JPH0784597A - Speech encoding device and speech decoding device - Google Patents

Abstract

PURPOSE:To provide the speech encoding device and speech decoding device which improve the quality of a reproduced sound at the time of a voiceless sound signal with small periodicity and a consonant with large variation while maintaining the quality of a voiced sound with large periodicity when the time- base compression and time-base expansion of a speech are performed. CONSTITUTION:The speech encoding device equipped with a pitch period detection part 700 which detects the pitch period of a speech input signal is equipped with a speech kind decision means 200 which decides which of a voiced sound, a voiceless sound, and a consonant the speech input signal is, a signal converting means 400 which converts the speech input signal to a positive or negative certain value according to whether the signal is positive or negative when the speech signal is the voiceless sound or consonant, and switch means 300 and 500 which apply the speech input signal as it is when the speech input signal is the voiced sound or the output of the signal converting means to the pitch detection part when the voiceless sound or consonant according to the output of the voice kind decision means.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder and a speech decoder, and more particularly to a time domain harmonic compression / expansion system (TDHS: Time Domain Harmonic).
The present invention relates to a voice encoding device and a voice decoding device for improving the encoding and decoding quality of unvoiced sounds and consonants when performing voice encoding and decoding by Scaling).

With the spread of digital lines in recent years,
In order to effectively utilize the line, various voice signal compression techniques have been proposed which efficiently compress the information amount while maintaining high quality of the voice signal. Among them, by utilizing the temporal periodicity (pitch period) of the voice signal, the transmitter side multiplies signals of several periods with similar characteristics by appropriate weighting, compresses it into a signal of one pitch period, and encodes it. TDHS is known which performs decompression opposite to compression while considering the front-back relationship of the compressed signal on the receiving side.

In the TDHS described above, unvoiced sound having a small periodicity is compressed on the time axis as well as voiced sound having a periodicity.

[0004]

2. Description of the Related Art In a speech coder that encodes and transmits a speech signal of a telephone or the like on the transmitting side and decodes it on the receiving side, conventionally, on the transmitting side, in a time base compression section (not shown),
The audio signal is time-axis compressed and then encoded by a coding unit (not shown). This output is multiplexed with the coded outputs of other channels and sent to the transmission line. Figure 8
FIG. 4 is a diagram for explaining an example of a time axis compression / expansion method, (a) shows a time axis compression processing on the transmission side, and (b) shows a time axis expansion processing method on the reception side.

In FIG. 1 (a), the pitch of the voice input signal is first detected. As a reference for observing the correlation of the signal at this time, the autocorrelation coefficient (equation 1) and covariance value (equation 2) shown below are used. Was used.

[0006]

[Equation 1]

[0007]

[Equation 2] However, S: input voice signal, F: evaluation formula. Next, two consecutive audio signal sequences (sections Pp ₁ and Pp ₂ ) divided by the extracted pitch period are compressed into one. That is, the signal of the previous cycle is multiplied by the windowing coefficient (1-Wc (n)), the signal of the subsequent cycle is multiplied by the windowing coefficient Wc (n), and both are added to obtain one signal Sc. The time axis is compressed as (n).

The autocorrelation count and covariance value, which are the criteria for pitch detection, are evaluated equally when the pitch is long and when it is short. However, as the pitch becomes long, intelligibility generally deteriorates, and therefore the optimum pitch may not always be the optimum pitch in terms of voice quality.

[0009] On the other hand, the same figure (b) showing the time axis expansion on the receiving side
In this case, pitch period information is sent separately from the compressed voice signal, and based on this pitch period information, a compressed signal for three pitch periods (section Pp ₁ ', Pp ₂ ', Pp ₃ 'in the figure) Among them, the signal for the preceding two pitch periods is multiplied by the windowing coefficient We (n), and the signal for the following two pitch periods is multiplied by the windowing coefficient (1-
We (n)) is added and the time axis is expanded to the original signal Se (n) for two pitch periods.

[0010]

In the TDHS described above, even for unvoiced sound with a small periodicity, or in the consonant part that is the beginning of a talk even if it is a voiced sound, it is possible to recognize a human voice within a short time at a relatively low level. Although it contains a lot of necessary information, it compresses and decompresses the same time axis as voiced sound with large periodicity.

