JP3375655B2 - Sound / silence determination method and device - Google Patents

Abstract

PURPOSE:To improve the precision of judging voiced/voiceless in a speech while using a low-precision judging rule. CONSTITUTION:Feature parameters extracted from a frame-divided speech signal are decided by 1st and 2nd many-valued logical decision parts 2 and 3 according to their likelihoods. An inference part 4 infers whether the speech signal is voiced or voiceless by using the outputs of the 1st and 2nd many-valued decision parts 2 and 3 and a decision result feedback part 7. A variable hangover generation part 5 varies a hangover time according to the result of the inference part 4 and a binary decision part 6 judges voiced or voiceless finally.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for determining the presence / absence of voice for use in digital voice communication and the like, and a device therefor.

[0002]

2. Description of the Related Art In recent years, in mobile communication such as car phones and mobile phones, a method of interrupting the transmission in a silent section of a voice is required in order to reduce power consumption. Development of methods is desired.

A conventional method for determining whether or not a sound is present will be described below. FIG. 9 is a diagram showing a conventional sound / soundlessness determination device. In FIG. 9, 9 is a parameter extraction unit, 10, 11
Is a first and second binary logic decision unit, 12 is a third binary logic decision unit, 13 is a hangover occurrence unit, and 14 and 15 are 1-frame delay units.

With the above-mentioned configuration, first, the parameter extraction unit 9 extracts some characteristic parameters such as power and the number of zero crosses, which are useful for determining the presence / absence of voice, from the input voice divided into frames. Next, according to the determination rule of each of the first and second binary logic determination units 10 and 11, binary determination of voiced / unvoiced is performed using a threshold value. For example, the first binary logic determination unit 10 performs binary determination based on the magnitude of power, and the second binary logic determination unit 11 performs binary determination based on the number of zero crosses. In the third binary logic determination unit 12, the first and second binary logic determination units 10, 1
1 determination result and 1 frame delay units 14 and 15
Based on the determination result of the previous frame output from
The presence or absence of voice is determined using a binary logic operation. In the hangover generation unit 13, when the silence determination by the third binary logic determination unit 12 continues for several frames, the final determination is transitioned from voiced to silence. The 1-frame delay unit 1
The feedback of judgment by 4 and 15 is used if necessary.

[0005]

However, in the above-described conventional voiced / soundless determination method, since there is no highly accurate parameter extraction method and highly accurate determination rule, it is not possible to perform determination by binary logic with a clear determination threshold value. However, there is a problem that an error is likely to occur in the determination of the determination rule, and an error is also caused in the final determination of voiced / unvoiced sound.

The present invention solves the above-mentioned problems of the related art. Even when a feature parameter and a determination rule that are not accurate, such as the power of voice and the number of zero crosses, are used,
It is an object of the present invention to provide an excellent voiced / unvoiced determination method capable of obtaining a more reliable determination result in the final determination.

[0007]

In order to achieve the above object, the present invention uses multi-valued logic having continuous values in the range of 0 to 1 in the judgment process on the way, for example, 0 is " Inference is performed using the maximum value and the minimum value of the values meaning "silent", 0.5 is "undecidable", and 1 is "voiced". The value judgment is performed.

Further, in each judgment, a data table recording the input / output relationship is provided so that a multivalued logic output can be obtained with a small processing amount even when the judgment characteristic is non-linear.

Further, for the uncertain silence determination, a long hangover time is generated to reduce the error of determining the sound as silence.

[0010]

With the above construction, the present invention does not perform clear threshold value processing in each judgment rule, outputs a value according to the certainty of the judgment, and performs processing according to the certainty in the final judgment. . That is, a more probable determination result is obtained by using a plurality of rules and using the most probable determination output among them.

Further, since the data table having the data table recording the input / output relation is provided, the processing can be performed with a small processing amount even when the parameter and the determination output have a non-linear characteristic.

Further, in the silent determination with low certainty, the hangover time is lengthened to delay the transition time from the final voice determination to the silent determination, thereby reducing the rate of false determination of voice. .

