JP2002169599A - Noise suppression method and electronic device - Google Patents

Abstract

(57) [Problem] To provide a noise suppression method capable of good noise suppression. A noise suppression method for dividing an input audio signal into frames in a predetermined time unit, dividing the divided frames into predetermined frequency bands, and performing noise suppression processing for each of the divided bands. In the band-specific gain determining step, the band-specific gain value when the determination target frame is determined to be a voice frame is smaller than the band-specific gain value when the determination target frame is determined to be a noise frame. A noise suppression method is characterized in that a gain value for each band is set so as to take a value.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for suppressing noise in electronic equipment such as a portable telephone which employs voice coding.

[0002]

2. Description of the Related Art In a device for performing voice communication such as a mobile phone,
A speech coding method such as a CELP (Code Excited Linear Prediction) method is used.

When such a device is used in an environment with a large amount of background noise, the background noise is taken in and coded, and the intelligibility of speech is reduced. Therefore, a technology (noise canceller) for removing or suppressing background noise and making the signal close to a signal of only sound and performing sound coding has been studied.

[0004]

Japanese Patent No. 2995737 discloses that an input signal is divided into frames of a predetermined period, noise / speech is determined for each frame, and a gain for each band is determined when the frame is determined to be a noise frame. A method is disclosed in which the value is set to a minimum, and when it is determined that the frame is an audio frame, a gain value exceeding the value is set to perform noise suppression.

By adopting this method, it is possible to obtain a noise-suppressed audio signal, but it is still insufficient in auditory sense. In other words, noise suppression is performed by judging voice / noise for each frame. However, there are bands in which voice components are included and bands which are not included in voice frames, and bands in which voice components are not included in voice frames. It is considered that the noise suppression is insufficient.

[0006] The present invention has been made in consideration of this problem, and an object of the present invention is to provide a noise suppression method capable of performing good noise reduction.

[0007] That is, an object of the present invention is to provide a noise suppression method which improves a noise suppression method in a voice frame and enables good noise reduction as a whole.

[0008]

The inventor of the present invention not only determines a gain value for each band when noise suppression is performed by distinguishing between a speech frame and a noise frame, but also determines a gain value for each band of a noise frame based on the minimum gain value. Also, it has been found that by setting the minimum gain value for each band of the audio frame to be smaller, the audibility of the audio signal after noise suppression is improved.

[0009] Even in a frame determined to be an audio frame, not all bands include audio components. The present inventors pay attention to this point, and regarding a band in which a speech component in a frame determined to be a speech frame is not included (or a speech component is small), the gain is smaller than a band-specific gain value of a noise frame. It has been found that by setting the gain and making the band in which the audio component is included in the audio frame stand out, a good audibility can be obtained.

That is, according to the present invention, an input signal is divided into frames in a predetermined time unit, the divided frames are divided into predetermined frequency bands, and noise is suppressed for each of the divided bands. In the suppression method, an audio frame determining step of determining whether the frame is a noise frame or an audio frame; and a band-specific gain determination for setting a band-specific gain value of each frame based on a result of the audio frame determining step. And a signal generation step of performing noise suppression for each band using the band-specific gain value determined in the band gain determination step, and then reconstructing a frame to generate a noise-suppressed output signal. In the band-specific gain determining step, the band-specific gain when it is determined that the frame to be determined is an audio frame. The noise suppression method is characterized in that the gain value for each band is set so that the gain value can be smaller than the gain value for each band when the frame to be determined is a noise frame. .

In other words, the gain value for each band of the band in which it is estimated that the voice component is not included in the frame determined to be the frame to be determined as the voice frame is determined, and the frame to be determined is determined to be the noise frame. By setting the gain value for each band so as to take a value smaller than the gain value for each band, the band including the audio component in the audio frame can be made more prominent, and as a result, noise suppression with good audibility can be achieved. The obtained output signal can be obtained.

In the noise frame, a constant gain value may be set for each band, or the gain value may be changed based on the difference between the power for each band and the noise power.

[0013] Further, with respect to the voice frame, the gain value for each band is set so as to increase as the index based on the difference between the power for each band and the noise power increases. Is also possible. A function that decreases continuously may be employed.

In the noise suppression method according to the present invention, there is a step of updating the estimated value of the noise power as a step before determining the above-mentioned gain for each band. The noise estimation value is updated under a predetermined condition.
The updating method disclosed in the noise suppression method disclosed in No. 3030 can be adopted.

This updating method uses an SNR (log value of signal energy / noise energy) for each band of each frame.
The voice metric, which is the sum of the weighted values, is used, and the deviation for each band (log value of signal energy−log value of average value of past signal energy)
Is a technique for updating a noise estimation value using a spectrum deviation that is a sum of absolute values of the noises. If the spectrum deviation falls below a threshold for a predetermined time (for example, one second), the estimated noise value is updated. You.

Further, instead of using the value of the spectral deviation as it is for the determination, the total deviation of the difference between the band power and the noise power between the previous frame and the past frame is normalized by the average value, and this normalized value is also calculated. By determining the noise section at this time, a method can be adopted in which noise having a large inter-frame variation compared to the above method can be recognized as noise.

That is, a significant value for each band in which a difference between the power for each band and the estimated noise power value for each band is weighted in a predetermined manner.
(suby) The sum of the difference between the current frame and the previous frame (s
um) is the ratio (r) normalized by the average value (sum_average)
Is a method for determining whether or not the current frame is a noise frame based on

Thus, the significant value for each band with respect to the past frame
By using the deviation of (suby) and normalizing the deviation with the average of the total deviations as the basis for the determination, it is possible to mitigate the variation for each frame. Can be. Therefore, it is possible to satisfactorily recognize noise having large inter-frame variations as noise.