As a result, a problem arises in that the correct encoding and decoding / reproducing process is not performed on unvoiced sound, and the decoded speech is deteriorated in the form of ambiguity of consonants. Further, even for voiced sound having a large periodicity, it is advantageous in terms of clarity of voice to compress it in a shorter cycle than in a large cycle.

Therefore, according to the present invention, the characteristics of the unvoiced sound and the consonant portion of the voice signal are used to compress the time base of the voice.
A voice encoding device and a voice decoding device that improve the quality of unvoiced sound signal with less periodicity and the reproduced sound quality at the time of consonant with large fluctuation while maintaining the quality of voiced sound with large periodicity when performing time axis expansion. The purpose is to provide.

[0013]

The above problems can be solved by the structure of the apparatus shown in FIGS. In FIG. 1 (claim 1), a pitch period detection unit 700 for detecting a pitch period of a voice input signal is provided, and a transmission unit for time-axis-compressing and encoding the voice input signal using the pitch period is provided. In the voice encoding device, a voice type determining unit 200 that determines whether the voice input signal is voiced sound, unvoiced sound, or consonant, and, in the case of the unvoiced sound or consonant, determines whether the voice input signal is positive or negative. When the voice input signal is voiced sound, it is output as it is, and when the voice input signal is unvoiced sound or consonant sound, the output of the signal conversion means 400 Switch means 300, 5 for adding to the pitch detector.
00 is provided.

(Claim 2) A pitch cycle detector 700 for detecting a pitch cycle of a voice input signal, and time-axis compression of the voice input signal for every two consecutive pitch cycle sections using the pitch cycle. 1. A voice encoding device having one time axis compression means 110 and having a transmission section for encoding and transmitting the output of the first time axis compression means, in a voiced or unvoiced sound for the voice input signal. A voice type determining means 200 for determining whether or not a consonant, and in the case of the unvoiced sound or consonant, inverts the sine wave function and the phase of the sine wave for the voice input signal in the two continuous pitch period sections. When the voice input signal is a voiced sound, the second time-axis compression means 120 for performing time-axis compression by multiplying the above function and the output of the voice type determination means, on the side of the first time-axis compression means, Again, unvoiced Is a consonant, switch means 105, 130 for switching to the second intermediate compression means side.
And.

(Claim 3) A voice decoding device for inputting a time-axis-compressed and encoded voice signal sent from the voice encoding device of the above-mentioned claim 2, for decoding and for expanding the time-axis. A voice type judging means 205 for judging voiced sound, unvoiced sound or consonant sound with respect to the time-axis-compressed sound signal, and a first time-axis expanding means 230 for expanding the time-axis in the case of the voiced sound. And in the case of the unvoiced sound or consonant,
Time is obtained by multiplying the time-axis-compressed audio signals in the front two and rear two pitch period sections of the three consecutive pitch period sections by a sine wave function and a function obtained by inverting the phase of the sine wave, respectively. When the voice signal is a voiced sound, the second time axis extending means 240 for axis extension and the output of the voice type determining means, to the first time axis extending means side,
Also, in the case of unvoiced sound or consonant, switch means 220, 250 for switching to the second time axis expansion means side are provided.

[0016]

In FIG. 1 (claim 1), when the voice input signal is unvoiced or consonant, the signal conversion means 400 converts the voice input signal into a positive or negative constant value according to the positive or negative. As a result, it is possible to improve the accuracy of calculation of the pitch period of unvoiced sounds / consonants whose level is originally small and the pitch period is not excessive.

(Claim 2) When the voice input signal is unvoiced or consonant, the sine wave function and the phase of the sine wave are respectively used as windowing coefficients for the voice input signal in the two continuous pitch period sections. By using the inverted function, as shown in FIG. 7 (a), it is made smaller in the adjacent portions of two consecutive pitch period sections Pp ₁ and Pp ₂ ,
By setting a larger value at both end portions, it is possible to maintain the continuity of signals in the portions adjacent to each other.

In FIG. 2 (claim 3), when the time-axis-compressed voice signal is unvoiced or consonant on the receiving side, three consecutive pitch period sections Pp ₁ ′, Pp ₂ ′, P
Using a sine wave function and a function obtained by inverting the phase of the sine wave as the windowing coefficient for the time-axis-compressed audio signal in the front two and rear two pitch period sections of p ₃ ', respectively. As shown in (b), the middle section Pp ₂ '
Is set to be larger and the sections Pp ₁ ′ and Pp ₃ ′ at both ends are set to be smaller, it is possible to maintain signal continuity in portions adjacent to each other. As a result, it is possible to provide high-quality reproduced voice for unvoiced sounds and consonants.

