JPS6249640B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a linear predictive speech analysis and synthesis device using an autocorrelation method, and more particularly to a speech analysis and synthesis device for obtaining approximately periodic speech in finite precision calculations. In general, a speech analysis and synthesis device that uses prediction coefficients obtained as a direct solution of a linear equation, partial autocorrelation coefficients that are a modification of the prediction coefficients, or linear prediction coefficients that exchange them, evaluates the approximation of the spectral envelope of the speech to be analyzed. In order to improve the linear prediction coefficient, multiple linear prediction coefficients are usually determined. One of the methods for obtaining this linear prediction coefficient is the autocorrelation method. The autocorrelation method cyclically measures the normalized prediction residual power, which expresses the approximation of multiple linear prediction coefficients and spectrum parameters, from the autocorrelation coefficients.
In the above cyclic measurement method, for example, the N+1-order linear prediction coefficient includes N linear prediction coefficients at the N-order, autocorrelation coefficients from the 1st order to the N+1 order, and
It is obtained using the normalized predicted residual power. Regarding the autocorrelation method, J. Makhoul: “Linear
Prediction: A Tutorial Review”, Proc.
It is explained in detail in IEEE63, P.561 (1975). In linear predictive analysis, the autocorrelation method is theoretically guaranteed to be stable. (e.g. 1976 IEEE
International Conferenceon ASSP April12−14
1976IEEE Cat.NO.76CH1067−8ASSP P462~
465, NEW LATICE METHODS POR
LINEAR PREDICTION”Introduction).
Furthermore, in linear prediction analysis, if stability is guaranteed, the absolute value of the partial autocorrelation coefficient is within 1, and if the absolute value of the partial autocorrelation coefficient is within 1, stability is guaranteed. Regarding the relationship between stability and partial autocorrelation coefficient in linear predictive analysis, see the paper by Mr. Kitawaki et al., "PARCOR-type speech analysis and synthesis system," Nippon Telegraph and Telephone Public Corporation Telecommunications Research Institute Research Practical Application Report, Vol. 27, No. 6. (1978), pages 1061-1078, for details. As described in the above-mentioned "PARCOR-type speech analysis and synthesis system", in the autocorrelation method, the n-th order partial autocorrelation coefficient Kn is obtained as Kn=W _o-1 /U _o-1 . However, W _o-1 and U _o-1 are respectively It is required as. Here, αi ^(n-1) is the i-th linear prediction coefficient at order n-1, and V _o-1 is the autocorrelation coefficient at delay n-i. Also U _o-
1 is the normalized prediction residual power in the n-1 order. As mentioned above, theoretically |Kn|<1, U _o-1 <
1 (n≧2)U _o-1 =1 (n=1), |W _o-1 |<
1, and |W _o-1 |<U _o-1 〓(W _o-1 = U _o-
1・Kn). In general, the normalized predicted residual power is relatively small for highly periodic voiced sounds and nasal sounds.
It is relatively large for unvoiced sounds with low periodicity. V _oi in the voiced stationary part of a male speaker
(n-i=1, 2, ......8), Kn (n=1, 2
......8), U _o-1 (n = 1, 2, ......8), W _o
-1 (n=1, 2,...8) measurement value examples are shown in the table.

