JPH059040B2 - - Google Patents

Description

[Detailed description of the invention]

[Industrial Field of Application] The present invention relates to a multipulse encoding device, and in particular, to a multipulse encoding device that generates audio waveforms synthesized independently for each analysis frame while synchronizing with the analysis frame period on the analysis side and maintaining continuity between analysis frames. The present invention relates to a multi-pulse encoding device that improves the multi-pulse train encoding efficiency in a multi-pulse encoding device that has means for linearly adding synthesized speech after each gain adjustment. [Prior Art] A multipulse encoding device is an encoding device that uses multipulses as sound source waveform information, and usually synthesizes audio waveforms using fixed-point finite precision calculations. In addition, in speech synthesis, the dynamic range of the speech synthesis filter, S/N (Signal/
In order to improve the noise ratio as much as possible, the gain is adjusted on the output side of the filter in accordance with the strength of the sound source. In such a multipulse encoding device, the multipulse sequence extracted on the analysis side is used as sound source information, and this is used as LPC such as α parameter and K parameter.
(Linear Prediction Coding)
It is sent to the synthesis side together with spectral envelope information based on coefficients to perform speech synthesis. These feature parameters sent from the analysis side to the synthesis side are transmitted in units of analysis frames or variable length frames, but in either case, speech synthesis is processed in units of analysis frames. Now, the time length of the above-mentioned analysis frame is set from the viewpoint of feature parameter analysis, transmission convenience, etc., and therefore, it is basically irrelevant to faithful transmission of feature parameters. For this reason, even with multipulse encoding devices that utilize multipulse trains and assume sound source waveform transmission, if the synthesis side continuously uses such uniform characteristic parameters for each analysis frame, the analysis frame boundaries will be lost. There is a problem in that the faithful continuity of the synthesized waveform that spans is basically ignored, and the waveform reproducibility inevitably deteriorates. In response to this problem, the following improvement measures have recently been introduced. In other words, it synchronizes with the analysis frame period on the analysis side, maintains continuity between analysis frames, and reproduces a synthesized waveform as a sum of impulse response waveforms by a multi-pulse train speech synthesis filter while taking into account the impulse response duration. This is a multi-pulse encoding device that basically eliminates the problem of discontinuity between analysis frames. The present invention is also directed to such a multipulse encoding device. [Problems to be Solved by the Invention] However, this type of conventional multipulse encoding device that eliminates the problem of discontinuity between analysis frames also has the following problems specific to the multipulse encoding device. remain. In other words, multipulse was originally searched for as a linear summation of impulse responses of a speech synthesis filter configured as an all-pole model that approximated the input speech waveform as much as possible, and was therefore a combination of multiple pulse trains that approximated the sound source waveform. is being searched as. There are various methods for this search, but in any case, it is composed of a collection of many pulses that can form multiple excitation points corresponding to the vocal fold vibration waveform, and therefore there is a limit to its encoding efficiency. There's a problem. The purpose of the present invention is to eliminate the above-mentioned drawbacks and to significantly reduce the number of multipulses by providing a means for searching for multipulses that have an excitation point that can represent a plurality of excitation points with one or at most a few excitation points. An object of the present invention is to provide a multi-pulse encoding device that greatly improves the encoding efficiency. [Means for Solving the Problems] The multipulse encoding device of the present invention provides an impulse response of a previous analysis frame that is synchronized with the analysis frame period on the analysis side and that enters the time domain of the next analysis frame in successive analysis frames. While maintaining continuity between analysis frames by resetting the audio signal to be added including the duration component, a maximum amplitude is given to the audio waveform independently synthesized for each analysis frame, and the gain is adjusted for each, followed by linear addition. In a multipulse encoding device having means for obtaining synthesized speech using a multipulse analysis, phase equalization is performed to determine filter coefficients of a phase equalization filter of an input speech signal for each analysis frame based on a multipulse sequence retrieved by performing multipulse analysis. filter coefficient determining means; and multi-pulse determining means for re-performing multi-pulse analysis for each analysis frame on the input audio signal subjected to phase equalization by the phase equalizing filter to determine an output multi-pulse. Equipped with. [Example] Next, the present invention will be described in detail with reference to the drawings.
FIG. 1 is a block diagram showing the structure of a first embodiment of a multipulse encoding device according to the present invention. The multi-pulse encoding device shown in FIG. 1 is composed of an analysis side 1 and a synthesis side 2. Furthermore, the analysis side 1 includes a window processor 101, a switch 102, a phase equalization film 103, a switch 104, and a coefficient calculator. 105, multipulse analyzer 106, LPC analyzer 107,
It is configured to include a multipulse encoder 108, an LPC coefficient encoder 109, and a multiplexer 110. In addition, the synthesis side 2 is a demultiplexer 2
01, LPC coefficient decoder 202, normalized multipulse decoder 203, maximum amplitude decoder 20
4, an LPC synthesis filter 205, a multiplier 206, and a waveform connector 207. The input audio signal input via the input line 1001 is processed by the window processor 101 into an LPF (Low Pass
After passing through a filter to cut out unnecessary high frequencies, it is sampled by an A/D converter at a predetermined sampling frequency of 8 kHz, and then quantized for each sample using a predetermined number of bits, 12 bits. This quantized audio signal is stored in the internal memory for a preset period of time, for example, 20 mSEC, or
Each 160 samples are cut out and stored using a preset window function, such as a rectangular function. The quantized audio signal stored in this way is also 20m
SECs are read out one after another and supplied to the multipulse analyzer 106 via the switches 102 and 104, which becomes the basic analysis frame length in multipulse analysis. Multipulse analyzer 106 extracts a multipulse train for each basic analysis frame using a known multipulse extraction technique. As a well-known multi-pulse train extraction technique, AbS (Analysis) is a so-called spectral domain evaluation technique.
