JP2539027B2 - Voice detection method - Google Patents

Abstract

PURPOSE:To detect a consonant effectively by regarding a section other than three sections, derived from the nature of a voice, as a noise section and calculating a zero-cross different frequency threshold value successively from the zero-cross frequency distribution in the noise section. CONSTITUTION:An input sample signal is inputted to a power calculation part 11 and a zero-cross frequency calculation part 21 to calculate power P(i) and a zero-cross frequency Z(i) for each specific block length. The power P(i) calculated by the power calculation part 11 and a power threshold value Pth(i) outputted by a power threshold calculation part 13 are inputted to a detection part 12, which outputs a power detection signal Jp according to the large-small relation between the both. The zero-cross frequency Z(i) which is calculated by the zero-cross frequency calculation part 21 and the zero-cross frequency threshold value Zth(i) which is outputted from a zero-cross frequency threshold calculation part 23 are inputted to a zero-cross frequency detection part 22, which outputs a zero-cross frequency detection accumulation signal Jz according to the large-small relation between the both. Consequently, the constant can be detected without reference to the nature of a noise and without increasing malfunction due to a noise.

Description

TECHNICAL FIELD The present invention relates to a voice detection system used in a digital voice insertion system or voice packet communication system in the field of digital communication.

[Prior Art] Conventionally, as a voice detector applied to a voice section detection system, one shown in FIG. 2 is known. According to FIG. 2, of the input sample signal A input to the input terminal 1, a signal A1 having a relatively large amplitude such as a vowel is input to the amplitude detection unit 2, and a signal A2 due to the frictional consonant is DC. After the offset is removed by the suppression circuit 4, a constant value a is added and the sign bit thereof is extracted and input to the zero-crossing detector 3.

In the amplitude detector 2, the comparison circuit 2a compares the absolute value of the signal A1 with the predetermined value θ, and the result is counted by the counter.
2b is increased or decreased, and the counter value of this counter 2b becomes the threshold THv.
When it becomes larger, the output αv from the threshold circuit 2C is output to the OR circuit 5 at a high level “1”.

On the other hand, the zero-crossing detector 3 determines whether the sign bit input by the OR circuit 3a matches the sign bit one sample before, and the counter 3b is incremented or decremented depending on whether the result is a match or a mismatch. This is equivalent to counting the number of times the input crosses (-a), and the counter value of the counter 3b is the threshold value.
When it becomes larger than THz, the threshold value circuit 3C outputs the output αz to the OR circuit 5 at a high level “1”.

The OR circuit 5 inputs the logical sum α of the outputs αv and αz from the threshold circuits 2C and 3C to the hangover control circuit 6, and the hangover control circuit 6 outputs the output α of the OR circuit 5 from the high level “1” to the low level. Even if the level becomes "0", the hangover time for continuing to output the high level "1" for a certain period of time is added, and the output terminal 7 outputs the output as α out. And, while this α out is at a high level “1”, it is judged to be sound, and if it is at a low level “0”, it is judged to be silent (Reference 1. 1976 IEICE General Conference,
1753 "A method of voice detection using zero-crossing frequency" Araseki
Table, Kazuo Ochiai).

[Problems to be Solved by the Invention] However, the method described above has the following problem because the zero-crossing threshold is a fixed value THz.

The frequency characteristics of noise vary depending on each subscriber line due to the influence of background noise and line noise on the subscriber side. When the influence of the line noise is strong, the noise has a white noise-like property, and conversely, when the background noise on the subscriber side is dominant, the noise has a property similar to the background noise. This background noise has a Hoth noise characteristic (Reference 2, "Audition and Speech", 3rd edition, p431-433 IEICE).

FIG. 3 shows an example of the zero-crossing distribution of vowels, consonants, white noise, and Hoth noise. As shown in the figure, when the Hoth noise characteristic becomes stronger, the maximum value of the number of zero crossings becomes smaller than that of white noise due to its low-frequency emphasis property, and more consonants with more number of zero crossings are detected. Is easier to do.

On the other hand, in the case of strong white noise, the number of zero-crossings is larger than Hoth noise and the maximum value is also large.For example, if this maximum value and a threshold value are used, it is disadvantageous in detecting consonants as compared with Hoth noise. Become.

As described above, the noise has various properties from white to Hoth, and the appropriate threshold value may change depending on the property. Therefore, in the case of a fixed threshold as in the past,
If a high threshold value is set assuming white noise, it will be a disadvantageous setting for detecting the consonant of the subscriber's voice on which Hoth noise is superimposed. On the other hand, if a low threshold is set assuming Hoth noise, there is a problem that the white line noise increases the probability that white noise will also be judged as a consonant in the subscriber line, resulting in increased malfunction. .

Since the object of the present invention is to adopt a fixed zero-crossing threshold value as in the conventional art, depending on the nature of noise, the threshold value is too high, which is disadvantageous for consonant detection, or conversely, the threshold value is too low. By eliminating the disadvantage of increasing malfunctions and setting the threshold adaptively according to the noise zero-crossing distribution, it does not depend on the nature of the noise and does not increase malfunctions due to noise, and In a situation where it is advantageous to detect a consonant such as Hoth noise, it is to provide a voice detection method capable of detecting a consonant without losing its advantage.

[Means for Solving the Problem] The voice detection method of the present invention calculates a power P _(i) of a predetermined block length of an input signal, and calculates the power P _(i) and the power threshold Pth _(i).
Power P _(i) is compared with the power threshold Pth _{(i) is} larger than the power determination means to determine the presence of voice, and the number of zero crossings Z _(i) of the predetermined block length NB of the input signal is calculated. zero-crossing times judging that number of zero-crossing times Z _(i) and zero-crossing frequency threshold Zth _(i) is compared with the number of zero-crossing times Z _(i) is a sound presence judgment is greater than the zero-crossing count threshold Zth _(i) Means, a sound determination unit that creates a voice determination as a voice detection signal when any one of the power determination unit and the zero-crossing number determination unit is a voice determination, and the power threshold Pth _(i) As a power threshold calculation means for giving a threshold value that adaptively changes to the average power Pav of the silent judgment section by the power judgment means, and the zero crossing number threshold Zth _(i) , the judgment by the power judgment means is a silent judgment, and Starting from the time when the determination by the zero-crossing number determination means becomes a voiced determination,
When there is sound due to power judgment within the first predetermined time,
The first section is from the start point to the time when the sound is judged, and the second section is the section which is judged as the sound by the power judgment, and the silence after the sound judgment by the power judgment is changed to the silence judgment. Within the first second predetermined time of the judgment section, the third
And a section other than the first, second, and third sections as a noise section, and in this noise section, a zero-crossing number threshold value Zth _(i) sequentially calculated based on the zero-crossing number distribution in the noise section. ₎ Which gives a zero crossing frequency threshold value calculating means.

[Operation] When the input sample signal is input, the presence / absence of sound is determined by determining the power P _(i) of the predetermined block length NB, the number of zero crossings Z _(i) , the power threshold Pth _(i) , and the number of zero crossings thresholds. It is determined as follows using Zth _(i) .

When P _(i) ＞ Pth _(i) or Z _(i) ＞ Zth _(i) , voiced P _(i) ≤Pth
_{When (i)} and Z _(i) ≤ Zth _(i) , there is silence, where the power threshold Pth
_(i) is adaptively set based on the power level in the section where the power judgment is silent (P _(i) ≤ Pth _(i) ).

On the other hand, the zero-crossing threshold Zth is updated so as to be more adaptive to the zero-crossing distribution of noise, and the updated interval is defined by the following characteristic (i) or (ii) by utilizing the nature of the voice. Becomes

(I) N blocks of continuous silence determination.

In this section, in order to improve the consonant detection capability, the zero-crossing number threshold Zth is updated to a smaller value unless the property of noise changes.

(Ii) A section in which the maximum value of the number of zero-crossings is not updated during N blocks after the sound is once determined by the number of zero-crossings determination, and there is no sound during the power determination during that period.

In this section, since it is determined that noise has occurred due to noise, the zero-crossing number threshold Zth is updated to a larger value so as to prevent malfunction due to noise that makes such speech determination.

However, (i) and (ii) are not applicable at a certain time immediately after the power determination transits from voiced to silence determination, because there is a possibility of a consonant portion occurring at the time of talking, ending, or the like.

In this way, the zero-crossing frequency threshold is set adaptively according to the noise zero-crossing frequency distribution, so that the detection sensitivity is not influenced by the noise zero-crossing frequency distribution.

[Embodiment] An embodiment of the present invention will be described below with reference to FIGS. 1 and 4.

FIG. 1 is a block diagram showing an embodiment of a voice detector for carrying out the method of the present invention.

The input sample signal is input from the input terminal 1 to the power calculation unit 11 and the zero-crossing number calculation unit 21, and the power P _(i) and the zero-crossing number Z _(i) for each predetermined block length are calculated. Here, the predetermined block length refers to the length of a section obtained by dividing the time axis of the input sample signal at predetermined intervals, which will be represented by NB hereinafter.

 Further, i means the block number.

The power P _(i) calculated by the power calculator 11 and the power threshold Pth _(i) output by the power threshold calculator 13 are input to the power detector 12, and the power P _(i) and the power threshold Pth _{( The} power detection signal Jp is output according to the magnitude relationship with _i) . Here, the power threshold calculation unit 13 calculates the power threshold Pth _(i) based on the generation pattern of the power P _(i) and the zero-crossing number detection signal Jz and the power detection signal Jp described later.

On the other hand, the number of zero crossings calculated from the zero crossing number calculation unit 21.
Z _(i) and the zero-crossing number threshold Zth _(i) output from the zero-crossing number threshold calculating unit 23 are input to the zero-crossing number detecting unit 22, and the zero-crossing number Z _(i) and the zero-crossing number threshold Zth _{( The} zero-crossing number detection signal Jz is output according to the magnitude relationship with _i) . Here, the zero-crossing number threshold calculation unit 23 calculates the number of zero-crossings Z _(i) and the power detection signal Jp.
And the zero-crossing number threshold Zth _(i) is calculated based on the generation pattern of the zero-crossing number detection signal Jz.

Then, the power detection signal Jp and the zero-crossing number detection signal Jz are input to the sound / silence determination section 31, and the sound detection signal J is taken out from the output terminal 2.

The power calculation unit 11 and the power detection unit 12 described above constitute the power determination means of the present invention, and the zero-crossing number calculation unit 21 and the zero-crossing number detection unit 22 constitute the zero-crossing number determination means of the present invention. .

Next, the processing and operation in each of the above-mentioned units will be described in (A)-
This will be described in detail in (G).

(A) The power calculator 11 calculates the power P of the input sample signal for each predetermined block length (NB) according to the equation (1).

However, NB is the number of samples in a predetermined block length, S
_(j) is the input sample value.

(B) The power detection unit 12 uses the power P _(i) and the current (currently set) power threshold Pth _(i) to generate the power detection signal Jp according to the equation (2).

(C) The power threshold calculation unit 13 updates the power threshold Pth by the following method based on the power detection signal Jp.

That is, it is determined that there is no sound by the power detection signal Jp (Jp
= 0) Every time a block is generated more than a predetermined number of blocks NBP (NB means P), the average power Pav of the section is increased.
And a fixed coefficient α (α>
It is calculated according to equation (3) using 1).

Pth = α · Pmin (3) (D) The zero-crossing number calculation unit 21 counts the number of sign inversions of the input sample signal generated in this block for each predetermined block length NB, and counts the number of zero-crossing numbers. Output as Z _(i) .

(E) The zero-crossing frequency detection unit 22 creates the zero-crossing frequency detection signal Jz according to the equation (4) using the zero-crossing frequency Z _(i) and the current zero-crossing frequency threshold Zth _(i) .

(F) The zero-crossing frequency threshold calculation unit 23 updates the zero-crossing frequency threshold Zth by the method (i) or (ii) described later.

(G) The voiced / non-voiced sound determination unit 31 finally determines the voiced / non-voiced sound according to the equation (11).

Now, in the zero-crossing number threshold calculation unit 23, the zero-crossing number threshold Zth
The two ways to update are:

(I) It is determined that there is no sound by the power detection signal Jp (P _(i) ≤ Pt
h _(i) ) and the number of blocks (Z _(i) ≤ Zth _(i) ) that have been determined to be silent by the zero-crossing detection signal Jz continuously occur for a predetermined number of blocks NBZ (NB means Z). For each determination section, the zero crossing number threshold Zth _{(i) is} calculated by the processing of the following equations (6) to (9 _).
Is calculated.