[0019]

FIG. 3 is a block diagram (transmission side) showing the configuration of a speech coder according to an embodiment of the present invention.

FIG. 4 is a block diagram (reception side) showing the configuration of the speech coding apparatus according to the embodiment of the present invention. FIG. 5 is a diagram for explaining a covariance value calculation method for voiced sound in the embodiment.

FIG. 6 is a diagram for explaining a pitch period extraction signal correction method for unvoiced sounds / consonants in the embodiment. FIG. 7 is a diagram for explaining the unvoiced / consonant part time axis compression / expansion method of the embodiment.

In FIG. 3, the voice input signal (SIN) is stored in the buffer 1 and then its level is calculated by the voiced sound / unvoiced sound / consonant determination unit 2. When the level is smaller than a predetermined threshold value, unvoiced sound or consonant sound, When it is loud, it is judged as voiced sound and the switching signal of switches 3, 5, 10 and 13 (SJUDG
E) is output.

When it is determined to be voiced sound, the respective switches are switched to the a side, and the voice input signal (SIN) of the constant section stored in the buffer 1 is input to the covariance calculation unit 6,
The pitch period of the certain section is obtained by using the above-mentioned covariance method (Formula 2) for the voice input signal (SIN).

As shown in FIG. 5, the operation for obtaining the pitch period is controlled by the pitch control section 9 so that the minimum one-pitch period candidate value Pmin is changed to the maximum one-pitch period candidate value P.
The candidate value of the pitch cycle is changed between max and the evaluation formula F (n) of the covariance in the candidate value of each pitch cycle is obtained. Figure 5
, N is an integer that satisfies the condition of Pmin = <n = <Pmax, and Sj represents sampling data (voice input signal) at a point j away from the point A. That is, n is changed from Pmin to Pmax and the covariance evaluation formula F (n) is calculated.
Ask for.

The maximum value search unit 7 weights the evaluation formula F (n) as shown in Formulas 3 and 4, and stores the weighted value of the evaluation formula Fw (n) in the RAM 8. Then, the maximum value among the values of each evaluation formula Fw (n) is obtained, and n at that time is set as the pitch period of the section.

[0026]

[Equation 3]

[0027]

[Equation 4] However, Fw: weighted evaluation formula, Wm: weighting function,
L: constant = K × Pmin, K: constant> = 0.

When the voiced sound / unvoiced sound / consonant judgment unit 2 judges unvoiced sound / consonant sound, the switches 3, 5, 10 and 13 are switched to the b side, and the voice input signal (SI
N) is input to the zero cross correction unit 4, and the voice input signal SI
The pitch period of a certain section is obtained by observing the correlation of N zero crosses only.

As shown in FIG. 6, when the voice input signal is positive, the positive constant value (k) is corrected and when the voice input signal is negative, the negative constant value (-k) is corrected to correct the voice input signal. Then, similarly to the voiced sound, the judgment is performed using the covariance formula (Formula 2). By performing the above correction, it is possible to improve the calculation accuracy of the pitch period of unvoiced sound / consonant whose level is originally small and the pitch period is not excessive.

Next, in the section judged as voiced sound, as shown in FIG.
The windowing coefficient shown in (a) is generated by the windowing coefficient generator 11,
In the section that is determined to be unvoiced, the windowing coefficient generator 12 generates the windowing coefficient shown in FIG. 7 (a), and the time axis compression unit 14 outputs a signal of two pitch periods (sections Pp ₁ and Pp ₂ ). Is time-axis compressed into a signal Sc (n) of one pitch period. The windowing coefficient in the case of unvoiced sound is a coefficient as shown in Expression 5.

[0031]

[Equation 5] However, Ww: windowing coefficient, p: pitch period, i: sample. Since this is a coefficient that differs depending on the pitch period, it is newly calculated each time the pitch period is obtained, but it may be prepared and used as a table for each pitch period in advance.

As a result, the windowing coefficient of unvoiced sound / consonant is made smaller by using Equation 5 in the adjacent portions of two consecutive pitch period sections Pp ₁ and Pp ₂ as shown in FIG. 7A. By setting a larger value at both end portions, it is possible to maintain the continuity of signals in the portions adjacent to each other.
The signal subjected to the time-axis compression processing is further compression-coded by CELP coding or the like in the encoding unit 15, multiplexed in the multiplexing unit 16 together with the information of the optimum pitch period read from the RAM 8 and sent to the transmission line.