-by Synthesis) processing, correlation function processing as so-called correlation region evaluation, etc., and in this embodiment, the latter method is used. This correlation region evaluation searches for a desired multi-pulse train by using the cross-correlation between the input audio signal and the impulse response of the analysis-evaluation synthesis filter on the analysis side and the autocorrelation of the impulse response. An analysis and evaluation synthesis filter used for this purpose is built into the LPC analyzer 107, and the multipulse analyzer 106 searches for multipulses while inputting impulse responses from the LPC analyzer 107. In this embodiment, this multi-pulse search is performed using a so-called isolated time that is preset by adding the impulse response duration of multi-pulses to the length of each analysis frame, excluding the first basic analysis frame. It is as follows. The LPC analyzer 107 analyzes the quantized audio signal for 30 m
A well-known LPC analysis is performed on this quantized audio signal while cutting it out using the SEC Hamming function and reading it out with a basic analysis frame length of 20 mSEC to obtain a predetermined order of LPC.
Extract the coefficients. These LPC coefficients are used as coefficients of the built-in synthesis filter to obtain the auditory weighted impulse response of the vocal tract filter. This impulse response is provided to a multipulse analyzer 106. The multipulse analyzer 106 first searches for the multipulse train of the foremost basic analysis frame through a cross-correlation calculation between the quantized audio signal and the impulse response for each basic analysis frame and an autocorrelation calculation of the impulse response. Now, the multi-pulse train of the basic analysis frame of the leading edge thus obtained is supplied to the LPC analyzer 107, and an impulse response is obtained by the built-in synthesis filter. This multi-pulse impulse response waveform is supplied to the multi-pulse analyzer 106 both in a portion spanning the basic analysis frame and in a portion connecting beyond the basic analysis frame. However, the impulse response connection time beyond the basic analysis frame is limited to a time range set in advance, taking into consideration the quality level of the synthesized speech. The multi-pulse analyzer 106 adds and synthesizes the quantized audio signal of the next basic analysis frame and the quantized data of the impulse response waveform of the previous basic analysis frame spanning the time domain of this basic analysis frame, taking into account the polarity, Quantized audio signals that maintain continuity between basic analysis frames are set one after another for successive basic analysis frames. In this way, the multipulse train is searched again while maintaining the continuity between the basic analysis frames. Although the multi-pulse train searched in this way maintains the continuity between the basic analysis frames, from the viewpoint of multi-pulse train coding efficiency, there is basically no difference from the conventional multi-pulse coding apparatus. Therefore, in this embodiment, the following phase equalization is applied to the output of the window processor 101 to solve this problem. The basic idea of this phase equalization is as follows. That is, there is a multi-pulse train that has just been retrieved, and when this is applied to a speech synthesis filter made up of an all-pole digital filter made up of LPC coefficients, the linear addition of its impulse response waveforms provides the desired sound source information. shall be taken as a thing. Such a digital filter can be easily constructed. In this case, it is assumed that a representative impulse is set based on these multi-pulse trains instead of the searched multi-pulse, and that this impulse response can also be obtained that substantially satisfies the sound source information. Such impulse settings can also be easily set. On the other hand, if there is a filter that converts the impulse response waveform obtained by a multi-pulse train into a representative impulse response waveform, then if the input audio signal is passed through this filter and then searched for a multi-pulse, the representative impulse is used instead of the multi-pulse train. It is clear that impulse is required. The purpose of phase equalization of the input audio signal is also to obtain such representative impulses with a significantly reduced number of pulses instead of multiple pulses by means of such a filter. Such a filter is a phase equalization filter. Note that a plurality of representative impulses are usually obtained for each basic analysis frame, corresponding to the pitch period and the like. The basic concept of phase equalization filters was introduced by Mr. Honda and Mr. Moriya, for example, in ``Speech Coding Using Phase Equalization Processing,'' Acoustical Society of Japan Speech Study Group Material S84-05, April 1984. has been done. Now, it is clear that the above-mentioned phase equalization filter can be easily realized by a transversal type digital filter, etc., and the phase equalization filter 10
3 is a digital filter for this purpose. The coefficient calculator 105 provides filter coefficients of the phase equalization filter 103, and the filter coefficients are determined as follows using the multipulse train supplied from the multipulse analyzer 106. Now, if the multi-pulse (or predicted residual waveform) is e(n) and the impulse response of the phase equalization filter is expressed as h(m) (m = 0, 1, 2,..., M), then the phase equalization residual is The difference waveform can be expressed as e _p (n) in the following equation (1). e _p (n)= _M 〓 ^m=0 h(m)e(n-m)...(1) Also, the pitch pulse train is expressed as e _M (n) in equation (2). e _M (n)= _∞ 〓 ^l=-∞ δ (n-n _l ) ...(2) δ in equation (2) is Kronetska's delta, and when n=n _l , that is, δ(o), the amplitude is 1. It becomes a pulse, and n _l −
n _l-1 is the pitch period. The coefficient calculator 105 inputs the multi-pulse e(n), calculates the autocorrelation coefficient of e(n), first calculates the pitch period by searching for the maximum value of this coefficient, and then calculates the pitch period with the same period as this pitch period. A pitch pulse train e _M (n) having a pitch pulse train e M (n) whose temporal position also almost coincides is selected from the multi-pulse train. The desired h(m) that best matches the above e _p (n) and e _M (n) is the phase equalized residual waveform e _p (n) shown by equation (1).
It is obtained by minimizing the squared error between and the pitch pulse train e _M (n) shown in equation (2). The squared error D is expressed by the following equation (3). D= _N-1 〓 ⁿ⁼⁰ (e _p (n)−e _M (n)) ² = _N-1 〓 ⁿ⁼⁰ ( _M 〓 ^m=0 h(m)e(n-m)− _{L- 1} 〓 ^l=0 δ(n-n _l )) ² ...(3) However, in equation (3), L is the number of pitch pulses included in the interval 0 to M-1. By finding the partial differential ∂D/∂h(k) based on equation (3) and _{setting it} to zero, the impulse response h( m) is found. ∂D／∂h(k)= _N-1 〓 ⁿ⁼⁰ {2( _M 〓 ^m=0 h(m)e(n-m)− _L-1 〓 ^l=0 δ(n-n _l ))・e(nk)} ...(4) Setting equation (4) to zero, the following equation (5) is obtained. _N-1 〓 ⁿ⁼⁰ _M 〓 ^m=0 h(m)e(n-m)e(n-k))h(m)= _L-1 〓 ^l=0 δ(n-n _l )e( n-k) ...(5) (k=0, 1, 2,..., M) From equation (5), h(m) can be found from a (k-1) element (k+1)-dimensional equation. This calculation formula can be expressed as a matrix equation as shown in equation (6) below.

