JPH08272394A - Voice encoding device - Google Patents

Abstract

PURPOSE: To provide a voice encoding device capable of gaining excellent tone quality of a voice even in bad environment with more noises. CONSTITUTION: A pie-filter 104 constituting a prestage of the voice encoding device is constituted of a segmentation means 200, a formant emphasis means 201 and a gain control means 202, and a voice signal supplied from an input terminal 100 to the segmentation means 200 and the formant emphasis means 201 is divided to the voice/a noise by the segmentation means 200, and its voice/ noise decision information is imparted to the gain control means 202, and the voice signal is supplied to the gain control means 202 after a formant is emphasized by an LPC coefficient anti-quantized by the formant emphasis means 201 at every sub-frame to be supplied to an auditory sense weighting filter 105 after the gain is controlled based on the voice/noise decision information.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to CELP (Code-Excit
ed Linear Predictive Coding That is, "code-driven excitation coding method") and MPC (Multi-Pulse Excited Linear)
PredictiveCoding, that is, "ABS" method (Analysis-by-Synthesi) such as "multipulse excitation coding method")
s That is, the present invention relates to a noise removal device of a voice coding system to which the “analysis method by synthesis” is applied.

[0002]

2. Description of the Related Art In recent years, it has been required to digitally convert a voice signal and encode it to effectively utilize a memory or the like provided in a device for communication or the like. Among them, "synthesis" of MPC, CELP and the like mentioned above is required. Analysis by “Ab
-Study on hybrid coding technology using S method is active.

Regarding CELP, which is a typical technique of this kind of technique, the following documents are listed. M.
R. Schroeder and BSAtal, "Code-excited linear pred
iction (CELP): High quality speech at very low bit r
ates "(Proc. ICASSP, p.937-940, 1985) As shown in this document, a linear predictive analysis (LPC; Line Predictive Coding) is calculated by performing a linear predictive analysis on an input signal,
The synthesized signal is obtained by aurally weighting the synthesized signal composed of white noise in the codebook and the pitch extracted from the long-term predictor, and the same as the aurally weighted synthetic signal. The index that minimizes the squared error from the perceptually weighted input signal is transferred to the output side together with the pitch and the LPC coefficient. As described above, this technique is a method of obtaining the index, pitch, and LPC coefficient that have the smallest distance from the input signal.
In other words, such a method is a method of selecting a sound source by obtaining a synthesized signal that is most nearly acoustically weighted to an acoustically weighted input signal,
It is abbreviated as “A-b-S method”.

According to this method, for example, when the input signal contains only the audio signal, a high quality audio signal is reproduced. However, in the operation in the actual environment, only the voice is not always included in the input signal. For example, in a car, engine noise of the car and noise from the road are included, and in a room, background noise such as noise generated from electric appliances such as an air conditioner is usually included. Therefore, there is a problem in that the processing quality of the voice itself is deteriorated if the input signal as it is is directly encoded.

Therefore, various methods have heretofore been proposed as methods for suppressing the noise component in the input signal as described above. For example, Suda, Ikeda, and Ikedo, “PSI-CELP error control and signal processing technology” (N.
(TT R & D vol.No.4, page.373-380, 1994), the method of using the "Kalman filter" is known as a pre-processing of the method of suppressing the noise component of the input signal. There is. However, in order for this Kalman filter to exert its predetermined function, complicated arithmetic processing is required, so if real-time processing is to be performed, dedicated large-scale hardware is essential.

Therefore, an object of the present invention is to provide a speech coding apparatus which can solve the above-mentioned problems and can obtain good speech quality even in a noisy and poor environment.

[0007]

In order to overcome the above-mentioned problems and achieve the object, the present invention takes the following means. That is, emphasizing the phonological factor (formant) contained in the input speech signal (hereinafter,
This concept is called "formant emphasis") and
The speech or noise is divided into a plurality of segment units (hereinafter, this process is referred to as "segmentation"), and the encoding process is performed as a signal obtained by suppressing the noise gain (hereinafter, referred to as "reference signal"). To do.

That is, according to the present invention, the "analysis method by combining" with the reference signal which is a signal in which the gain of noise is suppressed,
That is, a speech coder for performing coding processing by the A-B-S method, a means for emphasizing the formant of this reference signal, a segmentation means for dividing the speech part and the noise part of the signal, and this segmentation. And a means for suppressing the level of the noise part obtained by the means.

[0009]

The present invention has the following actions by the above-mentioned means. The input signal is segmented by speech / noise for each subframe, and after the so-called “formant enhancement” of the input signal, the amplitude of the noise part is suppressed, and the noise-suppressed signal is A−b− It is used as a reference signal used in the S method.