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a diagram showing a configuration for realizing the voiced / unvoiced determination method of the present invention. In FIG. 1, reference numeral 1 is a parameter extraction unit that extracts one or a plurality of characteristic parameters effective for determining whether a sound is present or not from voice data divided into frames. Reference numerals 2 and 3 are first and second multi-valued logic determination units that perform voiced / non-voiced determination according to their respective determination rules and output a determination result by multi-valued logic based on continuous values within the range of 0 to 1. Reference numeral 4 is an inference unit based on multivalued logic that infers a more reliable result from a plurality of determination results. A variable hangover generation unit 5 generates a variable hangover time according to the certainty of the determination result. Reference numeral 6 is a binary determination unit that finally performs a binary determination of whether there is sound or no sound. Reference numeral 7 denotes a determination result feedback unit that delays the determination result by one frame and feeds it back to the inference unit 4.

A case will be described in which the presence / absence of a sound is determined in the above configuration. The following equation is used in each step of this embodiment.

If power is large then voice ... rule If power is low then silence ... rule If (small zero cross number and previous frame is voice) then voice ... rule First, in the parameter extraction unit 1, the frame is extracted. From the voice of the j-th frame having a length of n (X _j (i); 0 ≦ i ≦ n−1), the power P _j and the number of zero crosses Z _j are set as the characteristic parameters.
Is obtained, and as a result, P _j and Z _j shown in FIG. 2 are obtained.

FIG. 3 (a) is a diagram showing the internal structure of the first multi-valued logic decision section 2, 2a being the input power P.
₆ is a data table that defines the relationship between _j and the output voiced / unvoiced determination value d _1j . 2b is the input power P _j
The data table reading unit reads out the corresponding voiced / unvoiced determination value d _1j from the data table in accordance with the above.
FIG. 3B shows the relationship between the power P _j and the voiced / unvoiced judgment value d _1j in the data table 2a. In the first multi-level logic determination unit 2, rules, a decision by, the data table reading section 2b, reads the data table 2a in accordance with the input power P _j, voice activity detection shown in FIG. 2 The value d _1j is obtained and output to the inference unit 4.

FIG. 4 (a) is a diagram showing the internal structure of the second multi-valued logic judgment section 3, and 3a shows the number of input zero crosses Z _j and the output sound / sound judgment value d _20j . It is a data table which defines a relationship. Reference numeral 3b is a data table reading unit for reading the corresponding sound / non-sound determination value d _20j from the data table 3a according to the input zero-cross number Z _j . FIG. 4B illustrates the relationship between the number of zero crosses Z _j and the voiced / unvoiced judgment value d _20j in the data table 3a. In the second multi-valued logic determination unit 3,
For the determination of "the number of zero crosses is small" in the rule, the data table reading unit 3b reads the data table 3a according to the input number of zero crosses Z _j , and obtains the voiced / unvoiced determination value d _20j shown in FIG. And outputs it to the inference unit 4.

FIG. 5A shows the judgment result feedback section 7
7a is a diagram showing the internal configuration of the output of the inference unit 4 d_j
The 1-frame delay unit 7c based on
Output d of the inference unit 4 in the game_j-1And the sound that is output
Silence judgment value @_21jIn the data table that defines the relationship with
is there. 7b is the input output d of the 1-frame delay unit 7c
_j-1According to the data table 7a
Sound silence judgment value @_21jRead Data Table Read
It is a department. In this judgment result feedback unit 7, the rule
Data table 7
The output d of the inference unit 4 of the previous frame according to b_j-1In response to the
The data table 7a is read, and the sound
Silence judgment value @_21jIs obtained and output to the inference unit 4.

FIG. 6 shows the internal structure of the inference unit 4. In FIG. 6, reference numeral 4a denotes a pre-calculation unit that calculates the following equation based on the output d _20j of the second multi-valued logic judgment unit 3 and the output @ _21j of the judgment result feedback unit 7. 4b
Is the output d _1j of the first multi-valued logic decision unit 2 and the pre-calculation unit 4a
Of the outputs d _2j and 0.5 of the maximum value detector 4c, the maximum value detector 4c outputs the output d _1j of the first multi-valued logic decision unit 2 and the outputs d _2j and 0.5 of the pre-calculator 4a. It is a minimum value detection unit that outputs the minimum value.

