JP3107905B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus, and more particularly to a speech recognition apparatus having an improved speech recognition rate in an environment having a large background noise.

[0002]

2. Description of the Related Art Conventionally, as a speech recognition apparatus, a power spectrum in each frequency band of an input speech has been calculated, and a power spectrum in each calculated frequency band and a frequency spectrum of a template input at the time of speech registration have been calculated. There is known a method of calculating a distance or a similarity to a power spectrum to recognize speech.

FIG. 5 shows a conventional speech recognition apparatus for recognizing speech using the above-mentioned method. In FIG. 5, reference numeral 1 denotes a sound input unit, 2 denotes a band spectrum calculation unit, 3 denotes a template storage unit, and 4 denotes a frame. Unit distance calculation means, 5 is generated content collation means, and 6 is recognition result output means. In FIG. 1, an audio signal is input from a sound input unit 1, and a power spectrum (hereinafter, referred to as a band spectrum) in each frequency band of the input voice is calculated by a band spectrum calculating unit 2.

[0004] A template storage unit 3 stores a band spectrum (hereinafter, referred to as a template) for each voice input at the time of voice registration. The distance between the spectrum of each input voice in each band and the template stored in the template storage unit 3 is determined for each frame of the input voice.

[0005] The occurrence content matching means 5 estimates the occurrence content from the distance between the input speech in frame units obtained by the frame unit distance calculation means 4 and each template, and outputs the estimation result from the recognition result output means 6.

[0006]

However, the above-described conventional speech recognition apparatus has a drawback that the recognition rate of input speech is greatly reduced in an environment where background noise is large. SUMMARY OF THE INVENTION The present invention has been made in order to improve the above-described drawbacks of the related art, and has as its object to provide a speech recognition device in which the recognition rate of input speech does not decrease even in an environment with a large background noise. .

[0007]

FIG. 1 is a diagram illustrating the principle of the present invention. The present invention is configured as shown in FIG. 1 in order to solve the above-mentioned problem. The invention according to claim 1 comprises a sound input unit 11 for inputting an audio signal mixed with noise, and a sound input unit 11.
A band spectrum calculating means 12 for calculating a band spectrum from the audio signal input from the LM, a distance calculating means 14 for calculating a distance between the template 13 and the band spectrum of the input audio signal, and a distance calculating means 14 Utterance content matching means 15 for comparing the utterance content based on the calculated distance ;
In a speech recognition device including a recognition result output unit 16 that outputs a result matched by the utterance content matching unit as a recognition result, a noise band spectrum estimating unit 17 that estimates a band spectrum of noise mixed at the time of speech input; And a parameter determining means for determining a parameter from an estimated value of the band spectrum of the input voice and the band spectrum of the noise.

[0008] Then, using the parameters determined by the parameter determining means 18, the distance calculating means 14, each
In each band, the band spectrum of the template 13
The band width of the input speech whose strength is determined as a parameter
Based on the spectrum and noise band spectrum estimates.
Template depending on whether it is inside or outside the section
For calculating distance between input signal 13 and band spectrum of input audio signal
Change the, calculated on the Symbol distance, which is constituted to recognize the noisy speech signal.

According to a second aspect of the present invention, there is provided a voice signal containing noise.
Input means 11 for inputting a signal and input from the sound input means 11.
Band spectrum to calculate band spectrum from input audio signal
Vector calculation means 12, template 13 and input voice
Similarity calculation to calculate the similarity between the signal and the band spectrum
Means 14 and the similarity calculated by the similarity calculating means 14
Speech content collation means 15 for collating speech content based on degree
And the recognition result by the utterance content matching means
And a recognition result output means 16 for outputting
In identification devices, the band scan of noise to be mixed at the speech input
Noise band spectrum estimating means 17 for estimating the spectrum
And the input speech band spectrum and the noise band spectrum
Deciding means for deciding parameters from estimated values
18 are provided. And the parameter determinant
Using the parameters determined by step 18, a similarity measure
The calculation means 14 calculates the template 13 in each band.
Between the template 13 and the band spectrum of the input voice signal depending on whether the band spectrum intensity of the Change the similarity calculation method to calculate the similarity,
It is designed to recognize mixed audio signals.
You.

According to a third aspect of the present invention, in the first aspect of the present invention, in each band, the band spectrum intensity of the template 13 is determined based on an estimated value of the band spectrum of the input voice and the band spectrum of the noise determined as parameters. A distance calculator 14 is provided for setting the distance between the template 13 and the band spectrum of the input audio signal to a constant value when the distance is within the determined section.