More specifically, a frame dividing step of dividing a transmission input signal into frames in a predetermined time unit; a frequency band dividing step of dividing each frame into a plurality of frequency bands; Calculating the power for each band (channel_power); calculating the difference (tmp) between the noise power estimation value for each band (noise_power) and the power for each band (chennel_power) for each frequency band; (tmp)
A significant value calculating step of calculating a significant value (y) obtained by adding a significant value (suby) for each band obtained by performing a predetermined weighting on a predetermined condition; and between a current frame and a previous frame, Calculating the sum of the absolute values (sum) of the differences of the significant values (suby) of the respective bands for the frequency band of; and calculating the average value (sum_average) of the sum of the absolute values (sum); A significant value normalizing step of calculating a ratio (r) obtained by normalizing the sum of values (sum) by the average value (sum_average) of the sum of absolute values.

Updating the noise power estimate has the following two steps.

That is, when the significant value (y) falls below a predetermined threshold value, the current frame is determined to be a noise frame, and the first noise power estimation value (noise_power) for updating the noise power estimation value for each band (noise_power) is updated. Updating the value; the ratio
When (r) continuously falls below a predetermined threshold for a predetermined period, the current frame is determined to be a noise frame, and the second noise power estimation value for updating the noise power estimation value for each band (noise_power) Update step.

The first noise power estimation value updating step is a case where noise estimation is performed satisfactorily and it is determined that the frame is a noise frame by a significant value determination. The second noise power estimation value updating step includes: Even if a significant value varies between frames and a good noise value cannot be determined with a significant value, forced updating can be performed.

As the average value used for normalization, an estimated value obtained by using leak integration of the sum of absolute values (sum) can be used. Further, it is also possible to use an estimated value of the average value (sum_average) obtained by using leak integration of a standard deviation of the sum of absolute values (sum).

When setting the gain for each band, the gain value for each band is determined by using a different function between the case of a voice frame and the case of a noise frame. Basically, it is calculated based on the difference between the power for each band and the noise power for each band (difference in logarithm: SNR). That is, a band having a large SNR in a voice frame is presumed to be a band containing a voice component.
The gain value of the band is set to a large value, the band having a small SNR is estimated to contain no voice component, and the gain value is set to a small value.

Incidentally, noise (Background Noise) is generally assumed to be stationary, but may fluctuate outdoors. In particular, the energy of noise generated when a vehicle passes by increases as the vehicle approaches. If the transmitted voice is input in this state, the suppressed voice may be distorted because the energy difference between the voice and the noise is small. Also, when the spectral shape of the noise is similar to the spectral shape of the voice, the suppression of the noise based on the noise energy makes it easier to interfere with the voice spectrum, so that the suppressed voice is distorted. Even if the noise energy fluctuates, S
It is also possible to suppress such an effect by adjusting the variable for determining the gain value while using the NR as a basis.

In the adjustment, a noise power estimating step of obtaining a signal power for each frequency band and estimating a noise power for each band based on the band power when determining the gain value for each band; A minimum value detecting step of detecting a minimum value of power over at least one of the band power and the band-specific noise power over a plurality of frame periods; and detecting the band power and the minimum value detecting step for each of the frequency bands. This can be performed by determining the noise suppression amount for each frequency band based on the difference obtained from the band-specific minimum value determining step for obtaining the difference from the obtained band-specific minimum value.

Further, using an adjustment value for generating an adjustment value common to different bands for each frame, a difference between a value obtained by adding the minimum value for each band and the adjustment value and the band power is obtained for each frequency band. It is also possible to determine the amount of noise suppression for each frequency band based on this difference.

In the noise section, an adjustment value common to the bands is determined based on a difference between the average value between the minimum values for the respective bands and the average value between the noise powers for the respective bands; Can be obtained by determining an adjustment value common to bands based on a difference between a minimum value among a plurality of band powers in one frame and a maximum value among a plurality of band-specific minimum values.

Note that, in the determination of a voice frame and a noise frame, a noise power estimation step of obtaining signal power for each frequency band and estimating noise power for each band based on the band power; Comparing the difference between the noise power and the band power for each band and comparing the difference for each band with a predetermined threshold value; a plurality of adjacent bands among the differences for each band arranged in frequency order The difference between the bands is determined to exceed the threshold value, the band-by-band difference is weighted by a predetermined weight, and then added to each other; And a determining step of determining whether the input signal is a voice section or a noise section based on the added value.

In the adding step, the difference for each band can be weighted so that the weight decreases as the frequency increases. In the determining step, the voice of the input signal is determined based on the added value. Section or
It is possible to determine whether it is a noise section or a transition section which is an intermediate area between both sections.

The present invention as described above is based on ACELP, EVR
The present invention can be applied to an electronic device such as a mobile phone adopting a digital voice coding method using various voice coding methods such as C, EFR, and AMR. That is, in an electronic device including an audio signal input unit (a direct input unit such as a microphone or a signal transmitted from an electronic file or the like) and an audio encoding unit, the audio signal of the audio signal input unit is received, An electronic apparatus comprising: a noise suppression unit that supplies a signal whose noise has been suppressed by the noise suppression method to a speech encoding unit.

The noise suppression unit can be executed by signal processing in the DSP, for example, in the same way as speech coding.

[0033]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The noise suppression method of the present invention can be used for general equipment that handles digital audio input. The case of a mobile phone will be described as an example.

FIG. 1 is a circuit block diagram of a digital portable telephone device equipped with the noise suppression method of the present invention.

In FIG. 1, a radio carrier signal transmitted from a base station (not shown) via a radio channel is received by an antenna 1 and then received by a receiving circuit (RX) 3 via an antenna duplexer (DUP) 2. And is mixed with the reception local oscillation signal output from the frequency synthesizer (SYN) 4 and frequency-converted to an intermediate wave signal. This received intermediate wave signal is sampled by an A / D converter (not shown) and then input to a digital demodulator (DEM) 6.

The digital demodulator 6 performs digital demodulation processing after establishing frame synchronization and bit synchronization with respect to the digital reception intermediate wave signal. The baseband digital demodulated signal obtained by this demodulation processing is input to a time division multiple access circuit (TDMA) 8, where a time slot addressed to itself is separated and extracted for each transmission frame. The information on the frame synchronization and the bit synchronization obtained in the digital demodulator 6 is input to the control circuit 18.

The digital demodulated signal output from the TDMA circuit 8 is subsequently sent to an error correction code composite circuit (CH-CO
D) It is input to 9 and subjected to error correction decoding processing. The corrected and decoded digital demodulated signal is supplied to a speech synthesizer (DE
C) The signal is input to 10 and subjected to voice decoding processing to reproduce a digital reception signal. This digital reception signal is D / A
After being converted back to an analog reception signal by the converter 11, the signal is supplied to a speaker 12 via an audio amplifier (not shown) and output as a loudspeaker.

On the other hand, the transmitted voice of the speaker is the microphone 1
A / D converter 1 after the sound is collected at 3 and converted into an electric signal
4 and is sampled at a predetermined sampling period and converted into a digital transmission signal. The digital transmission signal is passed through a noise canceller 17 to be described later, and then input to a voice coding circuit (COD) 16 where voice coding is performed.

The coded voice data output from the voice coding circuit 16 is input to an error correction code decoding circuit (CH-COD) 9 together with the control signal output from the control circuit 18 and is subjected to error correction coding. The digital transmission signal subjected to the error correction coding is input to the TDMA circuit 8. This T
In the DMA circuit 8, a transmission frame corresponding to the time division multiple access (TDMA) system is generated, and a process for inserting the digital transmission signal into a time slot allocated to the own device in the transmission frame is performed. Done. TD
The digital transmission signal output from the MA circuit 8 is input to a digital modulator (MOD) 7.

The digital modulator 7 generates a transmission intermediate wave signal digitally modulated by the digital transmission signal, converts the signal into an analog signal by a D / A converter (not shown), and inputs the analog signal to the transmission circuit (TX) 5. .

In the transmission circuit 5, the modulated transmission intermediate wave signal is mixed with the transmission local transmission signal output from the frequency synthesizer 4 and converted into a radio carrier frequency corresponding to a communication channel. The radio modulated wave signal is transmitted from the antenna 1 to the base station (not shown) via the antenna duplexer 2 after being controlled to a predetermined transmission power level indicated by the control signal TCS from the control circuit 18 in the transmission power amplifier. Is done.

The operation panel section 19 includes a key input section having a transmission key, an end key, a dial key, and various function keys, a liquid crystal display (LCD) and a light emitting diode (LE).
D) is provided.

The noise canceller 17 is, for example, a DSP (Di
The processing program is stored in a memory in the noise canceller or a memory attached to the control circuit 18. FIG. 2 is a block diagram showing a functional configuration realized by the processing program.

The digital audio signal output from the A / D converter 14 is first input to the frame division unit 21.
The frame division unit outputs, for example, a frame arranged into 128 samples (frame division step). At this time, the digital transmission signal may be divided into, for example, frames of 80 samples, and then the frame ends may be overlapped by windowing. This digital transmission signal frame is converted into a fast Fourier transform unit (FFT) 2
Enter 2

The FFT 22 performs a fast Fourier transform process on the input digital transmission signal frame, and sequentially performs 16 bands (k = 0, 1, 2,... 1) from a low band to a high band.
5) The frequency-divided transform coefficients are obtained. This transform coefficient need not be the same in each band. The band-divided transform coefficients are output to the band power calculator 26 (frequency band dividing step). <Band power calculation> The band power calculation unit 26 obtains the energy (mean square value of the conversion coefficient) for each band, takes the logarithm, and calculates the band power channel_power (m, k), [m is the frame number, and k is the band. No. (0 to 15)] (power calculation step for each band). This band power is output to the significant value calculation unit 27. <Significant value calculation> In the significant value calculating unit 27, the noise leak integrated value output from the noise leak integrated value updating unit 32 described later
noise_power (m, k) and the above-mentioned band power chanel_power
The difference tmp from (m, k) is obtained, and the difference tmp for each band is compared with a predetermined threshold value. Differences by band above arranged in frequency order
When it is determined that the difference tmp by band of a plurality of adjacent bands among the tmps exceeds the threshold value, the difference tmp by band is given a predetermined weight and added to each other. The conditional sum of the weighted values suby (m, k) (when it is determined that the difference tmp by band of a plurality of adjacent bands exceeds the threshold) is output as a significant value y (significant value calculating step) ).

An average value of the significant value y (y_average: an estimated value obtained by leak integration, for example, calculated by the following formula) is also output.

Y (m): significant value, conditional sum of suby (m, k) y_average (m) = y # average (m-1) × 0.9 + y (m) × 0.1 FIG. 6 is a flowchart showing the processing procedure of FIG. A flow for outputting the significant value y will be described with reference to FIG.

After resetting / initializing the frame number m to 0 in step 3a, the group number m is incremented and the significant value y and the band number k are incremented in step 3b.
And the continuous number flag (continuous number flag of the difference tmp for each band exceeding the threshold value) is initialized to “0”.

Next, in step 3c, for band k = 0,
The difference tmp between the band power and the noise leak integral value and the value suby (m, k) obtained by weighting the difference tmp for each band are calculated as follows.

Tmp = chanel_power (m, k) −noise_power (m, k) sub_y (m, k) = {200− (k−1) ² } / 100 × (tmp−1) where {200− (k−1) -1) ² ｝ is a weighting coefficient. In this case, the band is set to decrease as the frequency of the band increases, but can be changed as appropriate.

When the difference tmp for each band at the band k = 0 is calculated, the significant value calculation unit 27 determines the difference tm for each band at step 3d.
Compare p to a threshold value (eg, 1). If the threshold is exceeded, it is determined that there is a possibility that the sound is voice, and step 3
e, the process proceeds to step 3i via step 3g,
Set flg to 1. Next, after incrementing the band number k to k = 1 in step 3k, the process returns to step c to execute the same processing for the band k = 1.

Here, it is also assumed that the band difference tmp exceeds the threshold value after the band k = 0 even in the band k = 1. Since the continuous number flg is already 1, step 3e to step 3f
Then, the operation of y = y + suby (m, k−1) is executed. And set the continuous number flg to 2,
After step 3g, the process proceeds to step 3h to execute the following calculation.

Y = y + suby (m, k) Then, in step 3k, the band number k is further incremented to set k = 2, and the process returns to step 3c to execute the processing for band k = 2.

Thereafter, similarly, the difference tmp by band between adjacent bands
Each time exceeds the threshold, the suby (m,
k) is sequentially added to the significant value y obtained up to the immediately preceding band, and a weighted added value y of the band-based difference tmp is obtained.

When the difference tmp for each band becomes equal to or less than the threshold value in any band k = i, the significant value calculation unit 27
Moves from step 3d to step 3j, and the number of continuous flg
Is reset to 0.

When the processing is completed for all 16 bands (k = 0 to 15) constituting one frame,
The significant value calculation unit 27 proceeds from step 3m to step 3n, and outputs the significant value y and the weighted difference suby (m, k) (k = 0 to 15) calculated for each band after weighting.

In this way, a significant value y is obtained for each frame, and is used to determine whether the frame is a speech frame or a noise frame.

The significant value calculation unit 27 also counts a significant section for determining the noise power forced update. This processing will be described with reference to the flowchart of FIG.

First, an average value y_average (m) of the significant values y (m) is obtained.

In step 4a, frame number m = 0, sum
_Average (0) = 0.1, y_average (0) = 10, counter (0) = 0
, The group number m is incremented and a significant value y, sub (m, k) is input in step 4b.

Then, in step 4c, the average value of the significant value y is calculated. The average value can be set to an appropriate period based on the relationship between the memory capacity and the calculation amount (for example, 0.1 to
It is enough to take an average of about 0.3 seconds.
For example, an average is obtained by adding 0 frames), and in general, estimation is calculated as follows using leak integration. Needless to say, a method other than leak integration may be used for obtaining the average value.

Y_average (m) = y_average (m−1) × 0.9 + y (m) × 0.1 Next, in step 4d, the sum sum of absolute values of the difference between sub (m, k) and sub (m−1, k) Is calculated (step of calculating the sum of significant values for each band), and in step 4e, the ratio r is calculated by dividing the sum by the average value sum_average of the absolute sum sum (significant value normalization step).

Sum (m) / sum_average (m-1) This value may be directly used as r. However, in order to remove a specific value, the attenuation rate determined by r (m-1) (for example, 0.99 )
Is compared with the value obtained by multiplying the value, the larger value is adopted as r (m).

The ratio r serves as a criterion for adding a counter for calculating a significant value section. For example, the upper limit is set to 8. Therefore, if it is determined in step 4f that r (m) exceeds 8, r (m) = 8 is reset in step 4g.

Next, sum_average is updated in step 4h. The average value can also be set to an appropriate period based on the relationship between the memory capacity and the calculation amount (for example, 0.1 to
It is enough to take an average of about 0.3 seconds.
In general, it can be estimated and calculated as described below using leak integration. Needless to say, a method other than leak integration may be used for obtaining the average value.

Sum_average (m) = sum_average (m−1) ×
0.9 + sum (y) × 0.1 Note that sum_average may use an estimated value of the standard deviation.
Also in this case, an estimated value can be obtained by using the leak integral of the following equation, and this value is used instead.

Sum_average (m) = sqrt (sum_average (m−
1) ² × 0.9 + sum (m) ² × 0.1) Subsequently, the counter (m) of the significant section is calculated.

Y> 10 and counter (m) <100 and r (m)
When ≤ THR, 1 is added to counter (m). If this condition is not satisfied, counter (m) is reset to 0.

THR may be a fixed value, and y_avera
It can be changed by ge. In the present embodiment, the THR that changes according to the following equation is adopted.

THR = 1.7 + (y_average−40) / 200 where 1.7 ≦ THR ≦ 2.0 y_average> 100 THR = 2.0 y_average ≦ 40 THR = 1.7 40 ≦ y_average ≦ 100 THR = 1.7 + (y_average−40) / 200 If it is determined in 4i that y_average (m) exceeds 100, THR = 2.0 is set in step 4j. If it is determined in step 4k that y_average (m) exceeds 40, THR is set in step 4l. Is set to the variable value of the above equation. In other cases, THR is set to 1.7 in step 4m.

At step 4n, it is determined that the significant value y exceeds 10, and at step 4o, the counter counter is set to 10
If it is determined that the value is less than 0 and the ratio r is equal to or less than THR in step 4p, the counter counter is determined in step 4q.
Is added, otherwise the counter counter is reset to 0 in step 4r.

Similarly, if it is determined in step 4n that the significant value y is 10 or less, in step 4s the counter counter (m)
Is reset to 0, and counter is 100 in step 4o.
In the case of above (ie, 100), count 4
r (m) = counter (m−1).

With the above processing, for each frame m, coun
ter (m) and y_average (m) are output (step 4u). <Update judgment> These outputs (counter (m), suby (m, k), y
(m), y_average (m)), the update determination unit 31 determines whether or not the noise power value noise_power (m, k) for each band has been updated.
The noise leakage integrated value updating unit 32 updates the noise power value for each band.

The significant value y is 20 to 30 for normal voice.
And if noise estimation is well performed, y <
It will be about 15. Therefore, when y <15, for example, the following equation is used (first noise power estimation value updating step). noise_power (m + 1, k) = noise_power (m, k) × 0.9 + chan
nel_power (m, k) × 0.1 k = 0,1,..., 15 IS127 [Variable-rate speech coding system of US standard: “Enhanced VariableRate Codec, Speech Service O
ption 3 for Wideband Spread Spectrum Digital Syste
ms "(TIA IS127)], a normal noise power update may be performed.

If y is not accurately calculated for some reason, forced updating is performed based on the counter value (counter) (second noise power estimation value updating step). For example, counter (m) ≧ 100 and y <y_average (m) +5
At the time of updating, it is updated according to the above equation.

Subsequently, the gain for each band is determined by the gain determining unit for each band 33. At this time, the value is set for each band with reference to the significant value (y), the significant value (suby) for each band, and the like calculated by the significant value calculation unit. <Speech Weight Calculation> Upon receiving the significant value y output from the significant value calculation unit, the speech weight calculation unit calculates the speech weight sp used for determining the noise suppression gain. The voice weight sp is defined as a degree of voice included in one frame, 0 ≦ s.
It is a numerical value expressed in the range of p ≦ 6, where sp = 0 is a noise section,
sp = 6 represents a voice section. It should be noted that the numerical value, the step division, and the like can be set as appropriate.

FIG. 5 is a flowchart showing the procedure for calculating the speech weight sp in the speech weight calculator 28 and the processing contents.

First, after resetting the frame number m to 0 in step 5a, the group number m is incremented in step 5b. Next, in step 5c, the weighted addition value y is compared with an arbitrary threshold value “13”, and y <13
If so, it is determined that the frame is a noise frame, and the process proceeds to step 5d, where the speech weight sp (m) is set to sp (m) = sp (m-1) -0.5. As will be described later, sp (m) is set in the final step so that sp (m) = MAX (sp (m), 0)) and the minimum value becomes 0. Therefore, if noise frames are continuous, sp (m) Converges to zero.

On the other hand, if y ≧ 13, the frame is in the voice or transition period, and the process proceeds to step 5e to calculate the temporary voice weight zz = (y−13) × 1.5 + 1.

First, it is determined in step 5f that sp (m-1) ≦ 0.5. That is, the speech weight sp one frame before.
It is determined whether (m-1) is sufficiently smaller than 0.5 or less, and it is determined whether the frame is a noise frame.

When the previous frame is determined to be a noise frame, that is, when sp (m-1) is 0.5
In the following cases, the process proceeds to step 5g, where the voice weight sp (m) of the current frame is set to the temporary voice weight z.

This case is at the time of switching from noise to speech, and it is necessary to suppress the noise and raise the speech clearly so that the beginning of the word is not cut off. Therefore, the audio weight is set to take a large value.

On the other hand, the speech weight sp of the previous frame is
When (m-1) is not a noise frame (sp (m
In -1)> 0.5), the process proceeds to step 5h, where z> sp (m-1) +0.5 is determined. If z is greater than (sp (m-1) +0.5), in step 5i the speech weight sp (m) of the current frame
Is set to (sp (m-1) +0.5).

This case is a point in time when it is determined that the voice frame is in a transitional period. The importance is placed on continuity, and the rise of z from the previous frame is suppressed to 0.5.

On the other hand, if z is equal to or smaller than (sp (m-1) +0.5), the flow shifts to step 5j, where z <sp (m-1) -0.5 is determined, and z is set to (sp (m) If it is smaller than -1) -0.5), the voice weight sp (m) of the current frame is set to (sp (m-1) -0.5) in step 5k.

This case is also a point in time when it is determined that the speech frame is in a transition period, and the continuity is emphasized, and the drop of z from the previous frame is suppressed to 0.5.

If z is equal to or smaller than (sp (m-1) -0.5), the flow shifts to step 5m to set the speech weight sp (m) = z of the current frame.

Through the above steps, sp (m) = z,
sp (m-1) ± 0.5 is set to one of three values, and finally sp (m) = MIN (sp (m), 6) sp (m) = MAX (sp (m) , 0) determines the value of sp (m) = 0 to 6.

That is, steps 5f to 5
m, the temporary voice weight z calculated in the current frame is
It is corrected in consideration of the voice weight sp (m-1) set in the immediately preceding frame, output as sp (m) in step 5n, returns to step 5b, and returns sp to all m.
(M) is required.

By using the speech weight sp (m) obtained in this way, it is possible to adjust the speech / noise / transient region in consideration of the continuity between frames.

The speech weight sp (m) obtained by the speech weight calculator 28 is input to the noise minimum value estimator 29 and the band-specific gain determiner 33. <Estimation of Noise Minimum Value> The noise minimum value estimation unit 29 calculates the leak integrated value noise_power of noise in each band every 100 frames in which the speech weight sp becomes sp = 0.
Check the minimum value of (m, k). Then, this minimum value is calculated by the following 1
In the period of 00 frames, the noise minimum value noise_min
Used as (m, k). At the same time, an inter-band average value min_all of the noise minimum value of each band is obtained.

FIGS. 6 and 7 show the noise minimum value estimating unit 2.
9 is a flowchart showing the procedure and contents of a minimum value estimation process executed in FIG.

In the figure, the noise minimum value estimating unit 29 first resets the frame number m to m = 0 in step 6a, sets the frame counter value to fc = 96, and sets the noise minimum value to noise_min (k) = 36, the band is set to k = 0,.
.. Initially set to 15, respectively.

Further, noise_min_h (k) = MAX (noise_power (m, 2k), noise_p
ower (m, 2k + 1)), k = 0,..., 7 The noise-to-band average min_all of the noise minimum is noise_min_h (n):
Initially, min_all = Σnoise_min_h (n) / 8 where n = 0 to 7, which is the average of the sum of the values of n = 0, 1,..., 7.

That is, a value having a large noise power in an adjacent band is taken, and the average value is set as min_all.

Next, the minimum noise value estimating section 29 determines in step 6
After incrementing the frame number m in b, it is determined in step 6c whether the speech weight is sp = 0, that is, whether the frame is a noise frame.

If the frame is a noise frame, the flow returns to step 6b, where the frame number m is incremented, and the noise frame is determined in step 6c. sp =
When it is determined that the value is not 0, that is, when a voice frame or a transition area frame is detected, the minimum noise value estimating unit 29 proceeds to step 6d, where the frame counter fc is incremented and the band k = 0. Select

Then, after setting x = MAX (noise_power (m, 2k), noise_power (m, 2k + 1)) in step 6e, the process proceeds to step 6f to determine whether noise_min_h (k)> x. judge.

If Noise_min_h (k)> x, step 6
g, where the minimum noise value is noise_min_h (k) = x
Set to. Then, control goes to a step 6h.

On the other hand, if noise_min_h (k) ≦ x, the process directly proceeds to step 6h to select the next band k = 1, and the above-mentioned steps 6e to 6e until the band k = 8 is reached.
The process of setting the noise minimum value noise_min_h (k) in step 6g is repeated.

When the band k reaches 8, the noise minimum value estimating section 29 determines whether or not the frame counter fc has reached 100 in step 6j. Until the number of frames reaches 100, the process returns to step 6b to select the next frame, and the processes in steps 6c to 6i are repeated for the selected frame.

On the other hand, when the processing for the above 100 frames is completed, the noise minimum value estimating section 29 proceeds to step 7a, where the inter-band average (min_all) of the noise minimum value is calculated as nois.
e_min_h (n): Calculated as the average of the sum of the values of n = 0, 1,..., 7 as follows.

Min_all = Σ noise_min_h (n) / 8 n = 0-7 Further, noise_min (0) and noise_min (1) are respectively set as noise_min (0) = noise_min_h (0) noise_min (1) = 0.75 × noise_min_h ( 0) + 0.25 × noise
_Min_h (1) and the band is k = 1.

Further, the minimum noise value estimating section 29 proceeds to step 7b, where the remaining bands k = 8 to 8 are determined based on the eight noise minimum values previously determined for the bands k = 0 to 7. 1
Noise minimum value for 5 is noise_min (2k) = 0.75 × noise_min_h (k) + 0.25 × nois
e_min_h (k-1) noise_min (2k + 1) = 0.75 × noise_min_h (k) + 0.25 × no
It is calculated as ise_min_h (k + 1).

When the above operation is completed, the noise minimum value estimating unit 29 proceeds from step 7d to step 7e, where noise_min (14) = 0.75 × noise_min_h (7) + 0.25 × nois
e_min_h (6) noise_min (15) = noise_min_h (7) is calculated.

That is, the minimum noise value estimating section 29 interpolates 16 min_all based on 8 min_all in the above steps 7a to 7e.

When 16 min_all are calculated in this way,
In step 7f, the noise minimum value estimating unit 29 resets the frame counter fc to 0, sets the noise minimum value to noise_min_h (k) = 36, and sets the band to k = 0,.
・ Set to 7 again.

Then, in step 7g, the inter-band average value min_all of the minimum noise value calculated above and the minimum noise value noise_min (m, k), k = 0,...
Returning to step 6b, similar calculation processing of the minimum noise value and the average value between bands is repeated for the next frame (m = m + 1). <Band-Specific Gain Determination> The band-specific gain determination unit 33 determines the band power channe output from the band power calculation unit 26.
l_power (m, k), noise power noise_power (m, k) output from the noise leak integration value update unit 32, voice weight sp (m, k) output from the voice weight calculation unit 28, and noise minimum value estimation Noise minimum value noise_min (m, k) output from the unit 29
, The gain for each band gain (m, k) is determined.

First, the noise leak integrated value noise_power (m, k)
Noise_power (m, k): k =
Noise_all = Σnoise_power (m, k) / 16 k = 0 to 15 as an average of the sum of the values of 0, 1,..., 15.

Subsequently, band power channel_power (m, k)
Band minimum value min_band and noise minimum value noise_min (m,
The band maximum value max_band of k) is defined as min_band = MIN (channel_power (m, k), k = 2,.
.., 11) max_band = MAX (noise_power (m, k), k = 0,...)
・, 15)

Next, the adjustment value md common to the bands is expressed as md = (noise_all−min_all) × (1−sp / 6) + (min
_Band−max_band) × sp / 6. According to this equation, when sp = 0, that is, in a noise section, md = noise_all−min_al
l sp = 6, ie, in the voice section, md = min_band−max_ban
d, indicating that the transition region takes an intermediate value between these values.

FIGS. 8 and 9 show examples of frequency versus power characteristics for a noise frame and a speech frame, respectively.

In a noise frame, as shown in FIG. 8, the band power is close to the noise minimum. The value obtained by adding the adjustment value to the noise minimum value is obtained by changing the average value to the noise noise average value noise_all while keeping the spectral characteristics of the noise minimum value.

On the other hand, in the case of a voice frame, FIG.
As shown in (2), the value obtained by adding the adjustment value to the noise minimum value is adjusted so that the maximum value of the band coincides with the minimum value of the band power while maintaining the spectral characteristics of the minimum value.

The gain per band gain (m, k) is the band power cha
nnel_power (m, k), noise minimum value noise_min (m, k),
It is determined as follows from the adjustment value.

First, tmp = channel_power (m, k) −noise_min (m, k) −md−1.62
Set to 5.