〔Effect of the invention〕

As explained above, according to the present invention, audio waveforms synthesized independently for each analysis frame are linearly added after gain adjustment, while synchronizing with the analysis frame period on the analysis side and maintaining continuity between analysis frames. In the multi-pulse encoding device for obtaining synthesized speech using the multi-pulse encoding method, the multi-pulse encoding device is provided with means for determining the coefficients of a phase equalization filter from the searched multi-pulse train, equalizing the phase of the input speech for each analysis frame, and then performing multi-pulse analysis again. This has the effect of realizing a multi-pulse encoding device in which the number of multi-pulses is significantly reduced and multi-pulse encoding efficiency is greatly improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the structure of a first embodiment of the invention, and FIG. 2 is a block diagram showing the structure of a second embodiment of the invention. 1, 1'... Analysis side, 2, 2'... Synthesis side, 10
1...Window treatment device, 102...Switching device, 103...
Phase equalization filter, 104...Switcher, 105...
... Coefficient calculator, 106 ... Multipulse analyzer,
107...LPC analyzer, 108...multipulse encoder, 109...LPC coefficient encoder, 1
10... Multiplexer, 111... Equalization coefficient encoder, 201... Demultiplexer, 202...
... LPC coefficient decoder, 203 ... Normalization multipulse decoder, 204 ... Maximum amplitude decoder,
205... LPC synthesis filter, 206... Multiplier, 207... Waveform connector, 208... Equalization coefficient decoder, 209... Phase dispersion filter.

Claims (1)

[Claims] 1. Resetting of the audio signal in synchronization with the analysis frame period on the analysis side and including and adding the impulse response duration component of the previous analysis frame that enters the time domain of the next analysis frame in successive analysis frames. A multi-pulse encoding device having a means to give a maximum amplitude to a speech waveform independently synthesized for each analysis frame while maintaining continuity between analysis frames, adjust the gain for each, and then perform linear addition to obtain synthesized speech. a phase equalizing filter coefficient determining means for determining a filter coefficient of a phase equalizing filter of an input audio signal for each analysis frame based on a multi-pulse train retrieved by performing multi-pulse analysis; 1. A multipulse encoding device comprising: multipulse determining means for determining an output multipulse by re-performing multipulse analysis for each analysis frame on an input audio signal that has been subjected to phase equalization. 2. Means for dispersing the phase of the audio waveform independently synthesized for each analysis frame based on the filter coefficients of the phase equalization filter transmitted from the analysis side to the synthesis side so as to eliminate the effect of the phase equalization. A multi-pulse encoding device as claimed in claim 1, characterized in that: