JPH0285897A - Voice detecting system - Google Patents

Abstract

PURPOSE:To detect the existence of a voiced sound with a high detection rate by calculating a specific value of an input signal as a feature parameter, comparing a calculated value with dictionary data and discriminating it. CONSTITUTION:The number of intersections of reference axes of an input signal from a microphone 11, a value related to an amplitude distribution of a waveform and a value related to a power spectrum are calculated as feature parameters by a parameter calculating part 17. On the other hand, a dictionary data storage part 18 has a standard parameter prescribed by dictionary data with regard to a voiced sound and a specific noise. In this state, as for the input signal, a preprocessing for extracting a feature parameter related to a power spectrum is executed by a multi-channel band pass filter 14 and an A/D converter 15. Also, by an A/D converter 16, other feature parameter is brought to pre-processing, and thereafter, a calculated value of the calculating part 17 and the data of the storage part 18 are compared by a deciding part 17, and whether a voiced sound exists in the input signal or not is decided. In such a way, even if amplitude of a noise is large, the existence of a voice can be detected easily and with a high detection rate.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a voice detection method.

[Prior Art] Conventionally, there are many methods for detecting the presence of speech in a noisy environment. It may be used for preprocessing for recognition, but it is difficult to deploy it for general purposes under high noise conditions. There wasn't.

Note that a method for easily detecting the presence of voice in a noisy environment is to detect the number of times an input signal crosses a reference axis within a fixed time interval.

[Problems to be Solved by the Invention] However, the method using the conventional voice detection method described above generally uses the premise that the amplitude of noise is small compared to the amplitude of voice. If the amplitude is comparable to that of , the presence of voice cannot be detected.

Therefore, the present applicant has developed a method for detecting speech that can easily detect the presence of speech in a noisy environment.The number of reference axis crossings of the input signal and the peak value (the amount of non-dimensionalization of the amplitude level of the waveform) A method for detecting voiced sounds using and as a feature parameter,
■We proposed a method for detecting voiced sounds using the number of reference axis crossings of the input signal and the pressure reference amplitude time (the time when the amplitude of the waveform exceeds the threshold using the effective value within a fixed time interval) as characteristic parameters. ing.

Although the above voice detection methods (1) and (2) are more useful than conventional methods, there is a limit to the improvement in the voiced sound detection rate due to the following reasons.

That is, voice has a characteristic in that the power of low frequency components is large and the power of high frequency components is small. However,
Used as a feature parameter in the methods of ■ and ■ above,
The number of reference axis crossings and the values related to the waveform amplitude distribution (wave height value,
The reference amplitude time) approximates the frequency of the main frequency component of the input signal, and does not provide information regarding the frequency distribution of the input signal. Therefore, in methods (1) and (2), it is difficult to improve the voiced sound detection rate because the amount of information regarding the frequency distribution, which is one of the basic characteristics of voiced sounds, is insufficient.

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

[Means for Solving the Problem] The present invention according to claim 1 calculates the number of reference axis crossings of an input signal, a value related to the amplitude distribution of the waveform, and a value related to the power spectrum as characteristic parameters, and Compare the results with dictionary data for voiced sounds and specific noises,
It is designed to determine whether the input signal contains voiced sounds.

Here, the reference axis crossing number is the number of times a human input signal crosses a predetermined reference level, such as a zero level, and when the input signal includes a voiced sound, -M indicates a value within a certain range.

According to a second aspect of the present invention, a wave height value P expressed by the following equation is used as a value related to the amplitude distribution of the waveform. This peak value indicates a value within a certain range -M when the input signal includes a voiced sound.

P = 20X 1og + o (Vp/V, □),
(2) The maximum absolute value of the amplitude within a certain time interval ■r□ = Effective value of the amplitude within the same certain time interval The present invention as set forth in claim 3 provides, for example, the following as a value related to the amplitude distribution of the waveform: The wave height value P expressed by the formula is used. This peak value generally indicates a value within a certain range when the input signal includes a voiced sound.

P = 20X log + o (Vp/Va), ■
, (d) Maximum value of the absolute value of the amplitude within a certain time interval ■: Average value of the absolute value of the amplitude within the same certain time interval This method uses the time (referred to as pressure reference amplitude time) during which the effective value exceeds a threshold value within a fixed time interval.

This pressure reference amplitude time is
−M indicates a value within a certain range.

