JPH09244695A - Voice coding device and decoding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make an unpleasant sound based on a background noise not to be generated by smoothing the fluctuation of voice featured values only in a voiceless section timewisely or by performing a stronger smoothing in the voiceless section than in a voice section. SOLUTION: In a voice coding device, various parameters becoming featured values of a voice are calculated in a featured value calculating part from an inputted voice signal. Next, whether the input signal is the voice section including the voice or the voiceless section of only a background noise at this point in time is judged based on the featured values in a voice/voiceless section discriminating part. When the input signal is judged to be the voiceless section, the fluctuation of the calculated featured values is suppressed in a featured value smoothing part. Conversely, when the input signal is judged to be the voice section, the calculated values are used as they arc without performing the smoothing of the featured values. Then, the featured values are quantized in a parameter quantizing part. Moreover, when the voice is inputted, optimum indexes for reproducing a sound near to an original sound are retrieved by retrieving code book part in an optimum driving signal selecting part.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding apparatus and a decoding apparatus for suppressing abnormal noise due to background noise which is a problem in mobile phones and the like.

[0002]

2. Description of the Related Art As an example of a voice compression / decompression (encoding / decoding) device to which the present invention is applied, a CELP (Code
The basic configuration of a mobile phone based on Excited Linear Prediction) is shown in FIGS. 10 and 11. FIG. 10 shows a configuration example of a speech coding apparatus, and FIG. 11 shows a configuration example of a speech decoding apparatus. The encoding device shown in FIG. 10 includes a “feature amount calculation unit” that calculates a feature amount of a voice from an input original voice, a “parameter quantization unit” that quantizes them to compress data,
A “codebook section” that stores a plurality of indexes corresponding to drive signals used to synthesize speech later,
It is configured to include an "optimal drive signal selection unit" that selects an index in an optimal codebook for reproducing a sound close to the original sound. In the optimum drive signal selection section,
An index corresponding to the drive signal such that the synthesized voice becomes a voice close to the original voice is selected from the plurality of indexes in the codebook. That is, when the optimum drive signal selection unit receives an input voice, it selects a plurality of indices from the indices in the codebook and sequentially inputs these to a voice synthesis filter to select a procedure for creating a synthesized voice. By repeating the above, the index that produces the synthetic speech closest to the input speech is extracted and output from the synthetic speech of the selected index number.
At this time, the voice synthesis filter is driven according to the coefficient corresponding to the voice feature amount.

In order to decode the encoded data, the quantized voice feature quantity, the selected index, etc. are used in FIG.
1 is transmitted to the decoding device, the value of the feature amount is decoded in the “voice feature amount decoding unit”, and the “driving signal generation unit” drives from the codebook based on the index transmitted from the encoding device. The signal is created. In addition, the coefficient of the “voice synthesis filter” is determined using the feature amount representing the spectrum envelope shape of the voice, and the voice is synthesized by inputting the drive signal to the filter. Here, the drive signal corresponds to a sound source such as human vocal cord vibration, and the synthesis filter that creates a spectral envelope shape corresponds to a transfer function until the sound source emits due to the shape of the vocal tract. Therefore, as the quantization table and the index stored in the codebook when quantizing the voice feature amount, those suitable for synthesizing the human voice are prepared based on various standards.

On the other hand, in the voice compression / decompression system which compresses voice at a low bit rate by such a system, when a sound other than a human voice is input, the original sound is always faithfully reproduced. It is known that the sound quality is extremely deteriorated even when there is a background noise such as air conditioning noise or rain noise that is present in the immediate environment. One of them is swirling
It is called noise, and it is a phenomenon in which the background noise becomes a very unpleasant sound that fluctuates like “curcules”. The reason for this is that in the frame that is usually used to calculate the speech features that represent the spectral envelope shape,
This is because the amount of data is too small for stable calculation of the noise feature quantity, and the speech parameters fluctuate.
ar Predictive Speech Codes ", Proc. ICASSP 1995 etc.

[0005]

As one of the conventional techniques for solving such a problem, Japanese Patent Laid-Open No. 7-152395 is available.
As described in the publication, there is a method in which the spectral feature amount of background noise is obtained in the unvoiced section, and the noise is attenuated by a filter that suppresses the noise spectrum based on it. However, although the S / N ratio is improved by this method, it is impossible to completely eliminate the noise, and the feature amount calculated from the remaining noise fluctuates and sounds unpleasant. Not done. In addition, background noise on telephones is also information that conveys the realism of the surroundings, and the disappearance of the background sound causes another problem of giving an unnatural impression.