In the inference unit 4 having the above structure, the first scan
The rule "and" is executed as the step by the following formula
As a result of the rule, the voiced and silent judgment shown in FIG.
Constant value d _2jTo get

D _2j = @ _21j × (d _20j −0.5) +0.5 In the next step of the inference unit 4, all of d _1j , d _2j , and 0.5 are detected by the maximum value detection unit 4 b. Minimum value detector 4
It is given to c, and each output value is added. From this value 0.
5 is subtracted to obtain a voiced / unvoiced judgment value d _j . The maximum value detection unit 4b and the minimum value detection unit 4c here have a function of obtaining the most probable sound determination value and silence determination value, respectively.

FIG. 7 is a diagram showing the internal structure of the variable hangover generation unit 5. In FIG. 7, 7a is given by the following equation using the output d _j = x of the inference unit 4 and the value s ^′ of the output s of the variable hangover generation unit 5 in the previous frame.
It is a calculation unit that generates an output s. Reference numeral 7b is a one-frame delay unit that outputs the output of the calculation unit 7a of the previous frame (= the output s of the variable hangover generation unit 5) to the calculation unit 7a. In the processing in the following equation, for example, the time constant A
_{When m} = 0.1 and A _p = 0.9, the output s = f (x, s ^′ ) of the variable hangover generator 5 is as shown in FIG. 8 (a). The value in FIG. 8A indicates the output s.

S = f (x, s^') = S^'+ A_m× (1-
x) × (x−s^') (However, x ≦ s ^') …… s = f (x, s^') = S^'+ A_pXxx (x-s^') (Ta
But x> s^') …… According to the formula, even if there are consecutive silent frames,
When a sound-like frame appears, it is easy to move to voice
However, on the contrary, it is difficult to shift from voice to silence.
It This is not available when voice communication is actually performed.
Rather than judging a sound as a sound
It is more affected by the sound being interrupted when it is judged as silence
Is large.

In addition, x is a constant value (x <
0.5), and an initial value (s ^' > which represents a voiced judgment is given to s ^'.
The number of repetitions until the output s when 0.5) is reduced to 0.55 which is close to the boundary value of the voiced / unvoiced determination is shown in FIG.
As shown in (b), the hangover time is variable. Note that the numbers in FIG. 8B indicate the number of frames of silence when there is a sound (s ≧ 0.
5) is a value indicating whether to output.

Therefore, in the variable overhang generating section 5, when the number of frames (x <0.5) determined to be silent in the output of the inference section 4 continues after the frame determined to be voiced, the final As a result, the reference number of times to determine whether or not there is silence is changed. Therefore, as is clear from FIG. 8B, when the output value s ^′ of the previous frame is close to 1.00 (there is a certainty that there is an infinite sound), s ^′ is close to 0.5. In comparison, the number of frames for determining silence is large. Therefore, if there is a silence-like frame only a few times after being affected by noise after a series of voiced frames,
Although it is originally voiced, the risk of judging it as silence is reduced. It should be noted that when x is 0.5, it approaches the boundary value of the voiced / unvoiced judgment of 0.5 infinitely, but 0.5
It doesn't.

Then, in the binary decision section 6, the output s of the variable overhang generation section 5 is subjected to the final decision as to whether there is sound or not, with 0.5 as a threshold value.

When s ≧ 0.5, voice is present when s <0.5, and voice is not performed for each frame as described above. Even if the above is used, the first and second multi-valued logic judgment units make only judgments according to the certainty of voiced and unvoiced sounds, and the inference unit makes an inference considering these judgment results and the judgment result of the previous frame. This makes it possible to finally make a more reliable decision.

In the present embodiment, the power and the number of zero crossings in each frame out of the divided voices are used as parameters, but in addition to this, the ratio of the power between the previous frame and the current frame, or each frame is used. You may use the change of a spectrum etc. Furthermore, multi-valued logic determination may be performed on three or more parameters in advance.

Further, in this embodiment, the inference unit 4 of the previous frame is used.
In addition to this, the output of the variable hangover generating unit 5 of the previous frame may be fed back as the input of the inference unit 4.

[0031]

As is apparent from the above embodiment, the present invention multi-values the judgment value based on the parameter extracted from the voice with a continuous value within the range of 0 to 1, according to the judgment system. By outputting a value and performing inference based on the maximum value and the minimum value of a plurality of judgment results , even if a judgment rule with low accuracy is used, a judgment with high accuracy can be finally made.