The present invention as set forth in claim 5 provides that the audio frequency band of the input signal is divided into a plurality of channels by a multi-channel bandpass filter, and the total value of the power obtained from all the channels is determined as the value related to the power spectrum. The ratio of the total power values obtained from the low band channels is used. Here, voiced sounds are characterized in that the power on the low frequency component side is greater than the power on the high frequency component side. Therefore, when the input signal includes a voiced sound, the ratio indicated in "1" indicates a value larger than that of noise for -S. Note that the multi-channel bandpass filter is configured with a plurality of channels in which the audio frequency band is divided into a plurality of bands at equal intervals or logarithmic intervals.

[Operations] In the present invention as set forth in claim 1, speech in a noisy environment is detected as follows. In the present invention, voiced sounds (sounds that involve vibration of the vocal cords, such as vowels, semi-vowels, and nasal sounds, and almost all human sounds include voiced sounds) are considered to be speech. .

(1) For voiced sounds and specific noises, prepare dictionary data whose feature parameters are the number of reference axis crossings within a certain time interval of these signals, values related to the amplitude distribution of the waveform, and values related to the power spectrum of the waveform.

As dictionary data, for example, the following (a) (bl
(c) is used.

(a) A set of feature parameters for voiced sounds obtained from a large number of sounds.

(b) A set of a large number of characteristic parameters determined for a specific noise (for example, the ringing sound of a specific telephone).

(c) A set of feature parameters obtained for a large number of voices by adding voiced sounds and specific noise at a specific ratio.

Note that the data in (a), (b), and (c) above are numerical data obtained by converting acoustic data into feature parameters, statistical data such as an average value and variance obtained by statistically processing numerical data, or based on statistical data. It can be prepared in various forms such as discriminant data such as boundary equations.

(2) An input signal is sampled, and the number of reference axis crossings within a certain time interval of the input signal, a value related to the amplitude distribution of the waveform, and a value related to the power spectrum of the waveform are calculated as characteristic parameters.

(3) The feature parameters calculated in (2) above and the above (
The standard pattern defined by the dictionary data defined in 1) is
Comparisons are made in the parameter space, and pattern recognition is used to determine whether the input signal contains voiced sounds.

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

■Category “voiced sounds” defined by dictionary data (mentioned above)
A parameter space is constructed that is divided into two categories: a) (voiced sounds) or (C) voiced sounds added at a specific ratio) and the category "other", and the characteristic parameters of the human signal are Determine which category belongs to.

■Next, considering the possibility that the specific noise has a large amplitude and greatly affects the detection of voiced sounds, in addition to the above
, and determine which category the feature parameter of the input signal belongs to.

■As a result of the judgments in ■ and ■ above, voiced sounds are present in the input signal, provided that the input signal belongs to the category "voiced sound" in ■, and does not belong to the category "specific noise" in ■. Determine.

Therefore, in the present invention as set forth in claim 1, the basic characteristics of voiced sounds are Since the feature parameter that reflects the bias in the frequency distribution is used as the third parameter, even if the specific noise has a large amplitude and has a large influence on speech detection, the voiced sound category and the specific noise category can be differentiated. can be clearly separated in the parameter space, and the presence of speech in a noisy environment can be easily detected with a high detection rate.

According to the second aspect of the present invention, since the wave height value as described above is used as the value related to the amplitude distribution of the waveform, parameter values that faithfully reflect the sharp waveform that is a characteristic of voiced sounds are used. , 21g has the advantage of improving its identifiability.

According to the present invention as set forth in claim 3, since the above-mentioned wave height value is used as the value related to the amplitude distribution of the waveform, the amount of calculation can be reduced compared to the present invention as set forth in claim 2, and is advantageous. Parameter values that relatively faithfully reflect the sharp waveform that characterizes vocal sounds are used, which has the advantage of improving noise discrimination. According to the present invention as claimed in claim 4, in which a small amount of calculation means a fast response speed, the above-mentioned ultra-standard amplitude time is used as the value related to the amplitude distribution of the waveform. There is an advantage that the amount of calculation can be further reduced compared to the present invention described in item 2 or 3.

According to the present invention as set forth in claim 5, since the power ratio as described above is used as the value related to the power spectrum of the waveform, it is possible to use a parameter value that reflects the slope of the frequency distribution, which is a characteristic of voiced sounds. This has the advantage of improving distinguishability from noise.

[Embodiment] 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.

In FIG. 1, 11 is a microphone, 12 is an amplifier, 13 is a low-pass filter, 14 is a multi-channel band-pass filter, 15 is an A/D converter, 16 is an A/D converter, 17 is a parameter calculation unit, and 18 is dictionary data. A storage section, 19 a determination section, and 20 a result output section. In this embodiment, speech in a noisy environment is detected as follows.