[Table] In the autocorrelation method, as mentioned above, W _o-1 is is required. The distribution range of the linear prediction coefficient αi ^(n-1) cannot be clearly determined, but from experience, for example, α ¹² is distributed between −4.5 and +2 (for example, BISHNU S. ATAL and
LAWRENCE R.RABINER“Apattern
Recognition Approach to Voiced－Unvoiced－
Silence Classification with Applications to
Speech Recognition”IEEE TRANSACTIONS
ON ACOUSTICS, SPEECH, AND SIGNAL
PROCESSING, VOL ASP−24, NO.3, JUNE
1976, pp01-212 (Fig. 5), in many linear prediction type phonetic analysis and synthesis devices, the distribution range of αi ^(n-1) is set to +8 to -8, or +16 to -16, for example.
W _o-1 is measured by, for example, 16bist fixed-point arithmetic, assuming that the Now, considering a 16-bit fixed-point operation in which the distribution range of αi ^(n-1) is set to +8 to -8, the weight of the lowest bit is approximately 2.44 × 10 ^-4 , and for example, to obtain W7, 7 times Since the sum of products is required, the measurement error of W7 is approximately 2.44×10 ^-4 ×2×
√7=1.29×10 ^-3 . According to the data in the table above, U7 is 1.21×10 ^-3
This indicates that the value of U7 and the measurement error of W7 may be approximately the same size. As described above, although the autocorrelation method in linear predictive analysis is theoretically stable, it is not necessarily stable in finite precision calculations such as 16-bit fixed-point calculations. In order to alleviate the above drawbacks, reducing the order of the linear prediction coefficient and reducing the number of stages of the synthesis filter will improve the calculation accuracy of the normalized prediction residual power even for voices with relatively low stationarity such as unvoiced sounds and voiced sounds. On the other hand, the approximation of the spectral envelope of loud voices is significantly reduced, and the synthesized sound quality is degraded. An object of the present invention is to provide a speech analysis and synthesis device that enables highly stable speech analysis and synthesis without deteriorating the quality of synthesized speech. The present invention relates to a linear prediction type speech analysis and synthesis device using an autocorrelation method, and includes means for cyclically obtaining a plurality of linear prediction coefficients while monitoring partial autocorrelation coefficients, and a means for cyclically obtaining a plurality of linear prediction coefficients while monitoring partial autocorrelation coefficients, and It is comprised of a means for setting the partial autocorrelation coefficients of the measured order to 0 when the absolute value exceeds 1, and setting the partial autocorrelation coefficient of the measured order to a constant value. The feature of the present invention relates to a linear predictive speech analysis and synthesis device using an autocorrelation method, in which a plurality of linear predictive coefficients are cyclically determined while monitoring partial autocorrelation coefficients that have been measured or are being measured. If the absolute value of the autocorrelation coefficient exceeds 1, the partial autocorrelation coefficients after the measured order are set to 0, and
The purpose is to set the partial autocorrelation coefficient of the measured order to a constant value, such as a statistical average value. Therefore, it is possible to stably obtain linear prediction coefficients in finite precision calculations and to stably generate synthesized speech with good approximation. Next, embodiments of the present invention will be described in detail with reference to the drawings. The figure shows one embodiment of the invention. In the figure, the dashed-dotted line 102 indicates the analysis side configuration, and the dashed-dotted line 10
4 shows the composition side configuration. The sampled and quantized audio waveform data is sent to the autocorrelation coefficient measuring device 1 via the waveform input terminal 101.
06 and the sound source extractor 115. An autocorrelation coefficient measuring device 106 measures an autocorrelation coefficient and outputs the autocorrelation coefficient to a linear prediction coefficient measuring device 109 via an autocorrelation coefficient transmission line 107. A linear prediction coefficient measuring instrument 109 measures a first-order linear prediction coefficient, a first-order partial autocorrelation coefficient, and a first-order normalized prediction residual power from the autocorrelation coefficient. The partial autocorrelation coefficients are provided to the controller 112 via partial autocorrelation coefficient transmission lines 1 and 110. The controller 112 has an absolute value of the partial autocorrelation coefficient of 1.
If the absolute value is 1 or more, the measurement stop signal is transmitted to the measurement stop signal transmission line 1.
11 to the linear prediction coefficient measuring device 109, and the memory control signal is supplied to the constant memory 114 via the memory control signal transmission line 113. The linear prediction coefficient measuring device 109 stops measurement when a measurement stop signal is given, and when no measurement stop signal is given, it measures the autocorrelation coefficient, the first-order linear prediction coefficient, and the normalized prediction residual power. A more quadratic linear prediction coefficient, a quadratic partial autocorrelation coefficient, and a quadratic partial autocorrelation coefficient
Measure the next normalized prediction residual power. Thereafter, the linear prediction coefficient measuring device 109 measures the linear prediction coefficient cyclically until the controller 112 generates a measurement stop signal. Note that it is easily possible to set the maximum prediction order N1 in advance and automatically stop the linear prediction coefficient measuring device 109 after measuring the N1-order linear prediction coefficient regardless of the presence or absence of a measurement stop signal. The constant memory 114 stores a predetermined constant when a memory control signal is applied, for example, approximately - if the analysis order is quadratic.
0.98, if it is 3rd order or higher, 0 etc. is the constant transmission line 108
It is output to the linear prediction coefficient measuring device 109 via. When the linear prediction coefficient measuring device 109 is given a measurement stop signal and is further given the predetermined constant, it replaces the partial autocorrelation coefficient in the current analysis order with the predetermined constant, and also replaces the partial autocorrelation coefficient in the current analysis order with the predetermined constant. Let the correlation coefficient be 0. Linear prediction coefficient measuring device 1
The partial autocorrelation coefficient measured at step 09 is output to the quantizer 118 via the autocorrelation coefficient transmission line 2,116. The sound source extractor 115 measures sound source information such as pitch period, voiced/unvoiced discrimination signal, and power from the audio waveform data, and outputs the sound source information to the quantizer 118 via the sound source information transmission line 117. Quantizer 1
18 quantizes and multiplexes the supplied partial autocorrelation coefficients and sound source information, and outputs them to a decoder 119 via a transmission line 103. The decoder 119 decodes the partial autocorrelation coefficient and sound source information from the quantized and multiplexed signal, and sends them to a synthesis filter 120 and an excitation sound source generator 12, respectively.
Output to 1. The excitation sound source generator 121 generates an excitation sound source consisting of a pseudo impulse train, random noise, etc. from the sound source information, and outputs it to the synthesis filter 120. This is a voice synthesis filter whose synthesis filter coefficients are determined by partial autocorrelation coefficients and whose input signal is the excitation sound source, which resynthesizes the voice and outputs the waveform output terminal 1.
Output to 05. Note that the synthesis filter 120 can also be configured using linear prediction coefficients instead of partial autocorrelation coefficients. In this case, for example, linear prediction coefficients may be obtained from the decoded partial autocorrelation coefficients on the synthesis side and supplied to the synthesis filter. In the present invention, the absolute value of the partial autocorrelation coefficient is 1
It is possible to determine whether or not the partial autocorrelation coefficient exceeds the partial autocorrelation coefficient without directly monitoring the partial autocorrelation coefficient. For example, it can be determined whether the absolute value of the partial autocorrelation coefficient exceeds 1 by determining whether W _j1 is greater than U _j-1 or by determining whether the value of U j is negative. It can be determined indirectly.