As another conventional technique, in Japanese Laid-Open Patent Application No. 7-160294, a noise codebook is prepared on the decoding device side, and instead of a signal generated by an index or the like transmitted from the coding device. , In the unvoiced section, another signal selected from “Noise Codebook” is used. In this case, a method of replacing the signal from the “noise codebook” with the drive signal before input to the synthesis filter and a method of replacing with the synthesis signal that is the output of the synthesis filter are shown, but both of them are problematic. have. First, even if the signal before the synthesis filter is changed, the coefficient of the synthesis filter can be determined from the speech feature amount of the spectral envelope shape.
Since the cause of ng noise is the variation of the voice parameter, the synthesized voice output from the filter will eventually have an unpleasant variation. On the contrary, in the method of replacing the speech after the synthesis filter with the signal from the “noise codebook”, the original sound is reproduced unless all the noise signals having the spectral characteristics that may exist are prepared as the codebook. There is a problem that you cannot do it. Furthermore, the codebook search is the most computationally intensive part of the CELP method, and the codebook search can be performed not only by the encoding device but also by the decoding device by an extremely high-speed signal processing arithmetic unit. It will be necessary.

As another conventional technique, Japanese Patent Laid-Open No. 7-036485
In the gazette, a buffer for storing past speech signals is provided, and in the unvoiced section, the speech feature quantity is calculated using data with a larger number of frames than in the ordinary speech section, and the calculation result is calculated. It is trying to stabilize and suppress the fluctuation of the feature quantity. However, according to the above paper, 320 ms of voice data is used to stabilize the calculation results, and 2560 samples of data are required for 8 Khz sampling digital voice data. Therefore, this method requires a lot of memory resources only for the buffer. Therefore, an object of the present invention is to provide a speech coding apparatus and a decoding apparatus which do not generate an unpleasant sound based on the background noise without requiring a large memory capacity and without using a high-speed arithmetic unit. That is.

[0008]

A speech coding apparatus adopted by the present application to achieve the above object calculates a speech feature amount from a speech signal, compresses parameters constituting the speech feature amount, and transmits the compressed parameters. At the same time, in a voice encoding device for selecting an index corresponding to a drive signal for voice synthesis from a codebook according to the voice signal and transmitting the index together with the compressed parameter, a voiced / unvoiced section in the voice signal is identified. Either the coded voiced / unvoiced section discriminating means for smoothing temporally the variation of the voice feature amount only in the unvoiced section, or performing the smoothing stronger in the unvoiced section than in the voiced section. The speech coding apparatus is characterized by comprising coding smoothing means for performing smoothing processing. In this case, as the above-mentioned speech feature, a line spectrum pair is calculated, and in the unvoiced section, the value of each line spectrum pair is temporally smoothed and then the speech feature is quantized.

A first aspect of the present invention relating to a speech decoding apparatus receives a compressed parameter which is an output from a speech encoding apparatus and an index for generating a drive signal and the like, and outputs the speech from the parameter and the index. In a voice decoding device for restoring a feature amount and a driving signal and synthesizing a voice using them, a decoded voiced / voiced / voiced voice / voiceless part for identifying a voiced / unvoiced part in the voice signal from a compressed parameter from the voice encoding device. An unvoiced section identifying means and a smoothing process of either smoothing the variation of the speech feature amount temporally only within the unvoiced section or performing stronger smoothing than the voiced section in the unvoiced section. A voice characterized by comprising a decoding smoothing means and a voice synthesizing means for synthesizing a voice using each smoothed voice feature amount. It is configured as Goka device.
In this case, a line spectrum pair is calculated as the voice feature amount, and in the unvoiced section, the value of each line spectrum pair is temporally smoothed and then the drive signal is calculated to synthesize the voice.

The second aspect of the invention receives a compressed parameter which is an output from the audio encoding device and an index for generating a driving signal, and restores the audio feature amount and the driving signal from the parameter and the index. Then, in a speech decoding apparatus for synthesizing speech using them, a decoding voiced / unvoiced section identifying means for identifying a voiced / unvoiced section in the speech signal from a compressed parameter from the encoding apparatus, and an unvoiced section In the part determined to be the first, the first restored from the above codebook based on the above index
Creating a third drive signal in which some or all of the drive signals of are replaced by a second drive signal generated by another method,
The speech decoding apparatus is configured to synthesize speech based on the third drive signal. The second drive signal is generated by using a noise signal generated by a random number or the like, and the third drive signal is created by weighted addition of the first drive signal and the second drive signal restored from the codebook. be able to. Also, at the voice / unvoiced switching boundary, when creating the third drive signal,
The ratio between the first drive signal for weighted addition and the second drive signal such as a random number may be continuously changed.
Further, in the unvoiced part, the variation of the voice feature amount may be temporally smoothed, and the voice may be synthesized using the smoothed feature amount and the drive signal.

[0011]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of the present invention; It should be noted that the embodiments described below are examples embodying the present invention, and do not limit the technical scope of the present invention. 1 is a block diagram showing a configuration of a speech encoding apparatus according to an embodiment of the present invention, and FIG. 2 is a block diagram showing a configuration of a speech decoding apparatus according to an embodiment of the present invention. 3 is a graph showing the variation of the line spectrum pair related to the feature amount, FIG. 4 is a graph showing the variation of the line spectrum pair shown in FIG. 3 after quantization (without smoothing), and FIG. FIG. 6 is a block diagram showing the configuration of a speech decoding apparatus according to an embodiment of the present invention, and FIG. 7 is the above-mentioned diagram. 6 is a block diagram showing a configuration for smoothing a parameter of a feature amount in the speech decoding apparatus shown in FIG. 6, FIG. 8 is a graph showing a spectrum of background noise of original speech, and FIG. 9 is a spectrum of synthesized background noise. Fig. 10 shows a conventional CELP coding device. Block diagram illustrating the configuration of FIG. 1
FIG. 1 is a block diagram showing the configuration of a conventional CELP decoding device.

1 and 2 show the configurations of a speech coding apparatus and speech decoding apparatus according to an embodiment of the present invention.
First, in the voice encoding device (FIG. 1), various parameters that are the feature amount of voice are calculated from the input voice signal by the "feature amount calculation unit". Next, the "voiced / unvoiced section identifying unit" determines whether the input signal is a voiced section containing speech at the present time or an unvoiced section containing only background noise, based on the feature amount. When it is determined to be the unvoiced section, the calculated feature amount is suppressed in variation in the “feature amount smoothing unit”. On the contrary, when it is determined that the voiced section is used, the calculated value is used as it is without smoothing the feature amount. The subsequent steps are the same as those of the CELP system encoding device of FIG. 10 described in the above-mentioned related art, and the feature amount is quantized (compressed) in the “parameter quantization unit”. When a voice is input, the optimum drive signal selection section searches the codebook section, and the optimum index for reproducing a sound close to the original sound is searched from the “codebook section” among a plurality of indexes. In this way, the quantized feature parameter output from the encoder and the index resulting from the codebook search are transmitted to the decoder.

In the decoding device, as shown in FIG. 2, speech synthesis is performed based on them. That is, first, in the "speech feature amount decoding unit", the quantized parameters are inversely converted into the value of the voice feature amount. Next, the "voiced / unvoiced section identification unit" determines whether the voice state is a voiced section or an unvoiced section based on the feature amount. Of course, the result identified by the "voiced / unvoiced section identification unit" of the encoding device may be explicitly included as a parameter. When it is determined to be the unvoiced section, the calculated feature amount is suppressed in variation in the “feature amount smoothing unit”. On the contrary, when it is determined that the voiced section is used, the calculated feature value is used as it is without smoothing of. The parameter of the voice synthesis filter for synthesizing the voice is determined from the feature amount of the spectrum envelope shape of the voice subjected to such smoothing processing. In addition, the drive signal used as an input for synthesis is also restored from the index,
By inputting a drive signal to the "synthesis filter unit" determined by the feature amount parameter, a voice close to the original sound is synthesized. In many cases, the speech decoding device has a "post-filter section" (not shown) after the speech synthesis filter section.
And is output with the auditory quality of the synthesized speech being enhanced.