Further, in the present invention, a linear prediction coefficient is extracted from the input signal for each frame, formant enhancement and filtering are performed for each subframe based on the linear prediction coefficient, and the input signal is segmented by speech / noise. The signal that has been processed, the amplitude of the noise part is suppressed, and the noise is suppressed becomes a reference signal used in the A-B-S method.

As a result, noise can be suppressed for each subframe in real time by a simpler process compared to the prior art. In addition, the hardware configuration of the device itself is simplified by the process of applying the same linear prediction coefficient.

[0012]

【Example】

(First Embodiment) FIG. 1 shows the configuration of a first embodiment of a speech coder according to the present invention. As an input signal of this device, for example, it is sampled at 8kHz,
An audio signal divided into frame units of ~ 30ms
It is input from the terminal 100. This input signal is linearly predicted (L
The LPC analyzer 101 for performing the PC) and the pre-filter 104 (which pre-processes the input signal for transmission) are branched and input.

As shown in the figure, the LPC analyzer 101 quantizes the LPC coefficient obtained from the input signal.
The LPC quantizer 102 is connected to the quantizer 102, and the LPC quantizer 102 further includes an LPC dequantizer 103 for dequantizing the quantized LPC coefficient, and a multiplexer 119 for performing a multiplexing process as an output stage of the present apparatus. It is connected to the.

On the other hand, the pre-filter 104 includes a perceptual weighting filter 105 for perceptually weighting and the above L
It is connected to the PC inverse quantizer 103. The perceptual weighting filter 105 is further connected to a zero input filter 106 that outputs a reference signal.

Further, the LPC dequantizer 103 described above is
Not only the pre-filter 104, but also the auditory weighting filter 105, the zero-input filter 106, and the weighting synthesis filter 10 for synthesizing weighted synthetic speech
It is also connected to 7.

The respective constituent elements of the device of the present invention function as follows. First, the LPC analyzer 101 performs LPC analysis on the order of 10th to 12th for each frame,
LPC quantizer 102 as the linearly predicted LPC coefficient
Is output to

The LPC quantizer 102 is an LPC analyzer 10.
The LPC coefficient received from 1 is subjected to predetermined quantization processing to reduce the number of bits required for transfer to the output side, and then the multiplexer 119 and the LPC inverse quantum are used as the index ci of the LPC coefficient. It is supplied to the rectifier 103. The LPC dequantizer 103 dequantizes the index ci of the LPC coefficient. L dequantized in this way
The PC coefficient is supplied to the pre-filter 104, the perceptual weighting filter 105, the zero-input filter 106, and the weighting synthesis filter 107. At the same time, the input terminal 10
The audio signal input from 0 is pre-filtered by the pre-filter 104 for each sub-frame consisting of a quarter of the frame. Here, this pre-filter 104
LPC coefficients given to the perceptual weighting filter 10
5 and the zero input filter 106 and the equivalent to the weighting synthesis filter 107 are used.

Here, the processing unit used in each filter processing is the same subframe, whereby the structure of the apparatus itself can be simplified. Here, the detailed configuration and functional operation of the pre-filter 104 will be described with reference to FIG.

As shown, this pre-filter 104
Is composed of a segmentation means 200, a formant emphasis means 201, and a gain control means 202. The segmentation means 200 and the formant emphasis means 201 are connected to the input terminal 100 in series, respectively. Further, the segmentation means 200 is connected to the gain control means 202, and the gain control means 20
2 is connected to the perceptual weighting filter 105. Further, the formant emphasizing means 201 is the gain control means 2
02 is connected.

The pre-filter 104 has an input terminal 10
After processing the audio signal from 0, it is supplied to the segmentation means 200 and the formant emphasis means 201. The voice signal is divided into voice / noise by the segmentation means 200, and its judgment information is gain control means 20.
Given to 2. At the same time, the formant enhancement unit 201 enhances the formant for each subframe by using the dequantized LPC coefficient, and then the voice signal is supplied to the gain control unit 202 and gained based on the voice / noise determination information. Is controlled, the auditory weighting filter 1
It is supplied to 05.

Returning again to the block diagram of the apparatus shown in FIG. 1, the audio signal generated as described above is auditorily weighted by the auditory weighting filter 105, and then the zero input filter. It is input to 106 and then to the subtractor 120 as a reference signal.

On the other hand, synthetic speech is generated as follows.
First, the evaluator 108 gives an appropriate initial value of the index to the adaptive codebook 109, and the code corresponding to the initial value of this index is output from the adaptive codebook to the multiplier 112. Adaptive codebook 109
While the code is output from the multiplier 112 to the multiplier 112, the initial value of the index is given to the adaptive gain codebook 114 from the control means (not shown) via the terminal 117, and the corresponding gain is output from the adaptive gain codebook 114. It is then supplied to the multiplier 112. This multiplier 11
In 2, the code from the adaptive codebook is multiplied by the gain from the adaptive gain codebook, and the multiplication result is input to the weighting synthesis filter 107 via the adder 116. The signal input to the weighting synthesis filter 107 is synthesized as a weighted synthetic speech, subtraction processing is performed with the reference signal output from the zero-input filter 106 by the subtractor 120, and the signal is input to the evaluator 108 as an error signal. To be done.

The evaluator 108 has an adaptive codebook 109.
The index ia of the adaptive codebook and the gain and its index iga that minimize the squared error obtained from this error signal are obtained while updating the index given to This is the end of the adaptive codebook search process.

Next, the gains corresponding to the adaptive codebook index ia and the adaptive gain codebook index iga obtained by the adaptive codebook search process are fixed, and the stochastic codebook 110 and the stochastic gain codebook 115 are searched. Is performed in the same manner as the adaptive codebook search process. A sound source created based on the index and gain searched by the adaptive codebook, a code from the probability codebook 110, and a probability gain codebook 11
The sound source created based on the gain from 5 is added by the adder 116 to create a new sound source signal, which is synthesized by the weighting synthesis filter 107, from the error signal between the synthesized speech synthesized and the input signal. The index is of the probability codebook and the index igs of the probability gain codebook with which the obtained squared error is the minimum are obtained. The adaptive codebook ia, the adaptive gain codebook index iga, the probability codebook index is, the probability gain codebook index igs, and the LPC coefficient index ci determined as described above.
Are multiplexed by the multiplexer 119 and output to a recording medium or a communication path (not shown). Further, the sound source signal added through the adder 116 based on the above indexes is, igs, iga, and ia is input to the shift register 111, and the time is equivalent to one subframe (a quarter of one frame). Are input to the adaptive codebook 109 as a sound source signal that is shifted in a temporal sense, that is, a sound source signal of "past" for one subframe, and the adaptive codebook 109
Configure the code for.

Here, the operation of the pre-filter 104 will be described in detail. The input signal is input for each subframe, and the segmentation means 200 outputs the audio / voice.
It is divided into noise and the voice flag which is the output from the segmentation means 200 is turned on / off. For more information,
For example, when the voice flag is on, the input signal is voice, and when it is off, the input signal is noise. Next, the formant emphasizing means 201 emphasizes the input signal formant according to the following equation (1).

[0026]

[Equation 1]

In Equation 1, α _i is the i-th order LPC coefficient, N is the order of the LPC coefficient, η is the weighting coefficient of the zero filter, and ν is the weighting coefficient of the pole filter. The weighting coefficient η of the zero filter and the weighting coefficient ν of the pole filter have a relationship of 0.0 <η <ν <1.0. Further, as a preferable example, for example, η =
0.2 and v = 0.7.

By this filter, the formant is emphasized in the frequency plane, and the antiformant part is suppressed in the opposite direction, so that the intelligibility of the voice on which the noise is superimposed is improved. next,
The steps of the gain control process performed by the gain control means will be described based on the flowchart of FIG.

The gain control means 202 suppresses the amplitude of the voice in a predetermined section having background noise, but here, the area where the amplitude is suppressed stepwise before and after the noise section, that is,
Provide a "fade in / fade out" area. (See below).

First, it is determined whether the voice flag is on or off (S302). As a result, if the voice flag is off, the gain value is checked, that is, it is determined whether or not the gain is the lower limit value (eg, 0.5) of the silence suppression coefficient (S3).
03).

At this time, when the gain value is the lower limit value of 0.5 of the suppression coefficient, the gain is multiplied by the so-called "formant-emphasized" input voice (S30).
Four). On the other hand, if the lower limit value is not 0.5, the formant-emphasized input voice is multiplied by a gain reduced by a certain fixed value (S305).

Assuming that this constant value is linearly reduced from a state in which no gain is applied, that is, from 1.0 to the lower limit value 0.5 of the noise noise suppression coefficient of gain over 10 subframes, The amount of change per subframe is 0.05. Such a process is called a "fade out" process.

If the voice flag is ON, the gain value is inspected as in the OFF state (S304). If the gain value is the upper limit value of the voice part suppression coefficient, ie, 1.0, the formant enhancement is performed. The input voice that is input is output as it is.

