JPH0213999A - Voice signal detection method and apparatus - Google Patents

Abstract

PURPOSE: To correctly detect a speech signal among signals buried in noises by judging that there is the speech signals among the signals affected by the noises when the maximum energy or energy ratio of digital signals passed through filters is larger than its maximum threshold value. CONSTITUTION: The filter 44 and an energy calculation device 45 provide a parameter Xph and an energy calculation device 46 provides a parameter X; and the parameters X and Xph are applied to the 1st operand input and 2nd operand input of a divider 47 respectively to calculate a parameter R. Then when the maximum energy or energy ratio R of the digital signals passed through filters 43 and 44 is larger than its maximum threshold value, it is judged that there is the speech signal among the signals affected by noises. Consequently, the speech signal among the signals buried in the noises can correctly be detected.

Description

BACKGROUND OF THE INVENTION The present invention relates to a method and apparatus for detecting audio signals, which can be used in particular for alternating wireless electrical transmission on vehicles.

Description of the Prior Art Most prior art voice activity detectors have a minimum of two
It cannot operate correctly except in the case of a sufficiently high signal-to-noise ratio of about 0 dB. Moreover, this corresponds to working conditions in a quiet office-type environment.

In contrast, on a vehicle, speech/noise discrimination must take into account a very weak signal-to-noise ratio, typically less than 10 dB. Under certain conditions (eg, at high engine speeds in a vehicle with average soundproofing), the noise level may even exceed the signal level.

Finally, the level and type of noise to be discriminated will vary depending on the specific conditions of the vehicle (eg, the degree of soundproofing), but will also vary as a function of the route taken. A particularly unfavorable example is that of urban routes. The noise to be considered there is generally high level, non-stationary and usually highly variable.

An embodiment of a voice activity detector designed to operate in noisy environments was submitted for the applicant, 197
From patent application number 7974227 dated September 28, 1999. However, this detector cannot be used to optimize speech/noise discrimination, except for voice sounds, and decisions are made by comparing the audio signal exclusively to a threshold voltage value. This variable is automatically linked to the maximum amplitude value of the audio signal without taking into account the actual noise level. The result is a level of performance that is not sufficient to be able to operate properly in highly disturbed environments where speech signals are buried in noise.

SUMMARY OF THE INVENTION The aim of the invention is to overcome the above-mentioned drawbacks. To this effect, the subject of the invention is a method for the detection of a speech signal in a signal buried in noise. The above method consists of the following steps.

・Dividing the signal into several frames.

- Sampling each frame to obtain a digital signal containing a determined number n of samples.

- Preamplifying a digital signal to obtain a preamplified digital signal.

- Filtering the preamplified digital signal by a high-pass digital filter to obtain a filtered digital signal.

- Measure the maximum energy of the samples of the preamplified signal and the maximum energy of the samples of the filtered digital signal in each frame.

- Taking the energy ratio between the maximum energy of the sample of the digital signal passed through this filter and the maximum energy of the sample of the preamplified digital signal.

Calculating the long-term average value of the energy and energy ratio of the samples of the filtered signal between two limits.

Calculating four thresholds based on long-term average values, two of which are the maximum and form two lower limits of the conversation state for the filtered signal and energy ratio, respectively. However, two of them are the minimum signal and form two upper limits of the noise state to the filtered signal and energy ratio, respectively, and the maximum energy of the filtered signal and the energy ratio are defined by these thresholds. Compare with value.

- Determining the presence of a speech signal in the noise corrupted signal when the maximum energy or energy ratio of the filtered digital signal is greater than the respective maximum threshold.

and determining the absence of a voice signal in the noise-affected signal when the maximum energy or energy ratio R of the filtered digital signal is smaller than the respective minimum threshold.

A further subject of the invention is a device for carrying out the above method.

In the margin below (Example) The method according to the invention, illustrated in FIGS. 1 to 4, is carried out on a signal with about 20 milliseconds of noise introduced, and with 160 samples per frame to give a signal sample S. This is a practical example, sampled by percentage. As shown in steps 1 to 5 in FIG. 1, the digital signal S to be processed is first processed in step 1.
It is preamplified in step 2 to give a signal sample Sn, and then amplified in step 2 by a high-frequency two-wave generator with a cutoff frequency FC of 1200 Hz to give a signal sample sph (n). In the next steps 3 and 4, the following parameters are calculated.

X = mas (Sn) x, h = maxSph (rl) where n is an integer between 1 and 160. These calculations consist of looking for the sample with the maximum amplitude or maximum energy in each sequence of samples 5(n) and S+>h(n).

Step 5 consists of calculating the ratio R=xph/X of the two parameters Xpb and X calculated in steps 3 and 4.

The following steps 6 to 11 consist of calculating the parameters x1 and R1 according to the following relationships.

If Xph is larger than X, calculated in the previous frame in Fig. 1, then X,. If we write it as 1d, ・×1=xph
Otherwise・Xt=TX−Xtold+(1−TX)XphR
is larger than the R ratio calculated in the previous frame in Figure 1,
If we write it as Rlold, ・R,=R Otherwise, ・R+=Tr Rlold + (1-Tr) To the frame, instantaneous growth may be allowed, while its decrease will occur more slowly, with time constants equal to TX and Tr, respectively. According to Mr. B of the proposed implementation of the present invention, the value of the time constant is fixed at 0.75. This corresponds to approximately 70 milliseconds. The next steps 12 to 29 are shown in Figures 2 and 3.
As shown in the figure, it consists of determining four detection thresholds using long-term average values of parameters Xph and R.

The latter is first limited between a constant, maximum and minimum value in step 12 so as to prohibit excessive changes in the threshold. The limits of change of xph and R2 are respectively Xphi
They will be called nf, Xphsup, R,inf, and R,sup.

Steps 13 to 22 involve calculating two parameters x 2 and R2, and confirming the following relationship.

X2 = MAX (MIN (Xph, Xph, 5
up) + Lh, 1nf) R2= MAX (MI
N CR, R, 5up), R, 1nf) The long-term average values of parameters xp and R are Xmoy and Rm, respectively.
oy and is calculated in steps 23 to 28 in applying the following relationship:

x2 or the parameter R□ calculated in the previous frame in FIG. If we write it as Xmoy-0ld, which is larger than A, then
Xa r Otherwise・Xmoy” Td−Xmoy.old+(+a+(I
Td), X*R2 is calculated in the previous frame in Figure 3, and the parameter Rffi is
. , if larger than ・Rmov” Tm・Rmoy.old+(ta”(I
Tm) R2゜Otherwise, island. y"Td-Rn+oy.old+(+d"(1
-Td), R2 In these relationships, the rising time constant T,
is provisioned to rise exponentially slowly, while the falling time constant T, is provisioned to rise exponentially fast, and the considered mean quickly falls back to a level corresponding to the noise. . In the proposed embodiment of the invention, the values of these time constants are 0 for increasing
．． 95, or approximately 400 m5, and for machining, it is fixed at 0.2, or approximately 13 m5. Finally, the four values of the threshold are calculated in step 29, defined above by the following relationship: , and the value of Rmo y.

SXt conversation = a, Xmoy+Xph, 1nfS
X, Noise = b, X, , +Xph, 1nfSR, Conversation = a, R+may+R, 1nfSR4 Noise = b
, R,11,y+R,inf The values of the multiplication coefficients a and b are fixed to 1.8 and 2.5 in the example proposed by the present invention. It should be noted that it is desirable to keep in mind that if either of the parameters Xph or R is smaller than the corresponding lower limit value, the relevant decision will be taken automatically.

The apparatus for carrying out steps 1 to 5 of the method and calculating the energy ratios is shown in FIG.