Next, the concrete calculation method of each part will be shown.
Descriptions of commonly used speech signal processing algorithms are given in reference books such as "Basics of Speech Information Processing (Ohm Co.)", and regarding processing parts similar to the general CELP method, refer to "RCR Standard". Since it is described in detail in standards such as “27C”, it will be omitted here. The main parameters calculated by the “feature amount calculation unit” include the volume of the voice, the spectrum envelope shape, and the pitch. The calculation of the characteristic amount is performed by analysis using temporally continuous audio data X (n), n = 1, 2, ... N _A called a frame. First, as the volume parameter, the average amplitude A of the audio signal in the frame as shown in the following equation can be adopted.

[0015]

[Equation 1] Further, as the feature quantity of the spectrum envelope shape, linear prediction coefficients α _i , i = 1, 2,
…, There is M.

[0016]

[Equation 2] _{Wherein, R i, i = 0,1,2,} ..., the audio data X M is a continuous (n), n = 1,2, ... the autocorrelation calculated by the following equation using N _a. However, as the N _a pieces of voice data which are usually used for calculation of R _i here, those subjected to a Hamming window or the like are used.

[0017]

(Equation 3) It is well known that the method of solving this equation is efficient if the Durbin recursive solution method is used. Moreover, as a method for obtaining the pitch, there is an autocorrelation method or the like. If the autocorrelation function of the prediction residual signal obtained as the output through the speech X (n) is calculated by the inverse synthesis filter represented by the following Z-transform, the autocorrelation at the portion corresponding to the pitch of the speech is calculated. It is known that the function has a peak.

[0018]

(Equation 4) With respect to the pitch, in addition to such a calculation method, a method of selecting the drive signal with an evaluation function so that the drive signal has an optimum periodicity in the codebook search described later is also well known. As the means for detecting the unvoiced portion in the "voiced / unvoiced section discrimination", a method based on voice or a method focusing on the periodicity of the signal is used. In other words, when there are a lot of frames with relatively low volume and little fluctuation, it is judged as an unvoiced section, and because the voiced part has a signal with periodicity due to the pitch, the part with less periodicity is referred to as an unvoiced section. A method for making a determination is known. The peak value of the autocorrelation function of the prediction residual signal by the linear prediction coefficient obtained by the calculation of the feature quantity is often used as the parameter representing the periodicity.

In the "feature amount smoothing section", the spectral envelope parameter which greatly contributes to the fluctuation of the unpleasant sound caused by the background noise is smoothed in time. However,
When the above linear prediction coefficient is smoothed as it is, there is no guarantee that the synthesis filter formed by the smoothed parameters is stable. If the synthesis filter becomes unstable, the synthesized output signal becomes an oscillating sound, which is extremely inconvenient. As a method of solving this problem, the linear prediction coefficients αi, i = 1, 2, ...
It may be smoothed after conversion into equivalent parameters called i, i = 1, 2, ..., M. The conversion into a line spectrum pair can be performed by finding the root of the polynomial in the following equation determined from αi if M is an even number. P (z) = 1 + (α ₁ + α ₁₀ ) z ⁻¹ + (α ₂ + α ₁₀ ) z
^-2・ ・ ・ + (α ₁₀ + α ₁ ) z ^-10 + z ^-11 Q (z) = 1 + (α ₁ −α ₁₀ ) z ⁻¹ + (α ₂ −α ₁₀ ) z
^-2・ ・ ・ + (α ₁₀ −α ₁ ) z ^-10 + z ^-11

If the roots of these two polynomials P (z) and Q (z) are stable, they have roots of the form zi = e- ^j.wi , and the root of P (z) at this time is 0 < ω ₁ <ω ₃ <ω ₅ <… <ω
_M-1 <π and 0 <ω ₂ <ω ₄ <ω ₆ <... <ω _M <π are called line spectrum pairs. Line spectrum pair is 0 <ω _i
It is known that it is stable as long as the condition of <ω _j <π (i <j) is satisfied, so that the stability can be easily maintained even after smoothing if the original parameters are stable. For example,
Let the line spectrum pair of the k-th frame be ωi (k), i = 1,
2, ..., M, smoothing can be performed as in the following equation. Ω _i (k) = β · Ω _i (k−1) + (1−β) · ω
_i (k) (however, Ω _i (1) = ω _i (1) in the initial state) 0 <β <1, and the larger β is, the stronger the smoothing is. Obviously, if the line spectrum pair ω _i (k) before smoothing satisfies the above-mentioned stability condition, then Ω _i (k) after smoothing also satisfies the stability condition.