On the other hand, if the upper limit is not 1.0, the formant-emphasized input voice is multiplied by the gain increased by a certain fixed value (S307). Such processing is called "fade-in" processing.

Further, for example, if this constant value is linearly increased from the lower limit value of 0.5 to the upper limit value of 1.0 for the noise part suppression coefficient of gain over 10 subframes, as in the case of the fade-out described above, 1 The amount of change per subframe can be set to 0.05.

In the above example, the so-called fade-in / fade-out is linearly changed over the same number of sub-frames, but this may be different, for example, when the number of sub-frames applied for fade-in and fade-out is different, or is not always linear. It doesn't have to be.

As described above, by changing the gain over a plurality of subframes, it is possible to smooth the change of the input signal from voice to noise or from noise to voice. Further, even if the noise is erroneously determined to be speech by one subframe in the segmentation means 200 due to impulse noise while the noise is continuing, the amplitude of the noise is from 0.5 to 0.5. It is increased by 0.05 and multiplied by 0.55, and the impulse noise in the reference signal is also 0.55 times the actual impulse noise, resulting in impulse noise. Can be reduced.

As described above, by using the processed input signal as a new input signal, noise can be suppressed during the encoding process. Note that in the above configuration, the pre-filter 104 is independent of the subsequent coding devices. So turn on the prefilter /
It is also possible to turn it off, and when this prefilter is set to off, it functions as a conventional encoding device. By providing the switching means for that purpose, when the noise is so small that it does not affect the sound quality, the power consumption can be saved by turning off this pre-filter.

The segmentation processing of the voice part and the background noise part of the input signal is performed using a predetermined threshold value.
Hereinafter, a method that can appropriately perform the segmentation processing even when the background noise temporally changes by appropriately changing the threshold value will be described.

The threshold value is obtained by the sum of energy or amplitude of the input signal for each subframe. The energy is the sum of the squares of the values of the samples for the number of samples included in the subframe, and the sum of the amplitudes is the absolute value of each sample for the number of samples included in the subframe. Is the sum of

Next, the processing when the energy of the input signal for each subframe is used will be described. First, an example of a basic processing algorithm of segmentation processing is shown. (1) Let the value that is the energy Eth be a fixed threshold value (the upper limit value of the threshold value) of voice and noise. (2) The average energy for 10 subframes is set to the initial threshold value E.
Let th1. (3) The average energy of 10 subframes is calculated as the temporary threshold Eth2 while shifting by 1 subframe. (4) If the ratio between the temporary threshold Eth2 and the initial threshold Eth1 is Cth or less, it is determined to be noise, and the initial threshold Eth1 is set to the temporary threshold Eth.
Replace with 2. (5) If the ratio between the temporary threshold Eth2 and the initial threshold Eth1 exceeds Cth, it is determined to be voice. (That is, detection of rising) (6) When 5 subframes in which the temporary threshold Eth2 corresponds to less than Eth and exceeds the initial threshold Eth1 are accumulated, it is determined to be voice, and the initial threshold Eth1 is set to the temporary threshold Eth.
Replace with 2. (Note that such a state is referred to as a "speech state".) (7) When the subframes whose provisional threshold value Eth2 is less than or equal to the initial threshold value Eth1 are continuous for 5 consecutive subframes, it is determined to be noise and the initial threshold value is determined. Eth1 is replaced with a temporary threshold value Eth2. (Note that such a state is referred to as a “noise state”.) Next, the steps of the segmentation processing performed by the concrete segmentation means will be described with reference to the flowchart of FIG.

First, prior to execution of a series of processing, an initialization flag indicating the start of this processing is turned on. The initialization flag is set to be turned on (that is, the power is turned on) when the power of the device is turned on or turned off.

Then, the energy eng of the input signal for each subframe is calculated (S400). In the process corresponding to (1) above, by determining Eth, a threshold value Eth is found by finding a value by which speech and noise can be clearly separated from each other in terms of energy of the input signal for each subframe. This threshold Eth is the maximum value of the thresholds of voice and noise, and if it exceeds this value, it is unconditionally determined to be voice.

The process corresponding to (2) above is performed by the initial threshold value E.
This is a process for obtaining th1, and the initialization flag is inspected (S401). When the initialization flag is on, the energy eng of the input signal for each subframe is compared with the threshold Eth for separating voice / noise (S402). If the energy eng is larger than the threshold value Eth in this comparison, the voice flag is set, that is, turned on (S40).
7). Then, as the energy eng of the input signal, for example, 10 subframes are stored in the energy buffer provided in the segmentation means from the oldest one to the newest one (S403).

