JP2559475B2 - Voice detection method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a voice detection system.

[Prior Art] Conventionally, there are a number of methods for detecting the presence of speech in a noisy environment, such as detecting speech sections in communication as described in Japanese Patent Publication No. 57-12999, or using speech language contents. Although it is used for pre-processing of recognition, it is difficult to develop it for general use under high noise.For example, starting a response by voice of a hands-free phone while the ringing tone is ringing, etc. could not.

As a method of easily detecting the presence of voice in a noisy environment, there is a method of detecting the number of times the input signal crosses the reference axis within a fixed time interval.

[Problems to be Solved by the Invention] However, in the above-mentioned method using the conventional voice detection method, it is generally assumed that the noise amplitude is smaller than the voice amplitude, and the noise amplitude is lower than the voice amplitude. If the amplitude is almost the same, the presence of voice cannot be detected.

Therefore, the present applicant has set the number of reference axis crossings and the peak value (the dimensionless amount of the amplitude level of the waveform) of the input signal as a voice detection method that can easily detect the presence of the voice in a noisy environment. As a characteristic parameter, the method of detecting voiced sound, the number of reference axis crossings of the input signal, and the super-standard amplitude time (the time when the amplitude of the waveform exceeds the threshold value of the effective value within a certain time interval) as the characteristic parameter We propose a method to detect voiced sounds.

Although the above voice detection method is more useful than the conventional method, there is a limit in improving the voiced sound detection rate for the following reasons.

That is, in the case of (1), if there is an input signal with a particularly high amplitude level even at one location due to the influence of noise or the like within a certain period of time, the level becomes the peak value. The presence of such a section tends to increase the dispersion of the crest values, and becomes a factor that hinders the improvement of the voiced sound detection rate.

Further, in the case of, the super-reference amplitude time is only a value in the time axis direction of the waveform, and this parameter alone lacks the amount of information regarding the amplitude level of the waveform, making it difficult to improve the voiced sound detection rate. There is.

An object of the present invention is to easily detect the presence of voice in a noisy environment with a high detection rate even when the amplitude of noise is large and the influence on voice detection is large.

[Means for Solving the Problem] The present invention according to claim 1 is the number of reference axis crossings of an input signal and the effective value of the amplitude within the fixed time interval with respect to the maximum absolute value of the amplitude within the fixed time interval. The peak value represented by the ratio of the values and the time when the amplitude of the waveform exceeds the threshold value, which is the effective value within the fixed time interval, are calculated as characteristic parameters, and the calculation result is used as the voiced sound and the specific noise. It is designed to determine whether or not the input signal contains voiced sound by comparing with the dictionary data of the above, and the effective value of the amplitude within the fixed time interval with respect to the maximum absolute value of the amplitude within the fixed time interval. As the crest value represented by the ratio, the crest value P represented by the following equation, for example, is used.

P = 20 × log ₁₀ (V _P / V _rms ) where V _P : maximum absolute value of amplitude within a fixed time interval V _rms : effective value of amplitude within the fixed time interval According to the invention, a peak value P represented by the following equation, for example, is given as a peak value represented by the ratio of the average value of the absolute values of the amplitude within the fixed time interval to the maximum value of the absolute value of the amplitude within the fixed time interval. It is the one that is used.

P = 20 × log ₁₀ (V _P / V _a ), where V _P : maximum value of the absolute value of the amplitude within a fixed time interval V _a : the average value of the absolute value of the amplitude within the fixed time interval [Action] Claim According to the present invention described in Item 1, the voice in a noisy environment is detected as follows. In the present invention, voiced sounds (voices with vocal folds such as vowels, semi-vowels, and nasal sounds, and almost all voices uttered by humans include voiced sounds) are regarded as voices. .

(1) For voiced sound and specific noise, the number of reference axis crossings (the number of times the reference level crosses a predetermined reference level such as zero level) of those signals within a certain time interval, and the maximum absolute value of the amplitude within the certain time interval. A crest value represented by the ratio of the effective value of the amplitude within the certain time interval to the value, and the time when the amplitude of the waveform exceeds the threshold value of the effective value within the certain time interval (super-reference amplitude time) Prepare dictionary data that uses as a feature parameter.

Examples of the dictionary data include the following (a), (b),
(C) is used.

(A) A set of feature parameters for voiced sounds obtained from multiple voices.

(B) A set of a large number of characteristic parameters obtained for a specific noise (for example, a ring tone of a specific telephone).

(C) A set of feature parameters obtained by adding a result of adding a voiced sound and a specific noise at a specific ratio for many voices.

The data (a), (b), and (c) above are based on numerical data obtained by converting acoustic data into characteristic parameters, average values obtained by statistically processing the numerical data, statistical data such as variance, or statistical data. Can be prepared in various forms such as discriminant data such as a boundary equation determined by

(2) The input signal is sampled, and the ratio of the number of reference axis crossings within a certain time interval of this input signal and the effective value of the amplitude within the certain time interval to the maximum absolute value of the amplitude within the certain time interval The peak value represented and the super-reference amplitude time are calculated as characteristic parameters.

(3) The feature parameters calculated in (2) are compared with a standard pattern defined by the dictionary data defined in (1) in a parameter space, and pattern recognition is performed to determine whether or not the input signal includes a voiced sound. Determined by

The above pattern recognition using the dictionary data is performed as follows, for example.

The category "voiced sound" defined by the dictionary data (the voiced sound of (a) or the voiced sound category of the specific noise of (c) added at a specific ratio) and the category "others" are divided into two minutes. And determines which category the feature parameter of the input signal belongs to.

Next, in consideration of the possibility that the amplitude of the specific noise is large, which greatly affects the detection of voiced sound, in addition to the above, the category "specific noise" and the category "voiced sound"
Is defined, and it is determined whether the characteristic parameter of the input signal belongs to which category.

As a result of the above determination, it is determined that a voiced sound exists in the input signal on condition that the input signal belongs to the category “voiced sound” and does not belong to the category “specific noise”.

Therefore, in the present invention according to claim 1, since the three parameters of the reference axis crossing number, the peak value, and the super-standard amplitude time are used as the characteristic parameters, the reference axis crossing number and the amplitude within the fixed time interval are used. When using only the crest value represented by the ratio of the effective value of the amplitude within the fixed time interval to the maximum absolute value of and the two parameters, the absolute value of the absolute value of the amplitude within the fixed time interval is affected by noise. The tendency that the variance of the crest value represented by the ratio of the effective value of the amplitude within the fixed time interval to the maximum value becomes large and the voiced sound detection rate does not improve is complemented by using the super-reference amplitude time together, and When only two parameters of the number of reference axis crossings and the super-standard amplitude time are used, the fact that the amount of information about the amplitude level of the waveform is insufficient is considered within the fixed time interval with respect to the maximum absolute value of the amplitude within the constant time interval. Shaking Interplay The combined use of the wave height value represented by the ratio of the effective value. As a result, even when the amplitude of specific noise is large and the influence on the detection of speech is large, the category “voiced sound” and the category “specific noise” can be clearly separated in the parameter space, and the presence of speech in a noisy environment can be detected. It can be easily detected with a high detection rate.

According to the present invention as set forth in claim 1, since the crest value as described above is used, a parameter value that faithfully reflects a sharp waveform that is a characteristic of voiced sound is used, and noise discrimination is improved. There are merits. Further, in this case, since the effective value of the amplitude can be shared in the process of calculating the peak value and the super-reference amplitude time, the amount of calculation is small,
The detection work becomes easier.

According to the present invention as set forth in claim 2, since the crest value represented by the ratio of the average value of the absolute values of the amplitude within the fixed time interval to the maximum value of the absolute value of the amplitude within the fixed time interval is used,
The calculation amount can be reduced as compared with the present invention according to claim 1,
In addition, a parameter value that relatively faithfully reflects a sharp waveform that is a characteristic of voiced sound is used, and there is an advantage that noise discrimination is improved. Note that a small amount of calculation means a high response speed.

[Embodiment] FIG. 1 is a block diagram showing an example of a speech detection device used in the embodiment of the present invention, and FIG. 2 is a schematic diagram showing a parameter space formed by characteristic parameters of the present invention.

In FIG. 1, 11 is a microphone, 12 is an amplifier, 13 is a low-pass filter, 14 is an A / D converter, 15 is a parameter calculation unit, 16 is a dictionary data storage unit, 17 is a judgment unit, and 18 is a result output unit. . In this embodiment, speech in a noisy environment is detected as follows.

(1) For voiced sound and specific noise, 20 ms of those signals
Number of reference axis crossings between X ₁ , peak value X ₂ , and super-standard amplitude
Dictionary data having X ₃ as a characteristic parameter is prepared and stored in the dictionary data storage unit 16.

Here, as the peak value X ₂ , any of the following can be used.

The peak value P expressed by the following equation.

P = 20 × log ₁₀ (V _P / V _rms ) where V _P : maximum absolute value of amplitude within a fixed time interval V _rms : effective value of amplitude within a fixed time interval Crest value expressed by the following formula P.

P = 20 × log ₁₀ (V _P / V _a ), where V _P : maximum absolute value of amplitude within a fixed time interval V _a : average absolute value of amplitude within the fixed time interval In the case of using, the parameter value that reflects the sharp waveform that is the characteristic of voiced sound relatively faithfully is used, and there is an advantage that the discrimination of noise is improved.

When the above crest value is used, the amount of calculation can be reduced compared to the above crest value, and the parameter value that faithfully reflects the sharp waveform that is the characteristic of sexual sound is used. There is a merit that the property is improved.

As the dictionary data, for example, the following (a):
(B) and (c) are created.

(A) A set of feature parameters for voiced sound [a] obtained from a large number of voices.

(B) A set of a large number of characteristic parameters obtained for a specific noise (ringing sound of a specific telephone).

(C) Many results obtained by adding voiced sound [a] and specific noise at a voiced sound-to-specific noise ratio of -10 [dB] defined by 20 × log ₁₀ (S _rms / N _rms ) [dB] A set of characteristic parameters obtained for the voice of. Note that S _rms represents the effective value of the amplification of the voiced sound “A”, and N _rms represents the effective value of the amplitude of the specific noise.

(2) The input signal is sampled by the microphone 1i, this input signal is amplified by the amplifier 12 and passed through the low-pass filter 13 to cut the component of 4.2KHz or more, and the A / D converter 1
The signal is converted into a digital signal having a sampling frequency of 10 KHz and a conversion bit number of 16 bits by 4 and is sent to the parameter calculation unit 15.
The parameter calculator 15 calculates the reference axis crossing number X ₁ during 20 mS of the input signal, the peak value X _2, and the super-reference amplitude time X ₂ as characteristic parameters.

(3) The feature parameter calculated in (2) is compared with the standard pattern defined by the dictionary data defined in (1) in the determination unit 17 to determine whether the input signal includes a voiced sound. The result is output from the result output unit 18.

Here, the pattern recognition using the aforementioned dictionary data is
For example, the following is done on the parameter space in FIG.

In Fig. 2, three characteristic parameters of the number of zero crossings (reference axis level set to zero level), peak value, and super-standard amplitude time are taken on the X ₁ axis, X ₂ axis, and X ₃ axis, respectively. It is a thing. In FIG. 2, μ ₁ , σ ₁₁ , σ ₁₂ , and σ ₁₃ are voiced sounds (voiced sound [a] in (a) above) or specific noise in (c) above in a specific voiced to specific noise ratio. Sum of voiced speech dictionary parameters, standard deviation of X ₁ axis component, standard deviation of X ₂ axis component, standard deviation of X ₃ axis component, μ ₂ , σ ₂₁ , σ ₂₂ , σ ₂₃ Each represent a similar value for a particular noise dictionary parameter.

A category “voiced sound” defined by the dictionary data (voiced sound [a] in (a) above or a voiced sound category obtained by adding specific noise in (c) above at a specific ratio);
A boundary 1 that divides the category “other” into two is determined. In the boundary 1, the side including the average value mu ₁ of the dictionary data voiced is category "voiced". This boundary 1 is a concentrated ellipse indicating how much the voiced dictionary data is concentrated around the average value, and the ratio of the voiced dictionary data to the ellipse is changed by changing the length of the axis. be able to. In the case of this embodiment, the length of the axis is determined so that 90% of the voiced dictionary data falls within the ellipse. The broken line represents the concept of the category “voiced sound” defined by μ and σ. That is, in this process, it is determined which side of the boundary 1 the characteristic parameter of the input signal belongs to.

Next, in consideration of the possibility that the amplitude of the specific noise is large, which greatly affects the detection of voiced sound, in addition to the above, the category "specific noise" and the category "voiced sound"
The boundary 2 of is determined. At the boundary 2, the side including the average value μ ₂ of the specific noise is the category “specific noise”. The boundary 2 is a group of points having the same likelihood for the category “voiced sound” and the category “specific noise”. In the case of this embodiment, the standard deviation of the specific noise is the ring tone of the artificially created telephone, and the dictionary data of the voiced sound and the specific noise added at the specific voiced sound to specific noise ratio. Is generally smaller than the standard deviation of, the category "specific noise" is a closed space. The broken line represents the concept of the category “specific noise” defined by μ and σ. That is, in this process, it is evaluated which side of the boundary 2 the characteristic parameter of the input signal belongs to.

Above, the results of the determination of the input signal, in feature parameter space, belonging to the mu ₁ side of the boundary 1 in, when not belonging to mu ₂ side of the boundary 2 in and belongs to the category "voiced" input signal Is determined. That is, it is determined that a voiced sound exists in the input signal.

In the above embodiment, the three parameters of the reference axis crossing number, the crest value, and the super-standard amplitude time are used as the characteristic parameters. Therefore, when only the two parameters of the reference axis crossing point and the crest value are used, By using the super-reference amplitude time together, the tendency that the detection rate of voiced sound does not improve due to the large variance of the crest value due to the influence of noise is complemented.
Further, when only two parameters of the reference axis crossing number and the super-standard amplitude time are used, the lack of the information amount regarding the amplitude level of the waveform can be complemented by using the peak value together. As a result, even when the amplitude of specific noise is large and the influence on the detection of speech is large, the category “voiced sound” and the category “specific noise” can be clearly separated in the parameter space, and the presence of speech in a noisy environment can be detected. , Can be easily detected with a high detection rate.

In particular, as a result of an experiment in a very noisy environment with a voiced sound-to-specific noise ratio of -10 [dB], when a voiced sound is detected using the number of zero crossings and peak values as characteristic parameters, the detection rate is 50 [ %], When the voiced sound is detected using the number of zero crossings and the super-reference amplitude time as the characteristic parameters, it was 60%, but in the case of the above-mentioned embodiment, it was 90%, and the effect of the present invention was recognized. It was Further, in the process of calculating the crest value and the super-reference amplitude time, since there are many calculation parts that can be shared (for example, the effective value of the amplitude), the processing time of the above embodiment was almost the same as or.

In the above embodiment, a concentrated ellipse and a set of points at which the likelihoods of the two categories are equal are used as the boundary defining the standard pattern in the feature parameter space. General pattern recognition technique can be used. For example,
Instead of a set of points with the same likelihood for the categories "voiced" and "specific noise", Maharanobi
A group of points where the s distance and the Euclid distance are equal can be used.

[Effects of the Invention] As described above, according to the present invention, the presence of voice in a noisy environment can be easily detected with a high detection rate even when the amplitude of noise is large and the influence on voice detection is large. You can

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of a voice detection device used for implementing the present invention, and FIG. 2 is a schematic diagram showing a parameter space formed by characteristic parameters of the present invention. 11: microphone, 15: parameter calculation unit, 16: dictionary data storage unit, 17: determination unit, 18: result output unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location G10L 9/18 301 G10L 9/18 301A

Claims (2)

(57) [Claims] 1. A crest value represented by the number of reference axis crossings of an input signal, a ratio of the effective value of the amplitude within the fixed time interval to the maximum absolute value of the amplitude within the fixed time interval, and the amplitude of the waveform. Is calculated as a characteristic parameter, and the time over which the threshold value exceeds the effective value as a guide within a certain time interval.
A voice detection method that determines whether an input signal contains voiced sound by comparing it with dictionary data for voiced sound and specific noise. 2. A crest value represented by the number of reference axis crossings of an input signal and a ratio of a mean value of absolute values of amplitudes within a certain time interval to a maximum value of absolute values of amplitudes within a certain time interval, The time over which the amplitude of the waveform exceeds the threshold value, which is the effective value within a certain time interval, is calculated as a characteristic parameter, and the calculation result is compared with the dictionary data for voiced sound and specific noise. A voice detection method that determines whether a voiced sound is included.