Further, in the judgment from the parameters,
By providing a data table and reading the data table to make a judgment, the judgment result can be multivalued logically by a simple process even when the parameter and the judgment output have a non-linear relationship.

Further, in the case of suspicious silence determination, the hangover time can be lengthened, so that it is possible to reduce the rate of determining silence between words and endings.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a sound / sound determination device according to an embodiment of the present invention.

FIG. 2 is a diagram showing the output of each block in this embodiment.

FIG. 3A is a block diagram showing an internal configuration of a first multi-valued logic judgment unit in the present embodiment, and FIG. 3B is a diagram showing contents of a data table of the first multi-valued logic judgment unit.

FIG. 4A is a block diagram showing an internal configuration of a second multi-valued logic judgment unit in the present embodiment, and FIG. 4B is a diagram showing contents of a data table of the second multi-valued logic judgment unit.

5A is a block diagram showing an internal configuration of a determination result feedback unit in the present embodiment, and FIG. 5B is a diagram showing contents of a data table of the determination result feedback unit.

FIG. 6 is a block diagram showing an internal configuration of an inference unit in this embodiment.

FIG. 7 is a block diagram showing an internal configuration of a variable hangover generation unit in this embodiment.

FIG. 8A is a diagram showing an input / output relationship of a variable overhang generating unit in the present embodiment, and FIG. 8B is a diagram showing a relationship between an input value and a hangover in the variable overhang generating unit.

FIG. 9 is a block diagram showing a conventional sound / sound determination device.

[Explanation of symbols]

1 Parameter extractor 2 First multi-valued logic judgment unit 3 Second multi-valued logic judgment unit 4 Reasoning Department 5 Variable overhang generator 6 Binary judgment section 7 Judgment result feedback section

   ─────────────────────────────────────────────────── ─── Continued front page       (56) Reference JP-A-2-203397 (JP, A)                 JP-A-2-204799 (JP, A)                 JP-A-3-236100 (JP, A)                 Japanese Patent Laid-Open No. 4-42299 (JP, A)                 JP-A-60-209799 (JP, A)                 Japanese Patent Publication Sho 63-13200 (JP, B2)

Claims (7)

(57) [Claims] 1. A multi-valued logic with continuous values in the range of 0 to 1 is used to determine the presence or absence of voice using each of a plurality of parameters extracted from the input voice, and the maximum and minimum values of these plurality of determination results are determined. A voiced / unvoiced determination method which infers voiced / unvoiced by multivalued logic based on values . 2. The maximum value and the minimum value of the multivalued logic judgment result by a plurality of parameters and the inference result in the previous frame.
The voiced / unvoiced determination method according to claim 1, wherein voiced / unvoiced is inferred by multivalued logic based on the above. 3. The voiced / soundless determination method according to claim 1, wherein the multivalued logical determination is performed by using a data table defining a determination value for a parameter extracted from voice. 4. The voiced / unvoiced determination method according to claim 1, wherein the hangover time is varied according to the inference result of the multivalued logic. 5. A parameter extraction unit for extracting a plurality of parameters from audio data divided for each frame, and a multivalued logic with continuous values in the range of 0 to 1 based on each of the extracted plurality of parameters. A plurality of multi-valued logic judgment units for judging presence / absence of sound, an inference unit for inferring presence / absence of sound with a multi-valued logic based on outputs of the plurality of multi-valued logic judgment units, and an output of the inference unit. maximum variable hangover generator for varying the hangover, and a binary determination unit for performing binary determination on the output of the variable hangover generator, the inference unit includes a plurality of multi-valued logic determination value according to Value and minimum value
A sound / silence determination device , which outputs a value obtained by adding . 6. An inference unit is provided with a decision result feedback unit for inputting the output of the inference unit in the previous frame to the inference unit, and inference is performed from the maximum value and the minimum value of the outputs of the plurality of parameter extraction units and the output of the determination result feedback unit. The voiced / unvoiced determination device according to claim 5, which is performed. 7. The sound / sound determination device according to claim 5, wherein the multi-valued logic determination unit includes a data table defining output values for the parameters.