This energy buffer performs the same function as the shift register, and always stores the latest 10 sub-frame energies. If the energy eng is less than or equal to the threshold value Eth, the voice flag is turned off (S404), and it is checked whether or not the energy buffer is filled with energy for 10 subframes (S405). If this buffer is not filled with energy, the routine exits the routine with the voice flag off. That is, it branches to the end.

On the other hand, when the buffer is filled with energy, the average value is obtained from the energy for 10 subframes and held as the initial threshold value Eth1 (S406), and at the same time, the initialization flag is turned off.

The reason why the above-mentioned operation for obtaining the average value is necessary is to eliminate the variation in energy between the sub-frames as much as possible. Also, in this example,
The average value of 10 subframes is taken, but instead of this, the value of so-called "power of 2", such as "4", "8", "16", is fixed. Since decimal point calculation can be achieved by only shifting to the right, it is suitable for the purpose of reducing the amount of calculation.

The process corresponding to the above (3) is performed by the provisional threshold Eth.
In the process of obtaining 2, the oldest energy in the energy buffer provided in the segmentation means is first discarded, and the energy thus obtained is input to update the energy buffer (S408). Further, the average value is obtained as in the case where the initialization flag is turned on, and is stored as the temporary threshold value Eth2 (S409).

In the process corresponding to the above (4), first, similarly to the initialization operation, the sub-frame energy eng obtained here is compared with a fixed threshold Eth for separating voice / noise (S410). Subframe energy eng is threshold Eth
If it is larger than the above, the voice flag is turned on, and at the same time, "5" is set to the voice state counter that measures the number of subframes in the voice state, and "5" is set to the noise state counter that measures the number of subframes in the noise state. Yes (S407).

As will be described later, the voice state counter and the noise state counter respectively decrement (at the time of the state).
That is, subtraction) is performed. Also, the sub-frame energy en
When g is equal to or less than the threshold Eth for separating voice / noise, the temporary threshold Eth2 is compared with the initial threshold Eth1 (S41).
1). In this comparison, when the provisional threshold value Eth2 is larger than the initial threshold value Eth1, that is, when the signal energy tends to increase, "5" is set to the noise state counter (S4
12) Next, the ratio between the temporary threshold value Eth2 and the initial threshold value Eth1 is obtained, and the ratio is compared with a certain constant value Cth (S4).
13). If the value of the ratio is less than or equal to Cth, it is considered as an error between subframes in the noise part, the voice flag is set to OFF (S414), and the initial threshold Eth1 is replaced with the temporary threshold Eth2 (S415).

By performing such a series of processing,
Even if the background noise in the environment in which the device is used is increased temporally, if the rate of change is less than a certain rate, that is, Cth or less, the threshold is changed to follow the change in the background noise. It becomes possible.
This is based on the principle that the increase of background noise is extremely gentle with respect to the rising of voice.

In the process corresponding to (5) above, the ratio value is C
If th is exceeded, it is determined to be voice, and the voice flag is turned on (S416). By this processing, the rising edge of the input signal is detected.

In the process corresponding to (6) above, it is checked whether the voice state counter is "0" (S417), and if the voice state counter is not "0", the counter is decremented (S420). . On the other hand, if the voice state counter is "0", the initial threshold value Eth1 is replaced with the temporary threshold value Eth2 (S41
8), reset "5" to the voice status counter.

The above-described processing is processing for facilitating detection of the rising edge of the signal. At the time when the rising edge of the signal is detected, the initial threshold value Eth1 is no longer changed, and the voice state is used for detecting the end. Is continuous for 5 subframes or more, the new initial threshold Eth1 is set by rewriting the initial threshold Eth1 with the audio flag kept on.

The process corresponding to the above (7) is executed by the provisional threshold Eth.
When 2 is less than the initial threshold Eth1, that is, when the signal energy tends to decrease, it is determined to be noise and the voice flag is turned off (S421). It is checked whether the noise state counter is "0" (S422). If the noise state counter is not "0", the counter is decremented (S425). If the noise state counter is "0", the initial threshold Eth1 is set to the temporary threshold E.
Replaced with th2 (S423), "Noise state counter"
Set 5 "again.

Similar to the processing steps S417 to S420, this processing is processing for facilitating detection of the end of the rising signal. When a noise state continues for five subframes, the voice flag is turned off and the initial threshold value Eth1 is rewritten to detect the end point.
This is a process of setting 1 and preparing for a new rising.

Further, this pre-filter can be used alone as a noise suppression filter, for example, as a so-called "post-filter" after decoding. The operation of the segmentation means described above can be applied not only to speech coding but also to other speech processing, such as speech recognition.