On the receiving side, the signal sent from the transmission line is separated into coded data and optimum pitch period information by the multiplexing unit 17 in FIG. The encoded data is decoded by the decoding unit 18, and the time-axis compressed data RIN is output and stored in the buffer 19.

The voiced / unvoiced / consonant determination unit 20 calculates the level of the data RIN which has been time-axis compressed, and when it is smaller than a predetermined threshold value, it is determined as unvoiced sound or consonant, and when it is larger, it is determined as a voiced sound. 25 switching signals (RJUD
GE) is output. However, since the calculation of this level is performed in pitch cycle units, the threshold value is the value expressed by Equation 6.

[0035]

[Equation 6] However, Ps: threshold value, p: pitch period, K: constant> = 0. In the section judged as voiced sound, the windowing coefficient shown by the solid line in Fig. 8 (b) is generated by the windowing coefficient generator 23, and in the section judged as unvoiced sound or consonant, the windowing coefficient shown in Fig. 7 (b). The coefficients are generated by the windowing coefficient generator 24. This output is time-axis expanded by the time-axis expansion unit 21 and output as ROUT. The windowing coefficient of the unvoiced part is a coefficient as shown in Equations 7-9.

[0036]

[Equation 7]

[0037]

[Equation 8]

[0038]

[Equation 9] However, Ww: windowing coefficient, p: pitch period, i: sample, Ww _1: windowing a time axis extension coefficient 1, Ww _2: Time-base decompression windowing coefficient 2, Sc: time warping processed data , Se: Time-axis expanded data.

Since this coefficient is different depending on the pitch cycle, it is prepared as a table for each pitch cycle, but may be newly calculated for each pitch cycle. As a result,
As shown in FIG. 7 (b), the windowing coefficients of unvoiced sound and consonant are calculated by using Equations 7 and 8 as shown in FIG. 7 (b), and the middle section Pp of three consecutive pitch period sections Pp ₁ ′, Pp ₂ ′, and Pp ₃ ′. By making the value larger in ₂ ′ and smaller in the intervals Pp ₁ ′ and Pp ₃ ′ at both ends, it is possible to maintain signal continuity in the portions adjacent to each other. As a result, it is possible to provide high-quality reproduced voice for unvoiced sounds and consonants.

In the above embodiment, the compressed voice is analyzed,
Although the voiced sound / unvoiced sound / consonant is determined, a voiced sound / unvoiced sound / consonant determination signal may be transmitted from the transmission side through a transmission path. In that case, on the sending side, the judgment signal (SJUDJ
By multiplexing E) with optimum pitch period information and coded data in the multiplexing unit, the receiving side can omit the voiced sound / unvoiced sound / consonant determination unit 20 on the receiving side, and the multiplexing unit 17 has transmitted. The signal will be decomposed into an optimum pitch period, encoded data and a judgment signal (RJUDJE).

[0041]

As described above, according to the present invention,
(Claim 1) When the voice input signal is an unvoiced sound or a consonant, the signal converting means 400 converts the voice input signal into a positive or negative constant value according to positive or negative,
It is possible to improve the calculation accuracy of the pitch period of unvoiced sounds / consonants that originally have a small level and a few pitch periods.

(Claim 2) When the voice input signal is unvoiced or consonant, the sine wave function and the phase of the sine wave are respectively used as windowing coefficients for the voice input signal in the two continuous pitch period sections. As shown in FIG. 7 (a), by using a function obtained by inverting, the value is set to be smaller at the adjacent portions of two consecutive pitch period sections Pp ₁ and Pp ₂ and set to be larger at both end portions. , It is possible to maintain the continuity of signals in the portions adjacent to each other.

(Claim 3) On the receiving side, when the time-axis-compressed voice signal is unvoiced or consonant, the front two of the three consecutive pitch period sections Pp ₁ ′, Pp ₂ ′, Pp ₃ ′.
7 and a sine wave function and a function obtained by inverting the phase of the sine wave, respectively, are used as windowing coefficients for a time-axis-compressed audio signal in two pitch period sections and two rearward pitch period sections, respectively.
As shown in (b), it is made larger in the middle section Pp ₂ ',
By setting a smaller value in the sections Pp ₁ ′ and Pp ₃ ′ at both ends, it is possible to maintain signal continuity in the portions adjacent to each other. As a result, it is possible to provide high-quality reproduced voice for unvoiced sounds and consonants.