_{Zth (i + 1) = Zmin} (i) + D (i) ... (9) where, Zav _(i) the average number of zero-crossing times of the interval, Zmin _(i) the minimum value of the average number of zero-crossing times, D _{(i )} Is the fluctuation absorption offset value of the number of zero crossings, ε is a significant difference empirically determined, the setting value of the fluctuation range of Zav _(i) , and Do is the initial value of D _(i) .

The reason why the average number of zero crossings Zav _(i) is calculated by the equation (6) is to know the center of change in the number of zero crossings that varies in the silence determination section. Here, since a section that is a predetermined number (j) before the currently focused block (i) is always used as a section for capturing the average, noise generated instantaneously by averaging in this section It is prevented that the noise is erroneously determined and the sound is erroneously determined to be voiced.

The reason why the minimum value Zmin _(i ) of the average number of zero-crossings is calculated by the formula (7) is to increase the detection capability by setting the minimum value.

The offset value D _(i) is calculated by the equation (8) in order to absorb the fluctuation of the number of zero crossings and to cut off the influence of the fluctuation. The offset value absorbs undulation noise with a small change rate, rather than instantaneous noise with a large change rate.

In formulas (7) and (8), the minimum value Zm of the average number of zero-crossings
The absolute value of the difference between in _(i-1) and the average number of zero crossings Zav _(i) is taken, and the absolute value of the difference is Zmin _(i) , D _(i)
The possible values of are divided into cases because if the absolute value fluctuates up and down with reference to the set value ε, it is assumed that the noise property in that section has changed, and that it corresponds to the changed noise property. Is.

Further, in the equation (7), when the absolute value is ε or less, the minimum value Zmin _{(i -1)} or the average value Zav _{(i) of the} immediately preceding block is
Among them, the smaller one is selected in order to further improve the detection capability because the property of noise does not change. Similarly, in equation (8), when the absolute value is ε or less, the offset value D _(i-1) of the immediately preceding block is selected. For the same reason as above, the offset value of fluctuation absorption is lowered to improve the detection capability. This is to increase.

In equations (7) and (8), when the absolute value is larger than ε, the minimum value Zmin _(i) and the offset value D _(i) are initialized because the property of noise has changed. To return to.

The reason why the zero-crossing threshold value is obtained by adding Zmin _(i) and D _{(i) in the} equation (9) is to set an optimum threshold value according to noise.

(Ii) If Z _(j) > Zth _(j) , then Z _(k) from that point onward.
And Zth _(k) , the maximum value Zmax of the number of zero crossings that is not updated occurs for a predetermined number of blocks, and during that interval, the block that is determined to be voiced by the power detection signal Jp If none occur, the number of zero crossings in this section
While calculating equations (6) and (7) using Z _(k) , the offset value D _(i) is calculated using equation (10) using the maximum value Zmax of zero crossings, and is calculated using equation (11). Set the crossing threshold Zth.

Zth = Zmax (11) In equation (10), when the absolute value is ε or less, the difference between Zmax and Zmin is obtained. Zero-crossing threshold when shifting to the condition
This is because Zth _{(i + 1)} starts from the same Zmax as ₍ ii) (substituting the result of equation (10) into equation (9) yields Zth = Zmax) to smooth the transition. Similarly, the reason why the offset value is reset to Do when the absolute value is larger than ε in the equation (10) is to cope with the change in the property of noise.

The reason why the zero-crossing number threshold Zth is set to Zmax in the expression (11) is to prevent the noise from being determined by noise by increasing the threshold so far.

It should be noted that the calculation of the expression (10), which seems to be unnecessary in calculating the expression (11), is used when the condition (ii) is deviated to the condition (i). This is because the offset value D _(i) is obtained in advance.

However, the above methods (i) and (ii) are performed in the section of the first predetermined number of blocks NBZB (NB means ZB) of the silence determination section after the power detection signal Jp is changed from the sound determination to the silence determination. Prohibited section) does not apply.

 The intention of the above processing will be described below.

The purpose of the above-described updating process of the zero-crossing number threshold Zth is to estimate the maximum value Zmax of the zero-crossing number of noise and set it as the zero-crossing number threshold Zth.

Therefore, the problem is not to set the zero-crossing threshold Zth to an unnecessarily large value in a consonant portion that may have a larger number of zero-crossings than noise. Also, by setting the zero-crossing threshold Zth to a small value in a part such as a vowel part where the number of zero-crossings is relatively smaller than noise, it is possible to prevent the noise part from being mistakenly recognized as voiced. It is to be.

Therefore, it is possible to update the threshold value of the number of zero crossings in the noise section by using the following property of the voice. Note that FIG. 4 is referred to here for easy understanding. Therefore, FIG. 4 will be described first, and then the nature of voice will be described.

FIG. 4 shows an example of the operation of setting the zero-crossing threshold value according to this embodiment. FIGS. 4 (a) and 4 (b) show time changes of the block power, the power threshold Pth, and the zero-crossing number threshold Zth, for example, when a voice (including noise 5) is input, and FIG.
(D) shows an example of time change when only noise is input.

In FIG. 4, the method (i) is applied to the section (I), and the method (ii) is applied to the section (II). The section (III) is the above (i), (i
i) is a prohibited section that does not apply. Also, (i) (i
Suffixes i, j, k used in i) are, for example, in FIG. 4 (d), if the time point of interest is time T11, this time point is i, and the time points before this time point are i. The area of j, and the area after that is the area of k.

In FIGS. 4 (a) and 4 (b), Z _(i) > Zth _(i) at time T1, Zmax is updated after time T1, and at time T3 when the sound is judged by power judgment, zero crossing at time T2. The maximum number was Zmax. Since T3-T2 is shorter than the constant time TF, the zero-crossing number threshold Zth remains held.

The period from time T4 to T4 + TE (constant time), that is, time T5, which is the beginning of the section in which there is no sound according to the power determination, is a prohibited section for the processing related to the update of the zero-crossing number threshold Zth in this section.

In FIGS. 4 (c) and (d), Z _(i) > Zth _(i) at time T11, the maximum number of zero crossings is updated, and the number of zero crossings at time T12 becomes the maximum value Zmax, which value is constant time. TF retained (no zero crossing above it). Moreover, because the section power judgment did not produce sound, time T13
Then, the zero-crossing number Zmax at time T12 is set as the zero-crossing number threshold Zth.

In the present embodiment, the following properties of voice are used.

The consonant (symbol W in FIG. 4 (b)) at the beginning of the speech has a small power, but after the consonant part is finished, the vowel (fourth
In many cases, the symbol X) in FIG. A vowel has a larger power than a consonant, and that portion can be determined by the power detection signal Jp. Therefore, if the zero crossing frequency detection unit 22 produces a sound in the consonant portion, the vowel portion X appears within the fixed time TF, and the power detection unit 12 produces a sound portion.

On the contrary, the fixed time T after the voice determination by the number of zero crossings is detected.
When there is no voiced part due to power detection in F (Figs. 4 (c) and (d)), the voiced determination by the number of times of zero crossings is due to noise (symbol Y in Fig. 4 (d)). Therefore, the value Zmax of the number of zero crossings is set as a new threshold value.

In addition, since a vowel is not always accompanied in a consonant portion generated during talking or ending of a word with respect to the beginning of a speech, a silent determination section by power of a certain time TE after the voice determination by power detection (FIGS. 4A and 4B). In (), since the part V may be a consonant part, the updating process of the threshold value Zth of the number of zero crossings is not performed.

About 200 msec or more may be estimated as the value of the duration time TF, TE of the consonant (above reference 2.p290-289).

Furthermore, when the property of noise changes, the zero-crossing number distribution changes, so the average zero of the NBZ block that produces a difference of ± ε or more than the minimum value Zmin of the average zero-crossing number Zav of the NBZ block of the noise so far. When the average value Zav of the number of crossings is generated, the section is considered to be a noise section with changed properties, and the minimum number of zero crossings Zmin and the offset value D _(i) are initialized.

As described in detail above, according to the present embodiment, the power calculation unit 11 that calculates the power P of the predetermined block of the input sample signal and the zero crossing number calculation unit 21 that calculates the zero crossing number Z.
A power threshold calculation unit 13 that calculates the power threshold Pth, and a zero-crossing number threshold calculation unit 23 that calculates the zero-crossing number threshold Zth,
The power P is compared with the power threshold Pth, and the power P is the power threshold Pth.
Power detection unit 12 that outputs a power detection signal Jp that indicates sound when the value is large, and a zero crossing that outputs a zero crossing number detection signal Jz that indicates sound when the number Z of zero crossings is larger than a threshold value Zth of zero crossings. The number-of-times detection unit 22, a sound / non-sound determination unit 31 that outputs a sound determination as a sound detection signal J if any one of the above detections is a sound tail, and the power detection unit 12 detects the sound. Every time a block occurs only in the first predetermined block, the average power Pav of the section is calculated, and the minimum value Pm of the average power Pav is calculated.
Provide a power threshold calculation unit 13 that updates the power threshold Pth ₍ with a value obtained by multiplying this by a constant value α, and a zero-crossing count threshold calculation unit 23 that calculates the following zero-crossing threshold Zth. Are configured.