(Second Embodiment) Next, FIG. 2 is a functional block diagram showing a second embodiment of the speech coding apparatus of the present invention.

The present embodiment is characterized in that the formant emphasizing means (see FIG. 4) simultaneously performs the formant emphasizing process and the auditory weighting process. As in the first embodiment described above, sampling was performed at 8 kHz, and 20
The audio signal divided into frames of about -30 ms is input from the input terminal 100 to the LPC analyzer 101 and the pre-filter 104 '.

In the LPC analyzer 101, 1 is set for each frame.
LPC analysis is performed on the order of 0 to 12, and L is calculated as the LPC coefficient.
It is output to the PC quantizer 102. As shown in the figure, this LPC analyzer 101 is connected to an LPC quantizer 102 which quantizes the LPC coefficient obtained from the input signal,
The PC quantizer 102 is further connected to an LPC dequantizer 103 that dequantizes the quantized LPC coefficient and a multiplexer 119 that performs a multiplexing process as an output stage of the present apparatus.

On the other hand, the pre-filter 104 'is the above-mentioned LP.
It is connected to the C inverse quantizer 103, and is further connected to the quiescent filter 106 that outputs a reference signal. The LPC inverse quantizer 103 uses the pre-filter 10
4 and the zero-input filter 106 as well as the weighting synthesis filter 107 for synthesizing weighted synthetic speech.

The LPC quantizer 102 has an input terminal 100.
After the LPC coefficient received from the LPC analyzer 101 via the is quantized, the bit configuration of the audio signal is reduced to the number of bits required for transfer, and then the LP
The C coefficient index ci is supplied to the multiplexer 119 and the LPC dequantizer 103. The LPC inverse quantizer 103 performs an inverse quantization process on the index ci of this LPC coefficient. Then, the dequantized LPC coefficient is supplied to the pre-filter 104 ′, the zero-input filter 106, and the weighting synthesis filter 107.

Simultaneously with the above-mentioned series of processing, the audio signal input from the input terminal 100 is a quarter of a frame.
Pre-filter 104 ′ for each sub-frame composed of
Performs a pre-filtering process to emphasize the "phonological" factor, so-called "formant", for each subframe, while at the same time creating an acoustically weighted input signal.

That is, one of the features that is different from the above-described first embodiment in the construction is that the pre-filter 104 'is directly connected to the zero-input filter 106. Therefore, the output of the pre-filter 104 'is input to the zero-input filter 106.

The detailed structure and functional operation of the pre-filter 104 'according to the second embodiment will be described with reference to FIG. As shown, this pre-filter 104 'includes a segmentation means 200, a formant enhancement means 201', and a gain control means 202.
It consists of The segmentation means 200 and the formant emphasizing means 201 ′ are respectively connected to the input terminal 1
00 in series. Further, the segmentation means 200 is connected to the gain control means 202, and the gain control means 202 is connected to the quiescent filter 106. Further, the formant emphasizing means 201 ′ is connected to the gain controlling means 202 and directly outputs to the quiescent filter 106.

In this pre-filter 104 ', the audio signal from the terminal 100 is segmented by the segmentation means 20.
0 and the formant emphasizing means 201 '. The voice signal is converted into voice / voice by the segmentation means 200.
It is divided into noise, and the determination information is divided into gain control means 202.
Given to. Here, in the formant emphasizing means 201 ′, both the formant emphasizing process and the auditory weighting process are performed. Further, the same LPC coefficient is applied to the same subframe as a processing unit, and the processing is performed at the same time. That is, the processing performed by the formant emphasizing means 201 ′ according to the second embodiment is achieved by inputting the input signal to the filter function represented by the following Expression 2.

[0067]

[Equation 2]

In Equation 2, α _i is the i-th order LPC coefficient, η is the weighting coefficient of the zero filter, and ν ′ is the weighting coefficient of the polar filter. The auditory weighting processing and the formant enhancement processing are combined. It is set as a coefficient value. The weighting coefficient η of the zero filter and the weighting coefficient ν ′ of the pole filter have a relationship of 0.0 <η <ν ′ <1.0. As a suitable example, for example, η = 0.2, ν ′
= 0.64.

Note that, with normal auditory weighting, η = 1,
A value of ν = 0.8 is used. Further, in the formant enhancement for noise suppression, a value of η = 0.2 and ν = 0.8 is preferable, so that both effects are combined,
Formant emphasizing means 201 'for which such a value is set
At the same time, the formant is emphasized for each subframe, and at the same time, the acoustically weighted input signal is supplied to the gain control means 202, and the gain is controlled based on the voice / noise determination information, and then zero. It is supplied to the input filter 106.