(1) Regarding voiced sounds and specific noises, 20% of their signals
*The number of reference axis crossings x between S, the value X2 regarding the amplitude distribution of the waveform, and the value X regarding the power spectrum of the waveform
3 is prepared as a feature parameter, and is stored in the dictionary data storage section 18.

Here, as the value X2 regarding the amplitude distribution of the waveform, any one of the following ■, ■, and ■ can be used.

■The wave height value P expressed by the formula below.

P = 2 (IX log + o (vp/vr, *) plus v2 d Maximum absolute value of amplitude within a certain time interval V□1: Effective value of amplitude within the same certain time interval ■ Expressed by the following formula wave height P.

P = 20X log+o (Vp/Va) However, ■
, Maximum absolute value of amplitude within a time interval ■ Average value of absolute value of amplitude within a fixed time interval ■ Time when the amplitude exceeds a threshold value using the effective value as a guideline within a fixed time interval (super reference amplitude time).

When using the wave height value of (2) above, a parameter value that relatively faithfully reflects the sharp waveform characteristic of voiced sounds is used, which has the advantage of improving noise discrimination.

When using the wave height value of ■ above, the amount of calculation is reduced compared to the wave height value of ■ above, and parameter values that faithfully reflect the sharp waveform that is characteristic of voiced sounds are used. This has the advantage of improving identifiability.

When using the wave height value of (2) above, there is an advantage that the amount of calculation can be reduced compared to the wave height values of (2) and (3) above.

In addition, as for the value X regarding the power spectrum, the audio frequency band of the input signal is divided into multiple channels using a multi-channel bandpass filter, and the low band A ratio of the sum of the powers obtained from the channels can be used. When using this power ratio,
Parameter values that reflect the slope of the frequency distribution, which is a characteristic of voiced sounds, are used, which has the advantage of improving distinguishability from noise.

In addition, as dictionary data, for example, the following (a) and (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 determined for a specific noise (ringing sound of a specific telephone).

[c) Voiced sound [a] and specific noise, 20 x 10g+o (Sr+s++/Nrm*)
Voiced to specific noise ratio defined in [dB]3. O,-
A set of feature parameters obtained for a large number of voices by adding the results at 3°-6 and -10 [dB]. In addition, S r
ams represents the effective value of the amplitude of the voiced sound "a", and N1□
represents the effective value of the amplitude of the specific noise.

(2) An input signal is collected by the microphone 11, amplified by the amplifier 12, and passed through the low-pass filter 13 to extract only audio band components of 4.2 KHz or less. After this, the input signal is transferred to: ■ a path consisting of a multi-channel band-pass filter 14 and an A/D converter 15 that perform preprocessing for extracting feature parameters related to the power spectrum; and ■ values related to amplitude distribution and the number of reference axis crossings. The signal is branched and transferred to a path including an A/D converter 16 that performs preprocessing to extract two characteristic parameters. The multi-channel band pass filter 14 is
This is a bandpass filter that divides the frequency band from 250Hz to 4KHz into 25 channels every 176 octaves. The A/D converters 15 and 16 have a sampling frequency of 10 Kjlz and a conversion bit number of 16 bits. The digital values of the input signals and their frequency signals obtained by the preprocessing sections of these filters 14 and converters 15 and 16 are sent to the parameter calculation section 17. The parameter calculation unit 17 calculates a reference axis frequency xI during 2hS of the input signal, a value X2 regarding the amplitude distribution of the waveform,
A value X3 related to the power spectrum is calculated as a characteristic parameter.

(3) The feature parameters calculated in (2) above and the above (
The determination unit 19 compares the standard parameters defined by the dictionary data determined in 1) to determine whether the input signal includes a voiced sound, and outputs the determination result from the result output unit 20.

Here, pattern recognition using the aforementioned dictionary data is performed, for example, on the parameter space shown in FIG. 2 as follows.

In addition, in Figure 2, the three characteristic parameters of the number of zero crossings (the reference axis level is set to zero level), the value related to the amplitude distribution of the waveform, and the value related to the power spectrum are
This is taken for the 1st axis, the X2 axis, and the X3 axis. In FIG. 2, μmσ11% σ12, σ1. are the average values of the dictionary parameters of voiced sounds (voiced sounds [a] in (a) above, or voiced sounds obtained by adding the specific noises in (c) above at a specific voiced sound to specific noise ratio), respectively, and the X1 axis component , the standard deviation of the X2 axis component, and the standard deviation of the x33 component, and μ2, σ21, σ22, and σ23 each represent similar values for the dictionary parameters of the specific noise.