This device has a first filter 43, which has a transfer function H(z) = 1-0.86.z-'' and performs the preamplification of the signal shown in step 1. Depending on the output, first the second high frequency waveform filter 4
4 with a cutoff frequency of approximately 1200 Hz, and the second is connected to an energy calculation device 46.

The second high-pass filter 44 is also coupled at its output to an energy calculation device 45 similar to energy calculation device 46 . Filter 44 and energy calculation device 45
is the parameter X by executing Steps 2 and 3 of Kempo
ph, and the energy calculation device 46 provides the parameter
I will provide a. Parameters X and Xph are each divided by divider 4
7 is added to the first operand input and the second operand input, and the parameter R is calculated by step 5.

An embodiment of the energy calculation devices 45 and 46 is shown in FIG. The circuit includes a comparator circuit 48 coupled to a register 49 through a bypass circuit 50. This comparator circuit 48 has two inputs, the first input receiving signal samples S (n) provided by digital filter 43 or signal samples provided by digital filter 44 . The second input is connected to the output of register 49. The bypass circuit 50 is controlled by the input of the comparator circuit 48 and outputs signal samples 5(n
) or 5ph(n) is greater than the contents of register 49, the signal sample 5(n) for the input of register 49
Or bypass 5ph(n).

Otherwise, register 49 remains closed to itself.

One embodiment of an apparatus for carrying out steps 6 to 11 is shown in FIG. The device has a comparator circuit 51 coupled to an accumulator circuit 52 through a bypass circuit 53. A multiplier circuit 54 is connected by a first operand input to a first input of the comparator circuit 51 and receives the parameter 1-T, or 1-Tr, at its second operand input. These parameters are those mentioned in Steps 8 and 11 of Kenpo. Second multiplication circuit 55
is connected by a first operand input to the output of the accumulator 52 and receives at a second operand input the parameters TX or T, represented by Kenpo steps 8 and 11. The outputs of the multiplier circuits 54 and 55 are connected to the first and second operand inputs of the adder circuit 56, respectively, and the output thereof is connected to the first input of the bypass circuit 53. The output of accumulator 52 is further connected to a second operand input of comparator circuit 51. Steps 6 to 1
According to up to 1, the parameter Xph or R is applied to the first input of the comparator circuit 51 and the contents X, of the accumulator circuit 52. ld or R1゜1. compared to Step 6 or according to step 9, the parameter Xph or R is the content X of the accumulator circuit 52. ! , or Ro. If it is greater than ld, the bypass circuit 53 updates the contents of the accumulator circuit 52 with one of the parameters Xph or R according to steps 7 and 10. Otherwise, the bypass circuit 53 switches the output of the adder circuit 56 to the input of the accumulator circuit 52, and the contents of the accumulator are determined by the parameters x1 or R1, as determined by the relationships described above with respect to steps 8 and 11. Update. In these relationships, the product (1-TX)
Xph or (1-Tr) XR is obtained by the multiplication circuit 54, and TxxX-o1d or TRXRold
It is executed by Fa55 Lt times 1jt times. The sum of these products is created by adder circuit 56.

Steps 12 to 22 of the method are shown in FIG. 2 or performed by a threshold amplifier (not shown), the characteristics of which are shown in FIGS. 8(A) and 8(B).
is shown. These threshold amplifiers make it possible not to take into account excess values of the parameters x1 and R1. Due to these characteristics, each parameter x1 or R□ has two values, xlph, tnf and X1ph, 5
It is limited between Ijp or Rt, inf and Rt, sup. These characteristics are thresholds Xtp), , inf and xl
ph, 5ul) or R,,inf and R,,sup, the linear relationship of the parameters x1 and R1 allows the creation of the parameters x2 and R2. This parameter x
2 and R2 are those whose amplitudes are limited with respect to the values of parameter X and R8 outside these thresholds.