In the unvoiced section, the
Smoothed line spectrum vs. Ω _i(K), i = 1,
If a synthesis filter is constructed based on 2, ..., M,
Unpleasant fluctuations in sound quality in the scene noise part are reduced. smooth
Coefficients of the synthesis filter corresponding to the pair of line spectra after conversion
Is e^-j.QiIf you develop a polynomial rooted at
Since the above polynomials P (z) and Q (z) can be constructed,
Quantized linear prediction coefficient a corresponding to αi_i, I = 1,
2, ... M can be obtained as follows.

[0022]

(Equation 5) If the parameters are smoothed in the voiced section, the spectrum of the human voice will be distorted. Therefore, use the original ω _i without smoothing or reduce the parameter fluctuation due to noise to the extent that the voice is not distorted. For this purpose, light smoothing with a small β may be performed.

The smoothing effect is most effectively applied to the feature quantity representing the spectrum envelope shape, but may be applied to other feature quantities such as volume and pitch, of course. Regarding the “parameter quantization”, the “codebook”, and the “codebook search”, various ones are used in various standards, and those conforming to them can be used in this embodiment as well. For example, "parameter quantizer"
Then, the number of bits of the parameter is reduced by scalar quantization or “vector quantization” to compress the data. In general, a "codebook" generally consists of an adaptive codebook that stores past drive signals and creates a periodic waveform, and a fixed codebook that compensates for portions that cannot be represented by the periodic waveform alone. Target.

Of the drive signals that can be generated from this codebook, the "codebook search" is to select the optimum drive signal for synthesizing a voice close to the original input sound. Here, a method of selecting a waveform that minimizes the evaluation function representing the perceptually weighted distortion is usually used.
At that time, for the adaptive codebook that creates the periodic waveform, the pitch and gain for creating the periodic waveform from the drive signal selected in the past are determined and output as an index, and the fixed signal waveform is stored as a table. In the codebook section, the optimal combination of waveforms selected from these signals, code, gain, etc. are output as an index. The quantized parameters and codebook indexes output from the encoding device are input to the decoding device, and the value of the feature amount is restored in the “feature amount decoding unit”. In addition, the drive signal selected by the encoding device is generated from the codebook index. The restored feature amount can be smoothed by the decoding device as well as the encoding device. As a result, even if the data transmission source coding apparatus does not have the feature quantity smoothing function, the decoding apparatus can provide the smoothing effect. The characteristic of the “synthesis filter” is set by the feature amount representing the complex spectrum envelope shape. In the present embodiment, a _i , i = 1, 2, ..., M obtained by converting the smoothed line spectrum pair into a linear prediction coefficient is used,
Speech is synthesized by inputting a drive signal to a filter represented by z conversion of the following equation.

[0025]

(Equation 6) The speech synthesized in this way is usually processed by some kind of "post filter" in order to improve the auditory sound quality. As post filters, filters such as high-frequency emphasis, spectrum shaping, and pitch emphasis are often used.

In the embodiment described above, the decoding device is
The unvoiced section / voiced section is determined based on the voice parameters sent from the encoder, and the section judged to be unvoiced is generated from the codebook by the index sent from the encoder. Since the drive signal generated by the codebook is used as it is, only the drive signal generated from the codebook is suitable for the human voice, and an unnatural sound is produced when used against background noise.
An example in which this can be changed to a sound that sounds more natural noise is shown below. At this time, in order to mitigate the unnatural feeling that the quality of the background noise changes abruptly between the voiced part and the unvoiced part, the first drive signal restored from the codebook and other methods such as random numbers are generated at the boundary part. When the third drive signal is created by weighted addition or the like from the generated second drive signal, the weighting for adding both may be continuously changed. Furthermore, in the unvoiced portion, a method of smoothing the speech feature amount sent from the encoding device in time to suppress the variation of the feature amount caused by background noise, and the above-described drive signal processing By using them together, the discomfort of background noise can be further reduced.