The speech / noise determination in the segmentation means 200 is made based on a preset threshold value. This threshold value is set in advance to an appropriate value according to a predetermined representative environmental condition, and is stored in a predetermined storage means (not shown). In the encoding process, the user of the device of the present invention selects various environmental conditions at the time of use by using the manual switch, and the optimum threshold value under the conditions is read from the storage means.
Segmentation processing is performed.

As another processing method, the background noise level (in the input audio signal) is measured in conjunction with the operation switch attached to the device, and a predetermined threshold value corresponding to this level value is automatically selected. It may be operated as described above.

Referring again to FIG. 2, the prefilter 1
The audio signal that has been subjected to the pre-filtering processing in 04 'is input to the subtractor 120 as a reference signal when the signal is input to the zero input filter 106 and the zero state response is obtained. On the other hand, the process of generating the combined signal is the same as that described in the first embodiment.

As described above, in the second embodiment, the formant enhancement process and the perceptual weighting process are performed by the pre-filter 10.
4 ', the two functions in the above-described first embodiment are shared, so that they are collectively performed. Therefore, the arithmetic processing for one filter is substantially performed. Does not have to be performed, and as a result, it contributes to the reduction of the amount of calculation related to a series of processes.

(Modified Embodiment) The first embodiment of the present invention has been described above.
Although the second embodiment has been described, here, for example, the processing operation of the segmentation means 200 and the gain control means 202 of the second embodiment described above is performed instead of the same processing operation as that of the first embodiment described above. You may.

In the segmentation processing algorithm of the first embodiment, the threshold value Eth is set as a fixed value, but the threshold value may be relatively changed according to the signal level.

As an embodiment relating to the speech coding apparatus of the present invention, various modifications can be made within the scope of the gist of the present invention. The following inventions are included in this specification.

(1) "Analysis-by-Synt" analysis with a reference signal, which is a signal with suppressed noise gain
hesis (hereinafter, abbreviated as "A-B-S") in a speech coder that performs coding by means of a method for enhancing the formant of a reference signal, and a segmentation process for dividing a speech part and a noise part into a plurality of segments. And a means for suppressing the level of the noise part obtained by the segmentation means. Action 1: The input signal is segmented by voice / noise for each subframe, the input signal is formant-emphasized, the amplitude of the noise part is suppressed, and the noise-suppressed signal is used by the A-B-S method. Use as reference signal. Effect 1: Simple processing is applied, and noise can be suppressed for each subframe by real-time processing.

(2) It is characterized in that the same filter coefficient as the filter used in the means for enhancing the formant of the reference signal and the filter coefficient of the weighting filter used in the A-b-S method are used (1). The audio encoding device according to. Action 2: A linear prediction coefficient is extracted from the input signal for each frame, formant enhancement filter processing is performed for each subframe based on the linear prediction coefficient, and the input signal is segmented by speech / noise to generate a noise part. The signal whose noise has been suppressed is used as the reference signal used in the A-B-S method. Effect 2: Since the same linear prediction coefficient (LPC) is used, the hardware configuration of the device is simple.

(3) "Analysis-by-Synt" with a reference signal, which is a signal with suppressed noise gain
hesis (hereinafter, abbreviated as "A-B-S") in a speech coding apparatus for coding, a unit for emphasizing a formant of the reference signal and a unit for performing an auditory weighting filter unit, A speech coding apparatus comprising: a segmentation unit that performs a segmentation process that divides a noise unit and a noise unit into a plurality of segments; and a unit that suppresses the level of the noise unit obtained by the segmentation unit. Action 3: A linear prediction coefficient is extracted from the input signal for each frame, and a process in which formant enhancement filter processing and auditory weighting filter processing are integrated is performed for each subframe based on the linear prediction coefficient, and the input signal is voiced. / Segmentation processing is performed by noise to suppress the amplitude of the noise part, and the noise-suppressed signal is used as the reference signal used in the A-B-S method. Effect 3: The processing can be further simplified and the same effect as described in (1) to (2) can be expected.

(4) A fade-in / fade-out section is set as a predetermined area before and after the section determined to be the voice part by the segmentation means, which gradually suppresses the amplitude of the background noise voice. Do
The speech coding apparatus according to any one of (1) to (3). Action 4: When suppressing the amplitude of the noise part of the input signal, the gain is gradually increased or decreased in units of subframes to smooth the connection with the audio part and used as the reference signal. Effect 4: The noise effect of the voice part is made inconspicuous, and the effect is suppressed even when the background noise is generated in impulse.