One embodiment of an apparatus for calculating the mean value island or Rue, which will be described below by steps 23 to 28 of Margin Kempo, is shown in FIG. This device includes a subtraction circuit 57, a multiplication circuit 58, an addition circuit 59, and a register 60, which are connected in series in this order. Subtraction circuit 57 has a first operand input to which parameter x2 or R2 is added, and a second operand input connected to register 60. This device also includes a comparator circuit 61 with two inputs.
are respectively connected to two inputs of the subtraction circuit 57. The output of the comparator circuit 61 is sent to the bypass circuit 62
connected to the control input of the This bypass circuit has two inputs to which time constants Tm and T are applied. The output of the bypass circuit 62 is connected to the first operand input of the multiplier circuit 58, and the second operand input of the multiplier circuit 58 is connected to the output of the subtraction circuit 57. The output of the multiplier circuit 58 is further connected to the first operand input of the adder circuit 59, and the second operand input of the adder circuit 59 is connected to the first operand input of the subtracter circuit 57. This average value calculation device can perform the calculations of the method shown in steps 23 to 28. Step 23 or Step 26
, the parameter x2 or R2 is applied to the first comparison input of the comparator circuit 61 and the contents of the register 60 xr,
. 9. . ld and the respective values are stored in register 60.
, then the comparator 61 instructs the bypass circuit 62 to add a time constant T, , to the first operand input of the multiplier circuit 58. Multiplier circuit 58 has at its second operand input the contents of register 60 x 7°. 1d and the parameter x2 or R added to its first operand input
Receives the result of subtraction between the value of 2 and 2. The multiplication T performed in the multiplication circuit 58, (X,,,.old+(ld
-Xl) or T, (X, . . .td-R2) is added to the first operand input of the adder circuit 59 and added to the parameter x2 or R2 added to its second operand input. The result of addition performed by adder circuit 59 is transmitted into register 60. However, in step 23 or 26 the value of parameter x2 or R2 is found in register 60. ,. 1d or R1°,
. If it is not greater than the value of l, the bypass circuit 62 is commanded by the comparator circuit 61 to add the value of the time constant Td to the first operand input of the multiplier circuit 58. For condition 6 of these, the calculation is done as described above,
The time constant T, is replaced by the value of the time constant Td according to the relationship shown in steps 25 and 28 of Kenpo.

Speech threshold or noise threshold (SX, ``Speech'' and sx, ``Noise II, SL II Speech'' and sn.

The calculations according to the relationships established in step 29 of the Kempo ``Noise'') are performed by the circuits described in FIGS. 10(A) and 10(B). SXt “Conversation °° or SR.

The threshold value for "conversation" is calculated by a multiplier circuit 63 connected to an adder circuit 64. Multiplier circuit 63 receives at its first operand input x, vlQy or Rmoy provided by register 60 of FIG. 9, and also has a second operand input to which the parameter steal is added. The result of the multiplication is added to the first operand input of adder circuit 64 as well as to a threshold value Sph, inf which is applied to its second operand input. The output of adder circuit 64 will provide the SX1 "speech threshold" or the SR1' speech threshold.

Similarly, SX1 “Noise threshold or SR.

′“Noise°°threshold or both are multiplier circuit 6
5 and the adder circuit 66. Multiplication circuit 65
The first operand inputs of are the parameters Xmoy and R, given in register 60 of FIG. , receive. Multiplier circuit 65 also has a second operand input to which the parameter DA is added. Its output is connected to a first operand input of an adder 66, whose second operand input receives the value of the parameter threshold X ph,inf. The output of the adder circuit 66 is the threshold value sx, '“
Output "noise" and SR "noise". These thresholds are used in steps 30 to 40 of the method and in the first
According to the graphs shown in Figure 1 (A) and Figure 11 (B), x
This allows comparison of l and R1. The corresponding comparator is the first
This is shown in Figure 2. This circuit has a set of four comparator circuits numbered 67 through 70, each coupled to four inputs of speech/noise discriminator 71. Comparator circuit 67 compares parameter x1 with speech threshold SX1 "speech", comparator 68 compares parameter x1 with threshold SX1 "noise", comparator 69 compares parameter R1 with threshold SR , “conversation” °, the comparator 70 sets the parameter R1 to the threshold value SR+
The speech/noise discriminator 71 produces a voice activity signal DAV according to the state diagram shown in FIG.
, and there are unstable states represented by letters Ll to R4. Steady state DAVO is the "noise" state in which the voice activity detector is placed when there is no speech signal.

The stable state DAVI is also the state in which the voice activity detector is placed when the input contains an applied signal or a speech signal. When the detector is in the “noise” state DAVO, one of the two parameters
, ``conversation'' or SR, ``conversation °'', only goes to the conversation state DAVI.

If not, i.e. parameter x8 or threshold S
If the parameter R1 is smaller than the threshold value SR, "conversation", the determination as noise is maintained.

In contrast, when the voice activity detector is in the conversation state DAVI, one of the two parameters x1 and R1 is less than or equal to the corresponding noise threshold, i.e. xl equals the threshold SX + "noise" The detector can go to the noise state DAVI only if R1 is less than the threshold SR, "noise" °. Under these conditions, the detector passes through the unstable state L2 .This algorithm of state changes of signal DAV is performed in step 30 of FIG.
It is expressed from 39 to 39. After each change in the state of signal DAV, and after the initialization phase represented by step 40, the method returns to the operations of step 6 of FIG.

However, as shown in steps 41 and 42 of the diagram in FIG.
Valid only at the end of a period, calculated by a timing counter (not shown) marked . This timing counter is loaded with the maximum count value in steps 35 and 39 whenever the "conversation" state DAVI is determined, and the contents of the counter are loaded with the maximum count value whenever the determination DAVO occurs in step 36.
-reduced by a unit. This can allow speakers to systematically avoid going into a "noise" state during gaps in begging conversations, or while cutting off the end of a speech when its energy is low. It is.

Embodiments of the method according to the invention are not limited to the devices just described, but also to structures comprising computing means having a recorded microprogram, such as a read-only memory (ROM). It is quite clear that the same can be implemented equally well.

[Brief explanation of the drawing]

1, 2, 3 and 4 are flowcharts illustrating different steps of a method implemented in accordance with the present invention. FIG. 5 shows a device for calculating energy ratios, implementing steps 1 to 5 of the method according to the invention. Figure 6 shows the filtered or preamplified 5
2 shows an embodiment of a device for calculating the value of the sample with maximum energy in one frame of the signal of the figure; FIG. FIG. 7 shows an embodiment of an apparatus for carrying out steps 6 to 11 of FIG. FIG. 8(A) and FIG. 8(B) show step 1 of FIG.
2 to 22 are two graphs illustrating the method used to determine the threshold; FIG. 9 shows steps 12 to 22 in FIG.
Mean island explained. , and an embodiment of an apparatus for calculating R1°7. 10(A) and 10(B) show two circuits for threshold calculation according to the invention. FIGS. 11A and 11B show two graphs illustrating a mode of comparison with adaptive thresholds in accordance with the present invention. FIG. 12 shows an implementation of a comparison device for carrying out steps 30 to 40 of FIG. FIG. 13 is a state diagram illustrating a decision algorithm that allows determining whether an audio signal is present in an audio signal.

Claims (11)

[Claims] (1) A method for detecting a voice signal in a signal buried in noise, which consists of the following steps: (a) A step of cutting the signal into several frames. (b) sampling each frame to obtain a digital signal containing a determined number of n samples; (c) preamplifying the digital signal to obtain a preamplified digital signal. (d) To obtain the filtered digital signal,
Filtering the preamplified digital signal by a high-pass digital filter. (e) measuring, in each frame, the maximum energy of the samples of the preamplified signal and the maximum energy of the samples of the filtered digital signal; (f) taking the energy ratio between the maximum energy of the filtered digital signal and the maximum energy of the preamplified digital signal samples; (g) calculating a long-term average value of the energy and energy ratio of the samples of the filtered signal between two limits; (h) calculating four thresholds based on long-term average values, two of which are maximum values and two lower limits of the conversational state for the filtered signal and energy ratio, respectively; , two of which are the minimum signal and form two upper bounds of the noise conditions to the filtered signal and energy ratio, respectively, and the maximum energy of the filtered signal and the energy ratio and these Compare with threshold. (i) determining the presence of a speech signal in the noise corrupted signal when the maximum energy or energy ratio of the filtered digital signal is greater than a respective maximum threshold; (j) and determining the absence of a speech signal in the noise-affected signal when the maximum energy or energy cost R of the filtered digital signal is less than the respective minimum threshold. 2. The method of claim 1, wherein the digital signal is preamplified by a z-transforming high-pass digital filter (H(z)=1.86z^1). 3. The method of claim 2, wherein the high-pass digital filter has a cutoff frequency of about 1200 Hz. 4. The method of claim 3, wherein the measurement of maximum energy in each frame occurs for the sample of maximum amplitude. (5) Long-term average value of the maximum energy of the filter
5. A method according to claim 4, wherein determining _m is calculated at each current frame by applying a recursion relation of the form:
o_y_. X_m_o if larger than _o_l_d
_y=T_m・X_m_o_y_. _o_l_d+(1
-T_m). X_2 or the value of parameter X_2 is parameter X_m_o_y_. If it is smaller than _o_l_d, then X_m_o_y=T_d・X_m_o_y_. ＿o＿l
_d+(1-T_d). X_2 However, the value of X_2 is determined by two thresholds X_p_. s_u_p and X_p_. ＿i＿n
X_m_o equal to the value of the maximum energy sample X_p_h in each frame bounded between
＿y＿． _o_l_d is a long-term average value calculated in the immediately previous frame, T_m and T_d are time constants, and T_m is a larger time constant than T_d. (6) Average value of the maximum value of energy ratio R_m_o_y
is calculated in each current frame by applying a recursion relation of the form; that is, parameter R_2 is equal to parameter R_m_o_
y_. If it is larger than _o_l_d, R_m_o_y=T_m・R_m_o_y_. ＿o＿l
_d+(1-T_m)R_2 or parameter R_2 is parameter R_m_o_y_. If smaller than _o_l_d, R_m_o_y=T_d・R_m_o_y_
．． _o_l_d+(1-T_d)R_2However, R_m_
o_y_. _o_l_d indicates the long-term average energy ratio calculated in the previous frame. (7) The method of claim 6, wherein the four thresholds are calculated by applying the following relationship: SX_1 conversation=a. X_m_o_y+X_p_h_. ＿
i_n_fSX_1 noise=b. X_m_o_y+X_p
＿h＿． _i_n_fSR_1 conversation=a. R_m_o_
y+R_. _i_n_fSR_1 noise=b. R_m_o
___y+R_. _i_n_f However, a and b are constants. (8) The method according to claim 7, wherein ¥a¥=1.8 and ¥b¥=1.25. (9) Apparatus for detecting a speech signal in a signal corrupted by noise, comprising the following means: (a) in each frame, the maximum energy of the preamplified signal and the maximum energy of the filtered digital signal; The first means of calculating the ratio between the maximum energy. (b) A second means for calculating the maximum energy of the filtered signal and a long-term average value of the energy ratio. (c) third means coupled to the second means for calculating maximum and minimum adaptive thresholds for the filtered digital signal and energy ratio; (d) determining means coupled to the third means for determining whether an audio signal is present or absent in the digital signal; 10. The apparatus of claim 9, wherein each of said first, second, third, and fourth means is formed by microprogrammed computing means. 11. The apparatus of claim 10, wherein said microprogrammed calculation means are formed by a signal processor.