[0027]

[First Embodiment] First, an encoder outputs parameters for quantizing a speech feature amount calculated from original speech and an index for generating a drive signal from a codebook, and inputs them to a decoding device. Become. The value of the feature amount is restored in the "feature amount decoding" unit of the decoding device. In addition, the drive signal selected by the encoding device is generated from the codebook index. In the case of the decoding device, as shown in FIG. 6, it is possible to judge the voiced / unvoiced section based on the volume based on the quantized volume parameter sent from the encoding device. Further, the drive signal generated from the codebook corresponds to the pitch period vibration generated in the human vocal cords, and the periodicity can be detected from the autocorrelation function of this drive signal. Can be used for the determination.

In the "driving signal generating section", the first driving signal is generated from the "codebook" section based on the index of the codebook selected by the encoding device. In addition, in the present embodiment, a "random noise generating section" such as a random number is provided, and unlike the first drive signal, the second drive signal is generated by a natural sound source waveform as more normal background noise.
If the unvoiced portion is identified, the "driving signal processing section" replaces all or part of the first driving signal with the second driving signal to generate the third driving signal. As a specific example, assuming that the first drive signal is e1 (i) and the second drive signal is e2 (i), i = 1, 2, ..., N (N is the frame size), The third drive signal can be created by a linear sum with different weights. e ₃ (i) = g1 · e ₁ (i) + g2 · e ₂ (i) g1 and g2 are weights for adding e1 and e2 together. For example, if e1 and e2 are independent and have no correlation, By using the formula, e3 having the same level of power as the original drive signal e1 can be produced.

[0029]

(Equation 7) Of course, make the values of g1 and g2 smaller than this,
It is also possible to suppress the background noise in the unvoiced part. In the voiced part, the first drive signal is basically used as it is (K = 0). However, such a replacement of the drive signal causes a sudden change in the background noise quality at the boundary between the voiced part and the unvoiced part. To reduce it, it is possible to change K continuously when switching between the voiced and unvoiced parts. The generated drive signal is passed through a “synthesis filter” to synthesize voice. The feature of the synthesis filter is set by the feature amount representing the decoded spectral envelope shape.
Well-known parameters include parameters a _i , i = 1, 2, ..., M called linear prediction coefficients. Using this, a drive signal is input to the filter represented by the z conversion of the following equation to synthesize voice.

[0030]

(Equation 8) The parameters used in the synthesis filter, which represent the spectral envelope, are calculated from the original speech in the coding device, quantized, and passed to the decoding device. At that time, for the purpose of making it less susceptible to quantization errors, etc. , PARCOR coefficients and line spectrum pairs are often converted into parameters and then quantized. Since these are parameters equivalent to the linear prediction coefficient, they can be converted to each other. The basics of these audio signal processing are described in detail in reference books and the like (see “Basics of audio information processing”, Ohmsha, etc.), and therefore detailed description thereof is omitted here.

In particular, in the case of a line spectrum pair, the stability is assured even if it is smoothed in time. Therefore, by smoothing the parameters in the unvoiced part, the unnatural fluctuation of the background noise is suppressed, and further, It is also possible to reduce discomfort. FIG. 7 shows an embodiment in which such parameter smoothing is also performed. The synthesized speech is usually processed by some kind of "post filter" (not shown) in order to improve the audio quality. As post filters, filters such as high-frequency emphasis, spectrum shaping, and pitch emphasis are often used. (Refer to RCR standard 27C, etc.)

[0032]

The effects obtained by the present invention are shown below. Further, FIG. 3 shows the values of the line spectrum pairs calculated from the input voice in which the rain sound of about 20 dB exists as the background noise. The line spectrum pair was calculated using the data of 20 ms (160 samples) including the data before and after the frame. When a line spectrum pair is quantized to this value by a conventional method without smoothing, for example, it becomes as shown in FIG. On the other hand, according to the present invention, the line spectrum pair is smoothed in the encoding device and the decoding device, as shown in FIG. As a result, it was found that the fluctuation of the characteristic amount of the background noise part in the unvoiced section was suppressed, and even in the sensory test by hearing, the discomfort caused by the unnatural fluctuation of the sound quality was reduced.

The characteristics of the synthesis filter in the unvoiced part are
It represents the background noise spectrum characteristic. Therefore, even when synthesizing a voice with e3 as an input, the output spectrum is similar to the original background noise, as seen in FIGS. FIG. 8 and FIG. 9 are the spectrum of the air conditioning noise that is the background noise and the spectrum of the sound after being synthesized by the compounding device of the present invention. Of course, the background noise synthesized by this method is different from the original sound, but according to the sensory test, the background noise synthesized by the third method is better than the sound output by the conventional method using the first drive signal as it is. The result is that the sound synthesized by the drive signal is more natural and has no discomfort.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an encoding device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a decoding device according to an embodiment of the present invention.