According to a fourth aspect of the present invention, in the second aspect of the present invention, in each of the bands, the band spectrum intensity of the template 13 is based on an estimated value of the band spectrum of the input voice and the band spectrum of the noise determined as parameters. A similarity calculating means 14 is provided for setting the similarity between the template 13 and the band spectrum of the input audio signal to a constant value when the similarity is within the determined section.

According to a fifth aspect of the present invention, in the first aspect of the present invention, in each of the bands, the band spectrum intensity of the template 13 is based on the estimated values of the band spectrum of the input voice and the band spectrum of the noise determined as parameters. A distance calculating means 14 is provided for setting the distance between the template 13 and the band spectrum of the input audio signal to 0 when the distance is within the determined section.

The invention of claim 6 is the invention of claims 1, 2, 3, 4
Or, in the invention of claim 5 , in each band,
A parameter determining means 18 is provided for determining a section as a parameter based on the sum and difference between the band spectrum of the input speech and the estimated value of the band spectrum of the noise.

[0014]

According to the first and second aspects of the present invention, the sound input means 11 is provided.
The voice mixed with more noise is input, and the band spectrum calculating means 12 calculates the band spectrum intensity for each band. Further, the noise band spectrum intensity is estimated by the noise band spectrum estimating means 17.

The parameter determining means 18 determines parameters based on the band spectrum intensity of speech and the band spectrum intensity of noise, and calculates the distance or similarity calculating means 1 based on the parameters determined by the parameter determining means 18.
4 is the distance between the voice band spectrum and the template 13.
Alternatively, the similarity is calculated. The utterance content matching unit 15 collates the input speech with the band spectrum of the input speech based on the distance or similarity of the template, and outputs the label of the template having the shortest distance from the recognition result output unit 16.

As the parameters determined by the parameter determining means 18, a section is determined based on the band spectrum intensity of speech and the band spectrum intensity of noise as described in claim 1 or 2 , depending on whether or not the template is within the section. Change the formula to calculate distance or similarity,
The distance or the similarity when the template is present in the section is constant or zero as in claims 3, 4 or 5.
It can be determined as follows. Further, as in claim 6, the section can be determined by the sum and difference between the band spectrum intensity of the voice and the band spectrum intensity of the noise.

[0017] Determine the basis of the band spectral intensity and band spectrum intensity of the noise of the speech parameters, therefore calculates the distance or similarity in distance or similarity calculation unit 14 based on the parameter, taking into account the influence of the noise of each band And the degree of similarity can be calculated, and even under an environment with large background noise, speech recognition with a sufficient recognition rate can be performed.

[0018]

In the figure, 111 is a microphone, 112 is an analog / digital converter, 121 is a fast Fourier transform operation unit (hereinafter referred to as FFT operation unit), 122 is a band spectrum intensity calculation unit, 131 is a template storage unit, 14
1 is a DP local distance calculation unit, 142 is a calculation formula determination unit, 1
51 is a DP cumulative distance calculation unit, 152 is a minimum distance search unit,
161 is a label output unit, 171 is an environmental noise intensity storage unit,
181 is a section determination unit.

In FIG. 1, a microphone 111 is a means for inputting an acoustic signal, and the output of the microphone 111 is an analog / digital signal.
The digital signal is supplied to a digital converter 112 and converted into a digital signal. The FFT operation unit 121 is an operation unit that performs a Fourier transform on the audio signal output from the analog / digital converter 112 for each frame, and the output is provided to the band spectrum intensity calculation unit 122.

The band spectrum intensity calculation unit 122 is means for calculating the power spectrum intensity of each of the frequency bands of the input acoustic signal. The output of the band spectrum intensity calculation unit 122 includes a section determination unit 181 and a local distance for DP. The calculation unit 141 and the environmental noise intensity storage unit 171 are provided. The DP local distance calculation unit 141 is means for calculating a local distance for each frame for calculating the cumulative distance by dynamic programming. The calculation formula used here is given by the calculation formula determination unit 142. The environmental noise intensity storage unit 171 is a unit that stores the band spectral intensity before the voice is input as the environmental noise intensity.

The section determining section 181 is a means for determining the band spectrum section of the input audio signal in consideration of the environmental noise intensity by adding and subtracting the band spectrum of the input audio signal and the environmental noise intensity. Part 142
Given to. The calculation formula determining unit 142 includes a section determining unit 181
Is a means for determining the calculation formula used in the DP local distance calculation unit 141 based on the output of.

Based on the local distance calculated by the DP local distance calculator 141, the DP cumulative distance calculator 151 obtains an optimal path using dynamic programming, thereby obtaining the distance between the input audio signal and each template. The distance obtained here is given to the minimum distance search unit 152. The minimum distance search unit 152 is DP
Of the input audio signal obtained by the local distance calculation unit 141 and the distance between the templates, for searching for the template of the minimum distance. The label output unit 161 outputs the label of the template searched for by the minimum distance search unit 152. Output.

FIG. 3 is a diagram showing the configuration of the section determination unit 181 of FIG. 2. In FIG. 3, reference numeral 182 denotes an addition unit, and 183 denotes a subtraction unit. As shown in FIG. 3, the section determination unit 181 stores the band spectrum intensity Pi (i indicates the number of the band) of the input audio signal output from the band spectrum intensity calculation unit 122 and the environmental noise intensity storage unit 171. The added environmental noise intensity Ni is added by the adding unit 182, and is subtracted by the subtracting unit 183 to obtain the section [Pi−Ni, Pi +
Ni].

FIG. 4 is a diagram for explaining the operation of this embodiment. In FIG. 4, FIG. 4A shows an example of the band spectrum intensity in each band in a certain frame, and FIG. ) Is the band spectrum intensity Pi of the input audio signal when the horizontal axis represents time t and the vertical axis represents the band spectrum intensity P, and the value Ai = Pi−Ni obtained by subtracting the environmental noise intensity Ni from the band spectrum intensity Pi. The value Bi = Pi obtained by adding the spectrum intensity Pi and the environmental noise intensity Ni.
FIG. 9 is a diagram showing an example of the relationship of + Ni.

FIG. 3C is a diagram corresponding to FIG.
(In the figure, each band is indicated as i = 1 to i = n). Next, the operation of this embodiment will be described with reference to FIGS. The acoustic signal input by the microphone 111 is converted into a digital signal by the analog / digital converter 112 and provided to the FFT operation unit 121.

The FFT operation unit 121 performs a Fourier transform on the input acoustic signal for each frame, and a band spectrum intensity calculation unit 122 performs a band spectrum analysis for each predetermined frequency band as shown in FIG. Calculate the strength. At this time, first, the band spectrum intensity obtained from the acoustic signal before the sound is input is calculated as the noise intensity N of each band.
i is stored in the environmental noise intensity storage unit 171.

Next, when speech is input, the band spectrum intensity P
Calculate i. The band spectrum intensity Pi for the voice calculated by the band spectrum intensity calculation unit 122 is given to the section determination unit 181 and, as shown in FIG. 3, the environmental noise intensity N stored in the environmental noise intensity storage unit 171.
The interval [Pi-Ni, Pi + Ni] from the difference from i and the sum
Ask for.

FIG. 4 (b) is a diagram showing a time series of Pi-Ni (hereinafter, referred to as Ai), Pi + Ni (hereinafter, referred to as Bi), and Pi. In this section [Ai, Bi], As shown in FIG. 3C, the values are obtained for each of the bands 1 to n. The section determined by the section determination unit 181 is provided to the calculation formula determination unit 142, where the calculation formula used in the DP local distance calculation unit 141 is determined by any of the following Examples 1 to 4. .

In the following equation, Ti is the band spectrum intensity in each band of the template stored in the template storage unit 131, Ai and Bi are the above-mentioned sections, and Di is the local distance in each band. <Example 1> Ti <Ai OR Bi <Ti Di = (Ti-Pi) ² Ai ≤ Ti ≤ Bi Di = Ci where Ci is a constant <Example 2> Ti <Ai OR Bi <Ti Di = (Ti−Pi) ² Ai ≦ Ti ≦ Bi Di = (Ti−Pi) ² × Ci where Ci is a constant smaller than 1. <Example 3> Di <(Ti−Pi) if Ti <AiORBi <Ti ² Ai ≦ Ti ≦ Bi Di = 0 <Example 4> Ti <Ai Di = (Ai−Ti) ² Ai ≦ Ti ≦ Bi Di = 0 Bi <Ti Di = (Ti−Bi) ^{2 The} local distance calculation unit 141 for the DP
Template Ti for each voice stored in
Is read, the local spectrum Di between the template for each voice and the band spectral intensity of the voice input in each band i is calculated using the equations of Examples 1 to 4, and the sum ΣDi for each band i is calculated. , Local distance ΣD
Find i.

