidates 5, respectively.
This is an autocorrelation coefficient in pitch periods included in 003 and 5004. Assume now that 5003 has been determined as the optimal pitch period sequence.
Pitch periodic sequence candidates 5006,500 included in the 1st to 9th frames of pitch determination interval 2
7 and 5008, only candidate 5007 has a starting point within the slope limit based on point 5120, and 5007 is optimally determined as the pitch periodic sequence. Of course, if there is another candidate 5009 whose starting point is within the slope restriction based on the point 5120, the process is performed in the 9th frame and in the 20th frame of the pitch determination section 1 described above. Formula (2)
The pitch periodic sequence can be formed by .

Since the 10th to 12th frames are silent frame sections, there is no need to determine the pitch period. Pitch periodic sequence candidates 5010, 5011 and 5012 starting from the 13th frame are the
In the 20th frame, an equation equivalent to equation (2) is evaluated to determine the optimal pitch period sequence.

Note that a voiced frame near an unvoiced frame section has a low degree of voicing and is characterized by a sharp change in pitch period. Therefore, making the slope limit variable depending on the degree of voicing is effective in stably detecting the pitch period.

As explained above, the present invention evaluates and selects the optimum pitch period sequence using dynamic programming with a slope restriction corresponding to the rate of change of the pitch period, thereby creating a pitch period sequence having continuity of the pitch period. This makes it possible to easily determine the most likely pitch cycle sequence among the candidates, and also enables pitch extraction processing with a slight processing delay. That is, the present invention
The conventional method has the following pitch extraction error reduction effect for voiced sound steady parts and voiced sound transient parts where many pitch extraction errors occur. First, the voiced transient part exists in conjunction with the temporal past or temporal future of the voiced steady part, and the pitch period of the transient part is determined in relation to the pitch period of the steady part. If there is no pitch extraction error in the steady voice part, there will be almost no pitch extraction error in the voiced sound transition part. Next, in the voiced sound stationary part, the optimal pitch period sequence is determined from a series of pitch period sequence candidates through comprehensive evaluation using equation (2) above.
Double pitch period errors, etc. are removed. because,
Normally, when a maximum field search such as an autocorrelation coefficient is performed for each frame, the proportion of frames in which a trailing pitch period error occurs is about 0 to 30%, and each of the pitch period sequence and the trailing pitch period sequence This is because it is empirically known that when comparing the sum of autocorrelation coefficients of , the sum of pitch periodic sequences is almost certainly large. Further, the processing delay time of the present invention is, for example, at most
This is 200 mSEC, which is approximately equal to the processing delay time required for spectral envelope information analysis in a variable length frame type vocoder. Therefore, the present invention enables stable pitch extraction in real-time processing.

Note that in the present invention, the pitch extraction parameter is not necessarily limited to the autocorrelation coefficient. It is clear that the present invention can be easily implemented using cepstrum, waveform difference absolute values, and the like.

[Brief explanation of drawings]

Figure 1 is a waveform diagram for explaining the finger pitch cycle,
Figure 2 is a waveform diagram to explain pitch detection errors due to the influence of the first formant, Figure 3 is a waveform diagram to explain the drawbacks of the pitch extraction method that limits the pitch search range, and Figure 4 is a waveform diagram of the present invention. FIG. 5 is a block diagram for explaining an embodiment of the DP processor 400.
8 is a waveform chart for explaining the operation of No. 8. FIG. 4001...Waveform input terminal, 4002...A/
D converter, 4003...Window processor, 40
04...Autocorrelation coefficient calculator, 4005...Voiced/unvoiced discriminator, 4006...Local maximum value search device, 400
7...Transmission line, 4008...DP processor, 400
9...Transmission line, 4010...Pitch output terminal, 4
011... Pitch memory, 4012... Transmission line,
4013...Configuration range of the present invention.

Claims (1)

[Claims] 1. In a pitch extraction device for extracting the pitch period of speech, a pitch determination interval of a certain time length is used, and the pitch determination interval that is adjacent to the pitch period in the temporal past is the most temporally means for evaluating the continuity of the pitch period by a dynamic programming method that uses a new pitch period as the only starting point and has an inclination limit corresponding to the rate of change of the pitch period for the purpose of constraining the rate of change of the pitch period; 1. A pitch extraction device comprising means for determining a pitch periodic sequence uniquely from a plurality of pitch periodic sequence candidates at the end of a pitch determination interval. 2. The pitch extraction device according to claim 1, wherein the set value of the slope limit can be changed depending on the degree of voicing of the voice. 3. In the pitch extraction device according to claim 1, when there is an unvoiced section within the pitch determination section, a pitch period sequence is determined with the voiced section preceding the unvoiced section as the end, and , a pitch extraction device characterized in that a voiced portion following the unvoiced section is newly set as a starting point, and a plurality of pitch cycle candidates at the starting point are each set as starting points.