FIG. 3 is a graph showing variations in a pair of line spectra (values in the feature amount calculation unit).

FIG. 4 is a graph showing a variation of a pair of line spectra after quantization (values obtained without smoothing).

FIG. 5: Quantized line spectrum pair (smoothed value)
The graph which shows the fluctuation of.

FIG. 6 is a block diagram showing a configuration of a speech decoding apparatus according to an embodiment of the present invention.

7 is a block diagram showing a configuration for smoothing a parameter of a feature amount in the speech decoding apparatus shown in FIG.

FIG. 8 is a graph showing a spectrum of background noise of original speech.

FIG. 9 is a graph showing a spectrum of synthesized background noise.

FIG. 10 is a block diagram showing the configuration of a conventional CELP encoding device.

FIG. 11 is a block diagram showing a configuration of a conventional CELP decoding device.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Toru Sakata 1-5-5 Takatsukadai, Nishi-ku, Kobe-shi, Hyogo Prefecture Kobe Steel Research Institute, Kobe Steel Co., Ltd. (72) Inventor Takayuki Higata Nishi-ku, Kobe-shi, Hyogo Prefecture Takatsukadai 1-5-5 Kobe Steel Co., Ltd. Kobe Research Institute

Claims (8)

[Claims] 1. A voice feature amount is calculated from a voice signal, parameters constituting the voice feature amount are compressed and transmitted, and an index corresponding to a drive signal for voice synthesis is codebook according to the voice signal. A voice coding apparatus for selecting and transmitting the voiced / unvoiced section in the voice signal; and a voice feature amount only in the unvoiced section. And a coding smoothing means for performing a smoothing process for either smoothing the fluctuations in time or a stronger smoothing in the unvoiced section than in the voiced section. Speech coding device. 2. The speech coding apparatus according to claim 1, wherein a line spectrum pair is calculated as the speech feature quantity, and in the unvoiced section, the value of each line spectrum pair is temporally smoothed and then the speech feature quantity is quantized. 3. A compressed parameter which is an output from a speech encoding device, an index for generating a driving signal, etc. are received, a speech feature amount and a driving signal are restored from these parameters, indexes, etc. In a speech decoding apparatus for synthesizing speech using a decoding voiced / unvoiced section identifying means for identifying a voiced / unvoiced portion in the speech signal from a compressed parameter from the speech encoding apparatus, and in the unvoiced section Only after smoothing the variation of the above-mentioned speech feature quantity, smoothing processing that smoothes temporally or stronger in the unvoiced section than in the voiced section, and after smoothing And a voice synthesizing means for synthesizing a voice by using each voice feature amount of. 4. The voice according to claim 3, wherein a line spectrum pair is calculated as the voice feature quantity, and in the unvoiced section, the values of each line spectrum pair are temporally smoothed and then the drive signal is calculated to synthesize the voice. Decoding device. 5. A compressed parameter which is an output from the audio encoding device, an index for generating a drive signal, etc. are received, a voice feature amount and a drive signal are restored from these parameters, indexes, etc. In a speech decoding apparatus for synthesizing speech by using a voice parameter in the speech signal from the compressed parameters from the encoding apparatus,
Decoded voiced / unvoiced section identification means for identifying unvoiced parts,
In the portion determined to be the unvoiced section, a third drive signal obtained by replacing a part or all of the first drive signal restored from the codebook based on the index with a second drive signal generated by another method is used. A voice decoding device characterized by generating a drive signal and synthesizing a voice based on the third drive signal. 6. A noise signal generated by a random number or the like is used as the second drive signal, and the third drive signal is created by weighted addition of the first drive signal and the second drive signal restored from the codebook. The audio decoding device according to claim 5. 7. At the voiced / unvoiced switching boundary portion, a first weighting addition is performed when the third drive signal is created.
7. The speech decoding apparatus according to claim 5, wherein the ratio of the drive signal of 1 to the second drive signal such as a random number is continuously changed. 8. The unvoiced part smoothes temporally the variation of the voice feature quantity, and synthesizes the voice using the smoothed feature quantity and the drive signal. The speech decoding device according to any one of claims.