Here, when the calculation formula used in the DP local distance calculation unit 141 is the above example 1, FIG.
As shown in (b), when the template Ti1 is Bi <Ti1, the distance Di is calculated by the difference between the band spectrum intensity Pi of the voice and the template Ti1.
When the template Ti2 satisfies Ai≤Ti2≤Bi, the distance Di is a constant distance Ci.

Since the above-mentioned distance Di is obtained for each frequency band i as shown in FIG. 4C, for example, as shown in the band i = 1 in FIG. For i, the local distance Di is a constant distance Ci
The range in which the local distance Di is a constant distance Ci is narrow in, for example, a frequency band i in which the band intensity of the noise is weak as indicated by the band i = k in FIG. It will be.

That is, when the distance between the template Ti and the band spectral intensity Pi of the input voice is short, in a band where the noise band intensity is strong, the distance Di is set to a certain distance Ci.
Strict distance calculation is not performed, and, for example, in a frequency band i where the band intensity of noise is weak as indicated by a band i = k in the figure, the calculation of the distance Di is performed by the following formula: Di = (Ti−Pi) ² , the sum ΣDi is determined mainly by the distance of the band having a small noise band intensity, and is a value relatively unaffected by noise.

Similarly, when the local distance ΣDi is calculated using the equations of Examples 2 and 3, it is possible to obtain a local distance relatively unaffected by noise for the same reason. The local distance ΣDi obtained as described above is provided to the DP cumulative distance calculation unit 151, and the distance between each template and the band spectrum of the input voice is obtained by dynamic programming.

That is, the horizontal axis is the time series of the input voice,
When the vertical axis is the time series of the template, the local distance calculation unit 141 for DP is added to each grid point formed by the vertical axis and the horizontal axis.
Is applied, the optimal path with the smallest cumulative value of the local distance is determined, and the cumulative distance of the determined optimal path is defined as the distance between each template and the input voice. The minimum distance search unit 152 searches the cumulative distance obtained as described above for a template in which the distance between the band spectral intensity of the input voice and the template is the minimum distance, and outputs the label of the searched template as a label. Output from the unit 161.

As described above, in this embodiment, a section is obtained from the sum and difference between the band spectrum intensity of the noise and the band spectrum intensity of the input voice, and when the template is included in this section, the local distance is calculated. Since the cumulative distance is determined as being constant or 0, the distance between the template and the input voice can be determined relatively unaffected by noise.

In the above embodiment, the section in the section determining unit 181 is [Pi−Ni, Pi + Ni]. However, the section determined in the section determining unit is not limited to the above embodiment. For example, [Pi
−αNi, Pi + αNi] or using the function F (x), [F (Pi) −F (Ni), F (Pi) + F
(Ni)] can also be used.

Further, in the above-described embodiment, an example has been described in which the voice is recognized by obtaining the band spectrum intensity of the input voice and the distance between the template. However, the present invention is not limited to the above-described embodiment. The speech can be recognized using the band spectrum intensity of the speech and the similarity of the template (for example, the reciprocal of the distance). Furthermore, in the above-described embodiment, an example is shown in which dynamic programming is used as means for determining the distance between the band spectrum intensity of the input voice and the template. However, the present invention is not limited to the above-described embodiment. The distance or the similarity can be calculated by a known method.

[0038]

As is apparent from the above description,
In the present invention, parameters are determined from the estimated values of the band spectrum of the input voice and the band spectrum of the noise, and a distance or similarity calculation method is performed using the determined parameters.
The stage determines whether the band spectrum intensity of the template is within or outside the interval determined based on the estimated values of the band spectrum of the input voice and the band spectrum of the noise determined as parameters in each band. Since the distance or similarity is calculated by changing the distance or similarity calculation method with respect to the band spectrum of the input audio signal, the influence of noise for each band is taken into account. The distance or similarity can be calculated, and speech recognition with a sufficient recognition rate can be performed even in an environment with a large background noise.

[Brief description of the drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a diagram showing an embodiment of the present invention.

FIG. 3 is a diagram illustrating an example of a configuration of a section determination unit.

FIG. 4 is a diagram for explaining the operation of the embodiment of the present invention.

FIG. 5 is a diagram showing a conventional example.