In the zero-crossing number threshold calculation unit 23, blocks NB that are silent by the power detection unit 12 and are silent by the zero-crossing number detection unit 22 continuously generate the first predetermined block number NBZ, and the first block number NBZ is generated. In the silent section in which the predetermined block number NBZ has occurred, when the power determination is changed from the sound determination to the silence determination and the silence detection is continuously performed for the second predetermined block number TE or more, the first predetermined number of the section is determined. If the absolute value of the difference between this value Zav and the minimum value Zmin of the average number of zero crossings is less than or equal to a predetermined fixed value ε, the noise property changes. It is determined that the number has not occurred, and the average zero-crossing Z of the silent section in which the first predetermined number of blocks NBZ has occurred
av and the minimum value Zmin of the mean zero crossings are updated to the new minimum value Zmin of the number of mean zero crossings.
The sum of in and the current offset value D for absorbing the fluctuation of the number of zero crossings is newly set as the threshold value Zth of the number of zero crossings.

On the other hand, the average number of zero crossings Zav and the minimum value of the average zero crossings Zmi
When the absolute value of the difference from n is larger than the fixed value ε, it is judged that the property of noise has changed, and the minimum value Zmin
Is reset by the average zero-crossing count Zav in the silent section in which the first predetermined number of blocks NBZ has occurred, and then the zero-crossing count threshold Zth is updated by the sum of the minimum value Zmin of the average zero-crossing counts and the offset value D.

As described above, in the silence determination section, unless the property of noise changes, the zero-crossing number threshold Zth is reduced, and in the consonant part that may have a larger number of zero-crossing numbers than noise, the zero-crossing number threshold Zth is more than necessary. Since it is not set to a large value for, the ability to detect consonants can be improved.

On the other hand, when the noise property changes, the zero-crossing threshold
Since Zth is set to a reset value and initialized so that it can be adapted to the changed noise, the ability to detect consonants does not deteriorate even under changed noise.

Further, when the zero-crossing number Z becomes larger than the zero-crossing number threshold value Zth, the maximum value Zmax of the zero-crossing number that is not updated after the first predetermined block number NBZ is thereafter irrespective of the magnitude relation between the zero-crossing number and the threshold value. A block that is determined to be sound by power detection during a section in which a maximum value Zmax of the number of zero-crossings that has occurred and has not been updated for the first predetermined number of blocks NBZ has occurred
In the case where no NB is generated and the section in which the maximum value Zmax is generated has changed from the sound determination to the silence determination from the power determination and the silence determination has occurred for the second predetermined block number TE or more, Since it is considered that the voiced judgment by detecting the number of zero crossings is due to noise, the average number of zero crossings Zav and the minimum value of the average zero crossings Zmin in the section where the maximum value Zmax occurs.
The maximum value of the number of zero-crossings Zm regardless of the magnitude of the absolute value of the difference between
Let ax be a new zero-crossing threshold Zth.

In this way, under a certain condition, when the voiced judgment is made only by detecting the number of zero-crossings, consider the noise and set the threshold Zth of the zero-crossings to the maximum value Zmax, and set the threshold to a large value. Therefore, by setting the zero-crossing number threshold Zth to a small value in a part such as a vowel part where the number of zero-crossings has a relatively smaller value than noise, the noise part is erroneously determined to be voiced. Never.

As described above, according to the present embodiment, it is possible to determine the zero-crossing threshold value according to the zero-crossing distribution of noise from the input signal in which the noise is added to the voice. Moreover, even if the noise property changes, a threshold value can be set according to the changed new zero-crossing distribution of noise.

Next, the simulation result according to the present embodiment will be described.