[Brief description of drawings]

FIG. 1 is a principle diagram of the invention of claims 1 and 2;

2 is a principle diagram of the invention of claim 3, FIG.

FIG. 3 is a block diagram (transmission side) showing a configuration of a speech encoding apparatus according to an embodiment of the present invention,

FIG. 4 is a block diagram (reception side) showing a configuration of a speech encoding apparatus according to an embodiment of the present invention,

FIG. 5 is a diagram for explaining a covariance value calculation method for voiced sound according to the embodiment;

FIG. 6 is a diagram for explaining a pitch period extraction signal correction method for unvoiced sounds / consonants in the embodiment,

FIG. 7 is a diagram for explaining the unvoiced / consonant part time axis compression / expansion method of the embodiment;

FIG. 8 is a diagram for explaining an example of a time axis compression / expansion method.

[Explanation of symbols]

 105, 130, 220, 250, 300, 500 are switch means, 110 is a first time base compression means, 120 is a second time base compression means, 200, 205 are voice type determination means, and 230 is a first time base. Axis expanding means, 240 is second time axis expanding means, 400 is signal converting means, and 700 is a pitch period detecting section.

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Mitsuru Tsuboi, 1015 Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture, Fujitsu Limited (72) Inventor, Naoji Matsuo, 1015, Kamiodanaka, Nakahara-ku, Kawasaki, Kanagawa Prefecture, Fujitsu Limited ( 72) Inventor Shuei Eguchi 3-22-8 Hakataekimae, Hakata-ku, Fukuoka City, Fukuoka Prefecture Fujitsu Kyushu Digital Technology Co., Ltd.

Claims (3)

[Claims] 1. A voice code comprising a pitch period detection unit (700) for detecting a pitch period of a voice input signal, and a transmission unit for time-axis-compressing and encoding the voice input signal using the pitch period. A voice type determining means (200) for determining whether the voice input signal is a voiced sound, an unvoiced sound or a consonant; and in the case of the unvoiced sound or a consonant, the voice input signal is determined to be positive or negative. And a signal converting means (400) for converting into a positive or negative constant value, and by the output of the voice type determining means, when the voice input signal is voiced sound, it is as it is, and when it is unvoiced sound or consonant, the signal conversion means. And a switch means (300, 500) for adding the output of the above to the pitch detection unit. 2. A pitch period detection unit (700) for detecting a pitch period of a voice input signal, and a first period compression unit for time-sequentially compressing the voice input signal for every two continuous pitch period sections using the pitch period. A time axis compression means (110),
In a voice encoding device having a transmission unit for encoding and transmitting the output of the time axis compression unit, a voice type determination unit (200) for determining whether the voice input signal is voiced sound, unvoiced sound, or consonant sound, In the case of the unvoiced sound or the consonant sound, a second time period in which the sine wave function and the function obtained by inverting the phase of the sine wave are applied to the voice input signals in the two continuous pitch period sections to compress the time axis. When the voice input signal is a voiced sound, by the axis compression means (120) and the output of the voice type determination means, to the first time axis compression means side,
A voice encoding device, characterized by comprising switch means (105, 130) for switching to the second time axis compression means side in the case of unvoiced sound or consonant sound. 3. A speech decoding apparatus for inputting a time-axis-compressed and coded speech signal sent from the speech coding apparatus according to claim 2, decoding, and expanding the time-axis, Voice type determining means (205) for determining voiced sound, unvoiced sound or consonant sound with respect to the axially compressed voice signal, and first time axis expanding means (230) for expanding the time axis in the case of the voiced sound And in the case of the unvoiced sound or consonant, the sine wave function and the sine wave function for the time-axis-compressed audio signal of the front two pitch period sections and the rear two pitch period sections of three consecutive pitch period sections, respectively. When the voice signal is a voiced sound, the second time axis expansion means (240) for expanding the time axis by multiplying the function of inverting the phase of the wave, and the output of the voice type determination means, In the case of unvoiced sound or consonant on the side of axial extension means Speech decoding apparatus characterized in that a switch means (220, 250) for switching the time-axis expanding means side of said second.