■Category “voiced sounds” defined by dictionary data (mentioned above)
A boundary 1 is defined that divides the voiced sound in a) [ako, or the voiced sound category in which the specific noise in (c) above is added at a specific ratio] and the category "Other" into two. In boundary 1, the side that includes the average value μm of voiced sound dictionary data is the category "voiced sound". This boundary 1 shows how much voiced sound dictionary data is concentrated around the average value. By changing the length of the axis, you can change the proportion of dictionary data of voiced sounds that fall within the ellipse. In this embodiment, the length of the axis is determined so that 90% of the voiced sound dictionary data falls within the ellipse. The dashed line is μ
represents the concept of the category "voiced sound" defined by .

0th Next, considering the possibility that the amplitude of specific noise is large and this has a large influence on the detection of voiced sounds, in addition to the above ■, the category "specific noise" and the category "voiced sound" are added.
Define boundary 2. In boundary 2, the side that includes the average value μ2 of the specific noise is in the category "specific noise." This boundary 2 is a collection of points that have the same likelihood for the category "voiced sound" and the category "specific noise." In this example, the standard deviation of the specific noise is the dictionary data of the artificially created incoming ring tone of a telephone, which is a combination of voiced sounds and specific noises at a specific voiced to specific noise ratio. Since it 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 the process (2), it is determined which category of the boundary 2 the characteristic parameter of the input signal belongs to.

■ As a result of the above determinations ■ and ■, the input signal belongs to the μm side of boundary 1 in ■ on the feature parameter space, and ■
When the input signal does not belong to the μ2 side of boundary 2, the input signal is determined to belong to the category "voiced sound". That is, it is determined that a voiced sound exists in the input signal.

Therefore, in the above embodiment, compared to the case where only two characteristic parameters, the number of reference axis crossings and the value related to the amplitude distribution of the waveform, are used, the frequency, which is one of the basic characteristics of voiced sounds, is The characteristic parameters that reflect the bias of the distribution are
Since the amplitude of the specific noise is large and its influence on speech detection is large, the voiced sound category and the specific noise category can be clearly separated in the parameter space, and it is possible to clearly separate the voiced sound category and the specific noise category in the parameter space. Its presence can be easily detected with a high detection rate.

In particular, in the above embodiment, the voiced to specific noise ratio is −6[d
B] also showed a high detection rate of voiced sounds, -3 (dB
] was found to have a detection rate close to 100[%].

Note that in the above embodiment, a concentration ellipse and a collection of points having equal likelihoods for two categories were used as the boundary line defining the standard pattern on the feature parameter space, or other methods may be used in implementing the present invention. General pattern recognition techniques can be used. For example, instead of a collection of points that have equal likelihoods for the categories ``voiced sounds'' and ``specific noises,'' Mahara
A collection of points having the same nobis distance or Euclid distance can be used.

[Effects of the Invention] As described above, according to the present invention, the presence of speech 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 speech detection is large. Can be done.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an example of a voice detection device used to implement the present invention, and FIG. 2 is a schematic diagram showing a parameter space formed by characteristic parameters of the present invention. 11...Microphone, 14...Multi-channel band pass filter, 17...
・Parameter calculation section, 18... Dictionary data storage section, 19... Judgment section, 20... Result output section. Patent applicant: Sekisui Chemical Co., Ltd. Representative Hiroshi1) Kaoru

Claims (5)

[Claims] (1) Calculate the number of reference axis crossings of the input signal, values related to the amplitude distribution of the waveform, and values related to the power spectrum as characteristic parameters, and compare the calculation results with dictionary data regarding voiced sounds and specific noise, A voice detection method that determines whether the input signal contains voiced sounds. (2) Audio detection according to claim 1, wherein the value regarding the amplitude distribution of the waveform is a peak value expressed as a ratio of the effective value of the amplitude within the certain time interval to the maximum value of the absolute value of the amplitude within the certain time interval. method. (3) The wave height value expressed as the ratio of the average value of the absolute value of the amplitude within the certain time interval to the maximum value of the absolute value of the amplitude within the certain time interval is used as the value related to the amplitude distribution of the waveform. voice detection method. (4) The voice detection method according to claim 1, wherein the value regarding the amplitude distribution of the waveform is a time period in which the amplitude exceeds a threshold value based on an effective value within a certain time interval. (5) As a value related to the power spectrum, the audio frequency band of the input signal is divided into multiple channels using a multi-channel bandpass filter, and the total value of the power obtained from all channels is calculated by dividing the audio frequency band of the input signal into multiple channels using a multichannel bandpass filter. 2. The voice detection method according to claim 1, which uses a ratio of the sum of the powers.