[Explanation of symbols]

 Reference Signs List 11 sound input means 12 band spectrum calculation means 13 template 14 distance / similarity calculation means 15 utterance content collation means 16 recognition result output means 17 noise band spectrum estimation means 18 parameter determination means 111 microphone 112 analog / digital converter 121 FFT calculation unit 122 Band spectrum intensity calculation unit 131 Template storage unit 141 Local distance calculation unit for DP 142 Calculation formula determination unit 151 DP cumulative distance calculation unit 152 Minimum distance search unit 161 Label output unit 171 Environmental noise intensity storage unit 181 Section determination unit 182 Addition unit 183 Subtraction unit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-5-73090 (JP, A) JP-A-61-100878 (JP, A) JP-A-61-65299 (JP, A) JP-A-59-1985 99497 (JP, A) JP 2-4920 (JP, B2) JP-T-63-502304 (JP, A) US Patent 4630305 (US, A) (58) Fields investigated (Int. Cl. ⁷ , DB) Name) G10L 15/00-17/00 WPI (DIALOG)

Claims (6)

(57) [Claims] 1. A sound input means for inputting an audio signal mixed with noise, a band spectrum calculating means for calculating a band spectrum from an audio signal input from the sound input means, and a band spectrum of a template and an input audio signal. distance
A distance calculation means for calculating a voicing content collating means for collating the utterance contents based on the calculated distances by the distance calculating means and the recognition result output means for outputting the result of the collation by the utterance contents matching means as a recognition result A speech recognition apparatus comprising: a noise band spectrum estimating means for estimating a band spectrum of noise mixed at the time of speech input; and a parameter determining means for determining parameters from an estimated value of the band spectrum of the input speech and the band spectrum of the noise. provided, using the parameters determined by the parameter determining means, said distance calculation means, in each band, Te
The band spectral intensity of the template
Determined input speech band spectrum and noise band spectrum
Within or outside the section determined based on the estimated value of
Depending on whether or not there is
A speech recognition apparatus characterized in that a distance is calculated by changing a distance calculation method with respect to a vector, and a speech signal containing noise is recognized. 2. A sound input for inputting an audio signal mixed with noise.
Band spectrum from the sound signal input from the input means and the sound input means.
Band spectrum calculating means for calculating the band spectrum of the template and the input voice signal
Similarity calculating means for calculating the degree of similarity, and the similarity calculated by the similarity calculating means .
The utterance content collating means for collating the contents and the result collated by the utterance content collation means are regarded as the recognition result.
The speech recognition apparatus and a recognition result output means for outputting Te
To estimate fraud and mitigating risk band spectrum of the noise to be mixed at the speech input
Noise band spectrum estimating means for estimating the band spectrum of the input speech and the band spectrum of the noise.
Parameter determining means for determining a parameter from a fixed value
Provided, using the parameters determined by the parameter determining means, the similarity calculation means, in each band, the band spectrum intensity template, the estimation of the band spectrum of band spectrum and noise of the input speech is determined as a parameter Change the method of calculating the similarity between the template and the band spectrum of the input audio signal depending on whether it is within the interval determined based on the value or outside the interval, calculate the similarity, and recognize the audio signal mixed with noise A speech recognition device characterized by the following. 3. In each band, when the band spectrum intensity of the template is within an interval determined based on the estimated values of the band spectrum of the input voice and the band spectrum of the noise determined as parameters, the template and the input voice signal 2. A speech recognition apparatus according to claim 1 , further comprising a distance calculating means for setting a distance from the band spectrum to a constant value. 4. In each band, when the band spectrum intensity of the template is within an interval determined based on the estimated values of the band spectrum of the input voice and the band spectrum of the noise determined as parameters, the template and the input voice signal 3. A speech recognition apparatus according to claim 2 , further comprising a similarity calculating means for setting a similarity with the band spectrum of the predetermined value to a constant value. 5. In each band, when the band spectrum intensity of the template is within a section determined based on the estimated values of the band spectrum of the input voice and the band spectrum of the noise determined as parameters, the template and the input voice signal 2. The speech recognition apparatus according to claim 1 , further comprising a distance calculation unit that sets a distance from the band spectrum to 0. 6. Each of the band, according to claim characterized by having a parameter determining means for determining an interval as a parameter on the basis of the sum and the difference between the estimated value of the band spectrum of band spectrum and the noise of the input speech 1 , 2,
The speech recognition device according to claim 3, 4, or 5 .