(5) The segmentation means is
It is characterized by applying a predetermined threshold (1) to (
4) The voice encoding device according to 4). Action 5: When the energy or amplitude of the input signal is below a certain threshold, it is determined to be a noise part, and when it exceeds the threshold, it is determined to be a voice part. Effect 5: The voice part can be detected by a simple process.

(6) The segmentation means is
The method further comprises means for updating the threshold value applied to the segmentation process when the energy or amplitude of the reference signal is equal to or lower than a first threshold value having a fixed value and the rate of change is equal to or lower than a certain level. The speech coding apparatus according to (5). Action 6: If the energy or amplitude of the input signal is below a certain fixed threshold and the rate of change is within a certain range, the threshold is changed, and when the rate of change exceeds a certain range, the voice unit is activated. It is determined that the rising edge of. Effect 6: Even if the level of the background noise varies with time, the rising edge of the voice part can be detected.

(7) The segmentation means is
When the state below the first threshold value, in which the energy or amplitude of the reference signal has a fixed value, is maintained over a plurality of frames, it is provided with means for updating the threshold value applied to the segmentation processing. Do (
The speech coding apparatus according to 5). Action 7: When the energy or amplitude of the input signal is continuously lower than a certain fixed threshold value by a predetermined number of subframes, the threshold value is changed and it is determined as a noise part. Effect 7: Even when the level of the background noise is temporally varying, the fall of the voice part can be detected.

(8) The segmentation means is
The speech coding apparatus according to (6) to (7), characterized in that processing is performed by means for updating the threshold value. Action 8: If the energy or amplitude of the input signal is below a certain fixed threshold and the rate of change is within a certain range, the threshold is changed, and if the rate of change exceeds a certain range It is determined to be a rising edge, and when the energy or amplitude of the input signal is continuously lower than a certain fixed threshold value by a predetermined number of subframes, the threshold value is changed and it is determined to be a noise part. Effect 8: The voice part can be detected even when the level of the background noise varies with time.

[0085]

As described above, according to the speech coding apparatus of the present invention, the following effects can be obtained. It is possible to suppress the noise part in the input signal for each sub-frame by real-time processing by using a simpler arithmetic processing than the conventional one. Further, the noise sense of the voice part of the input signal is made inconspicuous, and at the same time, it has an effect of suppressing the generation of impulsive background noise. As a result, good voice quality can be obtained even when used in a noisy and poor environment.

[Brief description of drawings]

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

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

FIG. 3 is a block diagram showing the configuration of a prefilter according to the first embodiment.

FIG. 4 is a block diagram showing the configuration of a prefilter according to a second embodiment.

FIG. 5 is a flowchart showing a processing routine of gain control (of a prefilter) according to the first embodiment.

FIG. 6 is a flowchart showing a segmentation processing routine (of segmentation means) according to the first and second embodiments.

[Explanation of symbols]

100 ... Input terminal, 101 ... LPC analyzer, 102 ... L
PC quantizer, 103 ... LPC inverse quantizer, 104 ... Pre-filter, 105 ... Auditory weighting filter, 106 ...
Zero input filter, 107 ... Weighted synthesis filter, 10
8 ... Evaluator, 109 ... Adaptive codebook, 110 ... Stochastic codebook, 111 ... Shift register, 114 ... Adaptive gain codebook, 115 ... Stochastic gain codebook, 119 ... Multiplexer, 200 ... Segmentation means, 201 ... Formant Emphasis means, gain control means 20
2, S301 to S307 ... Gain control routine, S40
0-S426 ... Segmentation routine.

Claims (3)

[Claims] 1. A speech coding apparatus for coding by a "analysis method by combination" with a reference signal, which is a signal in which a gain of noise is suppressed, a means for emphasizing a formant of the reference signal, A speech coding apparatus comprising: a segmentation unit that performs a segmentation process that divides a noise unit and a noise unit into a plurality of segments; and a unit that suppresses the level of the noise unit obtained by the segmentation unit. 2. A speech coding apparatus for coding by a "analysis method by synthesis" with a reference signal, which is a signal in which noise gain is suppressed, in a speech coding apparatus, means for emphasizing a formant of the reference signal, and auditory weighting. Means for performing an integration process with the filter means, segmentation means for performing a segmentation process for dividing the speech part and the noise part into a plurality of segments, and means for suppressing the level of the noise part obtained by the segmentation means, A speech coding apparatus comprising: 3. A fade-in area as a predetermined area that gradually suppresses the amplitude of the background noise sound before and after a section determined to be a sound part by the segmentation means.
The speech coding apparatus according to claim 1, wherein a fade-out section is set.