[Brief explanation of the drawing]

The figure is a block diagram for explaining an embodiment of the present invention. 101... Waveform input terminal, 102... Analysis side,
103... Transmission line, 104... Combining side, 105...
... Waveform output terminal, 106 ... Autocorrelation coefficient measuring device, 107 ... Autocorrelation coefficient transmission line, 108 ...
Constant transmission line, 109...Linear prediction coefficient measuring device, 1
10... Partial autocorrelation coefficient transmission line 1, 1, 111
...Measurement stop signal transmission line, 112...Controller, 1
13... Memory control signal transmission line, 114... Constant memory, 115... Sound source extractor, 116... Partial autocorrelation coefficient transmission line 2, 117... Sound source information transmission line, 118... Quantizer , 119...decoder,
120... Synthesis filter, 121... Excitation sound source generator.

Claims (1)

[Claims] 1. An autocorrelation coefficient measuring means for calculating an autocorrelation coefficient from an input audio signal, and a linear prediction coefficient measuring device for cyclically measuring a plurality of linear prediction coefficients and partial autocorrelation coefficients using the autocorrelation coefficient. , a controller that outputs a correction signal when the partial autocorrelation coefficient exceeds 1; and a controller that outputs a correction signal when the partial autocorrelation coefficient exceeds 1;
Replace the value of the partial autocorrelation coefficient exceeding 1 with a predetermined constant whose absolute value is smaller than 1 and corresponds to the order, and set subsequent higher order partial autocorrelation coefficients to zero and output. means on the analysis side, transmitting the partial autocorrelation coefficient and the sound source information of the input audio signal as transmission parameters to the synthesis side, and the synthesis side synthesizes the input speech based on the transmitted parameters. A speech analysis and synthesis device featuring: