JP3244252B2 - Background noise average level prediction method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention predicts an average level (average power or average amplitude value) of background noise from a signal received by a sound receiver, for example, from a speech signal mixed with background noise. The present invention relates to a method and an apparatus for estimating a background noise average level applied to a speech signal detection and estimating a background noise average amplitude value in a reverberation canceling device used in a two-wire / four-wire conversion system and a loudspeaker system. .

[0002]

2. Description of the Related Art Background noise which is the square of the average amplitude of background noise is used.
Accurately predicting the average sound level depends on voice detection technology and
For sound information devices such as voice switches and echo cancellers
This is a very important technology. Generally using a microphone
When sound is received, background noise such as air conditioning
Mixed. At this time, conventionally, only the voiced sound section
If you try to detect
Noise level at threshold P _tAnd the microphone receiving signal
Level P_xFor P_x> P_tIs a voiced section
Is performed. But background noise
Is greater than or equal to a predetermined threshold,
There was a problem that it was judged to be between.

Therefore, a method has been proposed in which the background noise level is adaptively predicted and a threshold value is determined based on the predicted value. In this method, the microphone receiving signal is divided into fixed time intervals, an average power in each time interval is obtained, the average power is stored in a memory for a certain time, and a minimum one of the average powers is set as a background noise level. There is a method. However, this method has a problem that the background noise level is predicted to be smaller than the actual value when the background noise is instantaneously reduced.

Attention is also paid to the fact that the level difference of the average power in each time section is smaller than that of speech in the background noise. If the level difference of the average power in each time section is smaller than a certain value, the average power May be used as the background noise level. Also in this method, it is necessary to determine by trial and error how much the level difference of the average power in the time section should be, which lacks versatility.

[0005]

Conventionally, a method of predicting the level of the background noise by using the average power level for each time section has been proposed. When the minimum value of the average power for each time section is used, There is a problem that the predicted background noise level becomes too small, and the use of a predetermined threshold value lacks versatility. An object of the present invention is to provide a method and apparatus that can predict the level of background noise mixed in an information signal in a general-purpose manner even when an information signal such as voice and background noise are always mixed. The purpose is.

The background noise average level prediction of the present invention
A predetermined short-term average level of the force signal is obtained.
The short-term average level belongs to the level section divided into
The appearance frequency of each section, and for each section,
The appearance frequency is higher than the appearance frequency in the section
Find the section value that minimizes the section value in the peak section
The average level of the input signal is longer than the short
And the level corresponding to the minimum section value and the long
Long term average level is smaller than average level
The average level for a long time
The average level of the background noise is used.

Further, in order to follow a change in the background noise level, the past histogram data is multiplied by a forgetting factor every time the short-time average level is calculated. According to this configuration, the average value can be reliably obtained without being affected by the instantaneous reduction of background noise that is instantaneously reduced because the short-time average level histogram is used. Becomes

[0008]

FIG. 1 shows an embodiment of the device according to the present invention. The output of the microphone 10 is converted into a digital signal at a period of, for example, 1 / (8 kHz) by the AD converter 11, and the square root of the short-time average power and the long-term average power by the short-time power calculation unit 12 and the long-time power calculation unit 13, respectively. The square root is calculated, and the short-time average power square root is calculated by the histogram calculator 14 for the histogram.
As the basis of the histogram threshold P _t is determined by the adaptive threshold determining section 15, and the threshold P _t and long-term mean power square value are compared by the comparator unit 16, a sequence of the comparison result Then, the number of times equal to or smaller than the threshold value is counted by the counting unit 17, and when the counted value reaches a predetermined value, the long-term average power square root value at that time is output from the background noise average amplitude value determining unit 18.

According to the present invention using the apparatus shown in FIG.
4 shows a processing procedure of a background noise average level estimation method. Figure 1 and
Referring to FIG. 2 and FIG.
The method will be explained. First, the count value K of the counting unit 17 is set to 0.
To (S₁), Received by the microphone 10
And a sound receiving signal x converted into a digital signal by the AD converter 11.
(N) (n indicates sampling time) is short-time power
Supplied to the calculation unit 12 and the long-time power calculation unit 13
For a predetermined short time, for example, 5 to 10 milliseconds (
If it is difficult to distinguish between background noise and voice)
Average power P _s(N) and a predetermined long time,
For example, for 100-300 milliseconds (the longer the average accuracy, the better
(Although it also mixes audio signals)
P_L(N) is calculated by the following equations.

P _S (n) = α ₁ × P _S (n−1) + (1−α ₁ ) × x ² (n) (1) P _L (n) = α ₂ × P _L (n−1) ) + (1−α ₂ ) × x ² (n) (2) α ₁ and α ₂ are quantities of 1 or less relating to the time of power averaging, and T is the integration time (average time).
The relationship of 1 / (1−α) holds. Therefore, α ₁ <α ₂ <
It becomes 1. Furthermore, these average powers P _S (n) and P _L (n)
Square root _{√P S (n), √ P} L (n), that is obtaining the average amplitude (S _2, S _3). These calculations are performed for each n, that is, for each sampling cycle of the sound receiving signal.

Next, the histogram of the short-time power square root √P _S (n) is calculated in the histogram calculator 14 (S ₄ ). That is, a range of predicted √P _S (n), for example, half of the maximum voice amplitude (the background noise is usually smaller than voice) is divided into a plurality of amplitude sections (0, 1,..., N) at equal intervals δ.
Each time the short-time power square root ΔP _S (n) is calculated, the number of amplitude sections to which the value belongs is incremented by one. That is, the following calculation is performed.

[0012] _{h (int (√P S (n} ) / δ)) = h (int (√P S (n) / δ)) + 1 ... (3) int (A) is truncated after the decimal point of the value of A Indicates that the value is to be converted to an integer. Therefore, for example, when δ = 40, ΔP _S (n)
Is 50, h (int (50/40)) = h (1), and 1 is added to the first amplitude section, and ΔP _S (n)
Is 90, h (int (90/40)) = h (2), and 1 is added to the second amplitude section.

In order to make this histogram easily follow changes in the background noise level, a forgetting factor λ (for example, about 500 to 1000 milliseconds) of 1 or less is set for each of the amplitude sections 0 to N as shown in the following equation. multiplying respectively the number (S _5). h (i) = λ × h (i) (4) i = 0, 1, 2,..., N Next, the calculated histogram h (i) (i = 0,.
N) is transferred to the adaptive threshold value determination unit 15. In the adaptive threshold value determining unit 15, h (i) (i = 0,..., N)
, The peak h (i) is searched, that is, for the appearance frequency h (i) of each section, an amplitude section larger than the appearance frequencies h (i−1) and h (i + 1) of the preceding and following sections is determined. Then, the section number is searched and stored (S ₆ ). Further, a section number i which is the smallest in the amplitude section of the searched peak.
And this is set as w (S ₇ ). When the signal h and the noise h and the background noise are mixed as shown in FIG. 3A and the histogram h (i) of the short-time power square becomes as shown in FIG. 3B, the appearance rate h (1) of the first amplitude section is obtained. ) Is the peak and the section number i is the minimum, and w
= 1. As a result, the adaptive threshold P _t (n) is
In consideration of the effect of truncation by the int () function, the value is obtained by the following equation (S ₈ ).

P _t (n) = δ × (w + 1) (5) Here, taking the actual signal as an example, the determination of the adaptive threshold will be further described. FIG. 3A shows the signal x (n) observed by the microphone 10. x (n) is a signal obtained by adding an audio signal to background noise. In FIG. 3B,
The histogram calculated by the histogram calculator 14 with the amplitude section δ as 40 is shown. The adaptive threshold value determination unit 15 searches for the frequency having the smallest amplitude section number from among the appearance frequencies indicating the peaks, and this is found to be section 1 from FIG. 3B. As a result, the adaptive threshold becomes (5)
According to the equation, P _t (n) = 40 × (1 + 1) = 80 (6) This corresponds to that the amplitude section number 1 of the appearance frequency of the amplitude absolute values 40 to 80 is the minimum section number of the peak section.

In the comparing section 16, the adaptive threshold value determining section 1
Adaptive threshold value P determined in step 6_t(n) and long time power
-Long-term average power square root calculated by calculation unit 13
P_L (n) and (S₉), P_L(N) <P_t(N)
If there is, the count value k of the counting unit 17 is incremented by 1 (S_Ten), This
It is determined whether or not the counted value is equal to or greater than a predetermined value K (S_{1 1})
 . P in comparison unit 16_L(N) is P_t(N) Total if less than
The count value k of the number part 17 is reset to 0 and step S_TwoTo
Return (S₁₂), Step S₁₁And if k <K, step
_TwoReturn to Step S₁₁If k> K, that is, the predetermined time
P continuously_L(n) <P_t(N), P_L(N)
The scene noise average amplitude value determination unit 18 calculates the average amplitude value of the background noise
Decide (S₁₃). The value of K is time, for example, about 1 second
Degree, that is, the time when mutual conversation is clearly interrupted.
It is.

Each process in FIG. 2 is performed at the sampling period of the sound receiving signal as described above. In the long-time power calculation unit 13, in the initial state, P _L (n) up to that point is set to zero, and after a long time for averaging has elapsed, the processing according to the present invention is correctly performed. In the above-described embodiment, the average amplitude was predicted by taking the square root of the average power.
The average power of the background noise may be predicted. That is, in the present invention, the level means amplitude or power.

[0017]

As described above, according to the present invention, the background noise average level value can be adaptively adjusted without assuming the background noise level in advance, even when speech and background noise are mixed. It can be predicted as an adaptive threshold. Further, an accurate background noise average level value that is equal to or less than the adaptive threshold value can be measured. The present invention is capable of measuring an accurate background noise average level value,
Utilization for voice detection, noise reduction, and the like can improve the accuracy of voice detection and improve the amount of noise cancellation.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the apparatus of the present invention.

FIG. 2 is a flowchart showing an embodiment of the method of the present invention.

FIG. 3A is a diagram showing a signal in which voice and background noise are used to show the effect of the present invention, and FIG. 3B is a diagram showing a calculation result of a histogram calculator 14 in the present invention.

──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Junko Shimizu 1-1-6 Uchisaiwaicho, Chiyoda-ku, Tokyo Nippon Telegraph and Telephone Corporation (72) Junji Kojima 1-16-1 Uchisaiwaicho, Chiyoda-ku, Tokyo Japan (56) References JP-A-2-272835 (JP, A) JP-A-4-119028 (JP, A) JP-A-4-120927 (JP, A) JP-A-6-197049 ( JP, A) JP-A-7-74709 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) H04B 3/00-3/44

Claims (3)

(57) [Claims] A step of obtaining a predetermined short- time average level of an input signal; and a step of calculating an appearance frequency of a section to which the short-time average level belongs among a plurality of divided level sections . For each of the above sections , of the peak sections where the appearance frequency is higher than the appearance frequency in the sections before and after it,
Phase and the steps of obtaining a long-time average level than short above for the input signal, the level corresponding with the period value becomes the minimum and the long-time average level for obtaining the interval value which the interval value is minimized DOO compare, the long-term average and towards the level is low it is determined continuously for a predetermined time period the steps of the long-term average level and the average level of background noise, background noise average level prediction method having. 2. The short-time average level P _S (n) (n is
Sampling time) is calculated as α ₁ × P _S (n−1) + (1−
α ₁ ) × x ² (n) (x (n) is the above input signal)
The long-term average level P _L (n) is represented by α ₂ × P _L (n−
1) + (1−α ₂ ) × x ² (n) (α ₁ <α ₂ <1)
2. The method for predicting an average background noise level according to claim 1, wherein the calculation is performed . 3. A division and the short level calculation means for calculating a mean level of a predetermined short time for the input signal, and a long level calculation means for calculating a long-time average level than short above for the input signal, a plurality Of the short-term average level
And frequency calculating means which calculates the occurrence frequency of the section Le belongs, the ward of the peak interval frequency than frequency of appearance before and after the interval for each interval has a large
Adaptive threshold value determining means for determining a level corresponding to the interval value at which the inter-value becomes a minimum as a threshold value; comparing means for comparing the threshold value with the long-term average level; A background noise average level predicting device comprising: a background noise average level determination unit that outputs a long-term average level at that time as a background noise level when a value equal to or less than a predetermined value is continuously obtained.