Next, gain (m, k) (g
ain (m, K) ≦ 0) Change the method of decision. (1) When sp> 0, that is, when voice or a transient frame, gain (m, k) = ｛sqrt (1.4+ (0.7 × tmp) ² ) + 0.7 × tmp
−10｝ × 2 (2) When sp = 0, that is, for a noise frame, gain (m, k) = [sqrt (1.4+ (0.03125 × tmp) ² ) +0.03125
× tmp−10] × 2 This is obtained independently for k =,...

The function form for determining the gain can be set as appropriate. It is sufficient that the voice frame is lower than the noise frame in an area where the value of tmp is small.

The band-specific gain gain (m, k) obtained as described above is multiplied by a transform coefficient for each band in the multiplier 23, thereby performing noise cancellation.

FIG. 10 is a graph showing the relationship between tmp and gain. The audio frame (sp>) is indicated by a solid line.
0), and the dotted line is the case of a noise frame (sp = 0).

When tmp is lower than 0, the gain of the voice frame is lower than the gain of the noise frame. It can be considered that tmp is obtained by subtracting md and a constant (1.625 in the above example) from the SNR of the band, so that although there is a variation in the adjustment value md, the SN of the band is
When R is small, the voice section takes a smaller gain value than the noise section.

This is to actively suppress (small gain value) a band showing a small SNR in a voice section (this is a band that can be estimated to contain no voice component), thereby including a voice component in a voice frame. The result is to make the band stand out. This effect can be achieved by setting the gain smaller than the band gain value of the noise frame.

Such a gain value setting is not limited to the above-described hyperbolic function, and can be performed by various settings.

For example, as shown in FIG. 11, in the case of a voice frame and a transient frame, gain (m, k) = − 20 tmp <0 tmp × 2−200 0 ≦ tmp ≦ 100 tmp> 10 In the case of a noise frame, gain (m, k) =-18 can also be set.

The noise-cancelled transform coefficients for each band are subjected to inverse fast Fourier transform in IFFT 24 and returned to signal frames on the time axis, and then frame-synthesized in frame synthesizing section 25 to produce speech coding circuit 16. Supplied to

As described above, according to this embodiment,
Even in a frame determined to be a voice frame, a gain smaller than a band-specific gain of a frame determined to be a noise frame is set for a band determined not to include a voice component. Band) is emphasized, and as a result, an acoustically good noise suppression output signal can be obtained.

Further, the minimum value of the noise power in each band is obtained by the noise minimum value estimating circuit 29, and the spectrum shape of the noise minimum value is used for the determination of the gain for each band by the gain determining unit 33 for each band. For example, it is possible to realize a noise canceling process that is not affected by a short-term change of a noise spectrum such as when a car passes, and that hardly distorts a voice spectrum.

In addition, a frame in which the significant value y of each frame is large (usually determined to be speech) but the change of the difference of the band-based difference from the previous frame is small (however, the frame is normalized by the average value). When continuous (for example, 100 frames)
Judge as a noise frame and forcibly update the noise power estimation value. At the time of this forced update determination, the continuous section is counted using the value normalized by the average value of the spectrum deviation. can do. Therefore, even if there is a change in a significant value that does not make a good noise frame determination, the forced update is performed, so that a good noise power estimation value can be updated, and thus good noise suppression is performed.

In the above embodiment, the description has been made by taking the TDMA system portable telephone as an example. However, the same applies to the CDMA system (code division multiplex system). There is no change in the function of noise cancellation processed between the microphone and the speech coding circuit, and the present invention can be applied to any type of digital telephone.

The present invention is not limited to communication equipment such as telephones, but can be used for any electronic equipment (recording equipment, portable electronic terminal, etc.) using sound processing.

Each block shown in FIG. 1 is separately described for convenience of explanation, and each block does not need to be an individual element, and one or more functions, for example, a CPU , A DSP, a modem, a voice coding circuit, and the like may be combined into a one-chip LSI.

[0132]

As described above, according to the present invention, it is possible to sufficiently suppress noise components in a speech frame. Therefore, the obtained output signal is excellent in auditory sense. For example, if the output signal is used for a voice input unit of a mobile phone or the like, a high-quality voice signal can be supplied, which greatly contributes to the industry.

[Brief description of the drawings]

FIG. 1 is a circuit block diagram of a digital portable telephone device equipped with a noise suppression method of the present invention.

FIG. 2 is a block diagram showing a functional configuration realized by the noise canceller according to the embodiment of the present invention.

FIG. 3 is a flowchart showing a processing procedure of a significant value calculating unit according to the embodiment of the present invention.

FIG. 4 is a flowchart showing a processing procedure of counting a significant section for determining a noise power forced update according to the embodiment of the present invention.

FIG. 5 is a flowchart showing a processing procedure of voice weight sp of the embodiment of the present invention.

FIG. 6 is a flowchart showing a processing procedure of a noise minimum value estimating unit according to the embodiment of the present invention.

FIG. 7 is a flowchart illustrating a processing procedure of a noise minimum value estimating unit according to the embodiment of the present invention.

FIG. 8 is a diagram illustrating an example of frequency versus power characteristics in the case of a noise frame.

FIG. 9 is a diagram illustrating an example of frequency versus power characteristics in the case of a voice frame.

FIG. 10 shows tmp-ga according to the embodiment of the present invention.
FIG.

FIG. 11 shows tmp-ga according to the embodiment of the present invention.
FIG.

[Explanation of symbols]

 17 ・ ・ ・ Noise canceller

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI theme coat ゛ (reference) H04J 3/18 G10L 7/04 G H04M 1/00 9/18 MF term (reference) 5D045 DA20 5J064 AA05 BA01 BB07 BB08 BB12 BC02 BC05 BC06 BC07 BC17 BC18 BC21 BC29 BD02 5K027 AA11 DD12 DD18 5K028 AA04 AA12 BB04 EE07 HH01 PP03 5K052 AA01 BB02 BB21 CC06 DD01 EE11 EE31 FF12

Claims (12)

[Claims] 1. A noise suppression method for dividing an input audio signal into frames in a predetermined time unit, dividing the divided frames into predetermined frequency bands, and performing noise suppression processing for each of the divided bands. At: a voice frame determining step of determining whether the frame is a noise frame or a voice frame; and a band-specific gain determining step of setting a band-specific gain value of each frame based on a result of the voice frame determining step. A signal generation step of performing noise suppression for each band using the band-specific gain value determined in the band gain determination step, and then reconstructing a frame to generate a noise-suppressed output signal; In the gain determination step for each band, the gain value for each band when the frame to be determined is determined to be an audio frame is determined. A noise suppression method comprising: setting a gain value for each band so as to take a value smaller than a gain value for each band when the target frame is determined to be a noise frame. 2. In the step of determining a gain for each band, a gain value for each band of a band estimated to contain no audio component in a frame determined to be a frame to be determined is determined as a frame to be determined. 2. The noise suppression method according to claim 1, wherein the band-specific gain value is set so as to take a value smaller than the band-specific gain value when is determined to be a noise frame. 3. A gain value for each band is determined based on a difference between a signal power and a noise power for each band. 2. The noise suppression method according to claim 1, wherein a value determined when a voice frame is determined is set to a value smaller than a value determined when the frame is determined to be a noise frame. 4. In the voice frame determination step, a significant value (sub) for each band obtained by weighting a difference between the power for each band of the target frame and the estimated noise power value for each band in a predetermined manner.
4. The noise suppression method according to claim 1, wherein whether or not the current frame is a noise frame is determined based on a significant value (y) obtained by adding y) under a predetermined condition. 5. A noise suppression method for dividing an input voice signal into frames in a predetermined time unit, dividing the divided frames into predetermined frequency bands, and performing noise suppression processing for each of the divided bands. At: a frame dividing step of dividing an input signal into frames in a predetermined time unit; a frequency band dividing step of dividing each frame into a plurality of frequency bands; and calculating a power per channel (channel_power) for each frequency band. A power calculation step for each band; a noise power estimation value (noise_power) for each band for each frequency band;
nnel_power) and a significant value (y) obtained by adding a significant value (suby) for each band obtained by performing a predetermined weighting on the difference (tmp) under a predetermined condition. Value calculating step; sum of absolute values (su) of differences between significant values (suby) for respective frequency bands between the current frame and the previous frame.
m) calculating a sum of significant values for each band;
Calculate the average value (sum_average) of (sum) and calculate the sum of the absolute values.
a significant value normalizing step of calculating a ratio (r) obtained by normalizing (sum) by the average value (sum_average) of the sum of absolute values; and a current value when the significant value (y) falls below a predetermined threshold value. A first noise power estimation value updating step of determining a frame as a noise frame and updating the noise power estimation value (noise_power) for each band, and the ratio (r) continuously exceeds a predetermined threshold for a predetermined period. A noise power estimation value updating step comprising: a second noise power estimation value updating step of determining the current frame as a noise frame when the value falls below the threshold value and updating the noise power estimation value for each band (noise_power); Is set so that the gain value for each band when it is determined that the target frame is a noise frame can be smaller than the gain value for each band when the target frame is determined to be a noise frame. Gain for each band to be performed Noise suppression method characterized by having a constant step. 6. A noise power estimating step of determining signal power for each frequency band and estimating noise power for each band based on the band power; A minimum value detecting step of detecting a minimum value of power over at least one of a plurality of band-wise noise powers over a plurality of frame periods; and for each of the frequency bands, the band power and the band detected by the minimum value detecting step. 6. The noise suppression method according to claim 1, wherein a noise suppression amount for each frequency band is determined based on a difference obtained from a minimum value determination step for each band for obtaining a difference from the minimum value. 7. The noise suppression amount uses an adjustment value for generating an adjustment value common to different bands for each frame, and for each of the frequency bands, a value obtained by adding the minimum value for each band and the adjustment value, and a band for the adjustment value. 7. The noise suppression method according to claim 1, wherein a difference from power is determined, and a noise suppression amount for each frequency band is determined based on the difference. 8. The adjustment value is: in a noise section, an adjustment value common to bands is determined based on a difference between an average value between the minimum values for the bands and an average value between noise powers for the bands; 8. An adjustment value common to bands is determined in a section based on a difference between a minimum value among a plurality of band powers in one frame and a maximum value among a plurality of band-specific minimum values. The described noise suppression method. 9. A noise power estimating step of determining the power of a signal for each frequency band and estimating noise power for each band based on the band power; A comparison step of determining the difference between the band-specific noise power and the band power and comparing these band-specific differences with a predetermined threshold value; An adding step of, when it is determined that the difference by band exceeds the threshold value, adding these differences to each other after performing predetermined weighting; and adding the difference by band obtained by the adding step. 9. A noise suppression method according to claim 1, wherein the determination is performed by a determination method comprising: determining whether the input signal is a voice section or a noise section based on the value. Method. 10. The noise suppression method according to claim 9, wherein in said adding step, weighting is performed on each band difference so that the weight decreases as the frequency increases. 11. The method according to claim 9, wherein the determining step determines whether the input signal is a voice section, a noise section, or a transition section which is an intermediate area between the two sections, based on the added value. 11. The noise suppression method according to any one of claims 10 to 10. 12. An electronic device having an audio signal input section and an audio encoding section, wherein the electronic apparatus receives an audio signal from the audio signal input section and outputs a signal whose noise has been suppressed by the noise suppression method according to claim 1 to 11. An electronic device comprising: a noise suppression unit that supplies a signal to an encoding unit.