FIG. 5 shows an example of the zero-crossing number threshold characteristic according to the present embodiment when the power ratio R (= P _H / (P _W + P _H )) of the Hoth noise power P _H and the white noise power P _W is changed. . It can be seen that the stronger the Hoth characteristic (R → 1), the lower the threshold value, which makes consonant detection more advantageous.

Also, in FIG. 6, R = 1, that is, a voice with Hoth noise added,
The simulation results of this example are shown for English "Four" (Fig. 6 (a)).

As a comparative example, a conventional fixed threshold method based on white noise is shown. Here, the block length is 8 msec, N = 32,
Noise level −50 dB0, voice level −33 dB0 (consonant part −49 dB
0). It can be seen that dropout of consonant sounds with a fixed threshold value according to the conventional example can be prevented by adapting the threshold value according to the present embodiment (FIG. 6 (b)).

As described above, according to the present embodiment, the loss of the initial consonant is eliminated, so that the reliability of the communication can be further improved when applied to a digital voice insertion system or a voice packet communication system especially in the field of digital communication.

[Effects of the Invention] The present invention has the following effects.

(1) In the speech detection method according to claim 1, in a section other than the three sections derived from the nature of the speech, that is, a vowel section determined to be voiced by the power judgment, and a vowel often followed. A consonant section and a section other than a consonant section that is not necessarily accompanied by a vowel, such as a consonant section that occurs at the end of a speech, is used as a noise section, and the zero-crossing threshold value is sequentially calculated in this noise section based on the zero-crossing frequency distribution in the noise section. Therefore, compared to the conventional case where the threshold value is fixed, the threshold value is too high, which is disadvantageous in the detection of consonants, or conversely, the threshold value is too low, which increases malfunction due to the nature of noise. The consonant can be effectively detected without being brought.

(2) A zero-crossing frequency threshold value is determined from the input signal in which noise is added to the voice in accordance with the noise zero-crossing frequency distribution, and the change in the noise property is detected to detect a new zero-crossing noise. Since the threshold is set again according to the frequency distribution so that the threshold is adapted to noise, compared to the conventional case where the threshold is fixed, due to the nature of noise, the threshold is too high and consonants are generated. There is no disadvantage to the detection of noise, and conversely, the threshold is too low to cause an increase in malfunction, and it does not depend on the nature of noise, does not increase the tail motion against noise, and is similar to Hoth noise. In a situation where it is advantageous to detect a consonant, the consonant can be effectively detected without losing its advantage.

Also, in particular, when no voiced part is generated by power detection within a certain time after the voiced determination by the zero-crossing number detection, the voiced determination by the zero-crossing number detection is often due to noise. By using the property of, in that case, the zero-crossing threshold is set to the maximum value.
Consonants can be effectively detected without increasing malfunction due to noise.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a voice detector for implementing a voice detection system according to the present invention, FIG. 2 is a block diagram showing an example of a conventional voice detector, and FIG. 3 is a vowel, a consonant,
FIG. 4 is an explanatory diagram showing an example of each zero-crossing frequency distribution of white noise and Hoth noise, FIG. 4 is a timing diagram showing an operation example for explaining the zero-crossing frequency threshold setting according to the example of the present invention, and FIG.
FIG. 6 is an explanatory diagram showing an example of the zero-crossing threshold value characteristic according to the present embodiment when the power ratio of Hoth noise power and white noise power is changed, and FIG. 6 shows the present embodiment and the conventional example for speech to which Hoth noise is added. FIG. 6 is an operation explanatory view showing a simulation result comparing the above. Reference numerals 11 and 12 denote a power calculation unit and a power detection unit that constitute the power determination unit, 13 a power threshold value calculation unit, 21 and 22 a zero-crossing number calculation unit and a zero-crossing number detection unit that configure the zero-crossing number determination unit, and 23. Is a zero crossing frequency threshold calculating means, 31 is a voice / silence judging means, P is a power, Pth is a power threshold, Z is a zero crossing frequency, Z
th is the threshold value of the number of zero-crossings, Zav is the average number of zero-crossings, Zmin is the minimum value of the average number of zero-crossings, Zmax is the maximum value of the number of zero-crossings, J
Is a sound detection signal, T1 is a start point, S1 is a first section, S2 is a second section, S3 is a third section, NB is a predetermined block length, NBZ
Is a first predetermined block number, TE is a second predetermined time or a second predetermined block number, TF is a first predetermined time, ε is a predetermined fixed value, and D is a current value for absorbing fluctuations in the number of zero crossings. The offset value and Do are fixed values.

 ─────────────────────────────────────────────────── ─── Continued Front Page (72) Inventor Yasuo Shoji 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (56)

Claims (2)

(57) [Claims] 1. A power determining means for calculating a power of a predetermined block length of an input signal, comparing the power with a power threshold value, and making a voiced determination when the power is larger than the power threshold value, and a predetermined block length of the input signal. The number of zero crossings is calculated, and the number of zero crossings is compared with a threshold value of zero crossings, and when the number of zero crossings is greater than the threshold value of zero crossings, the number of zero crossings is determined to be voiced. When either one of the zero-crossing number determination means is a voiced determination, a voiced / non-voiced determination means for creating a voiced determination as a voiced detection signal, and the power threshold of a silence determination section by the power determination means. A power threshold calculation means for giving a threshold that adaptively changes to the average power, and a time point at which the determination by the power determination means is a silent determination and the determination by the zero crossing number detection means is a voice determination as the zero-crossing number threshold. Starting point and Then, when the sound is determined by the power judgment within the first predetermined time, the section from the start point to the time when the sound is determined is the first section, and the section determined to be the sound by the power judgment is the first section. The second section is defined as the second section, and the first, within a second predetermined time, of the silence determination section after the sound determination by the power determination is changed to the silence determination is defined as the third section, and the first, second, and third sections are included. A section other than the section is defined as a noise section, and a zero-crossing number threshold calculation unit that gives a zero-crossing number threshold value that is sequentially calculated based on the zero-crossing number distribution in the noise section in the noise section is provided. Voice detection method. 2. The zero-crossing number threshold calculating means in the voice detecting method according to claim 1, wherein the power determination is a silence determination and the block which is a silence determination by the zero-crossing number determination produces a first predetermined number of blocks. In addition, when the power determination changes from the sound determination to the silence determination in the silent section in which the first predetermined number of blocks have occurred, the second predetermined block number or more silence determination continuously occurs If the absolute value of the difference between this value and the minimum value of the average number of zero-crossings is equal to or less than a predetermined fixed value, the silence generated in the first number of the predetermined blocks is calculated. The smaller value of the average zero-crossing count and the minimum value of the average zero-crossing count of the section is updated as a new minimum value of the average zero-crossing count, and this minimum value and the current offset value for absorbing fluctuations of the zero-crossing count are set. Also added Is used as a new zero-crossing frequency threshold value, and when the absolute value of the difference between the average zero-crossing frequency and the minimum value of the average zero-crossing frequency is larger than a fixed value, the minimum value of the average zero-crossing frequency is set to the first value. After resetting the average number of zero crossings in the silent section generated by a predetermined number of blocks and resetting the offset value with a predetermined fixed value, the sum of the minimum value of the reset average zero crossings and the offset value A means for updating the threshold value of the number of crossings, and a zero crossing that is not updated for a first predetermined number of blocks after that, regardless of the magnitude relation between the number of the zero crossings and the threshold value, when the number of zero crossings becomes larger than the threshold value for the number of zero crossings. In the section in which the maximum value of the number of times occurs and the maximum value of the number of zero-crossings that is not updated for the first predetermined number of blocks occurs, no block is determined to be voice by the power determination, and maximum In the case where the number of silence determinations is equal to or greater than the second predetermined block number after the sound determination is changed from the power determination to the silence determination in the section in which the power generation determination is performed, the average of the first predetermined block numbers in the section in which the maximum value is generated is generated. If the absolute value of the difference between the number of zero crossings and the minimum value of the average zero crossings is equal to or less than a predetermined fixed value, the smaller one of the average zero crossings and the minimum value of the average zero crossings is updated as the minimum value, A means for newly setting the difference between the maximum value of the number of zero crossings and the minimum value of the average number of zero crossings as an offset value, and setting the maximum value of the number of zero crossings as a new threshold value of the number of zero crossings, and the average number of zero crossings and the average number of zero crossings When the absolute value of the difference from the minimum value of is larger than the fixed value, the minimum value of the average zero-crossing is reset by the average number of zero-crossings, the offset is reset to a fixed value, and the maximum value of the number of zero-crossings is updated. A zero crossing frequency threshold Speech detection method, characterized in that.