JP2003241792A - Device and method for speech recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To prevent a speech recognition rate from decreasing even in noisy environment having a frequency spectrum close to that of a speech. <P>SOLUTION: A speech taking-in part 1 takes in at least a speech spoken by a user and generates a speech signal, a noise superposition part 2 inputs a noise signal having a gradient of signal level to variation in frequency from a superposed noise data output part 3 and superposes it on the speech signal taken in by the speech input part 1 to reduce the signal peak of a speech other than the speech spoken by the user, and a speech recognition part 4 performs speech recognition processing with a speech signal on which the noise superposition part 2 superposes the noise signal. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition apparatus and method for performing voice recognition processing when performing word recognition, sentence recognition, emotion recognition, etc. using voice, and more particularly to a voice recognition device similar to the voice of a television or radio. The present invention relates to a voice recognition device and method suitable for use in an environment where a noise signal exists.

[0002]

2. Description of the Related Art Conventionally, in a voice recognition system,
It is known that voice recognition processing is performed on a voice signal captured by a microphone or the like. In such a voice recognition system, in order to reduce erroneous recognition of voice recognition, noise removal processing such as spectrum subtraction is performed as preprocessing of the voice recognition processing, and learning data of the voice model and voice data for recognition are used. The noise data was superposed on. As such a voice recognition system, for example, Japanese Patent Laid-Open No. 9-1980.
The technique disclosed in Japanese Patent Publication No. 79 is known.

This speech recognition apparatus has a structure in which the same noise data is superimposed on the learning data of the speech model and the speech data for recognition, and the frequency spectrum of the speech data for recognition on which noise data is superimposed and the speech model of the speech model. The speech recognition rate is improved by making the frequency spectrum of the learning data similar.

[0004]

However, in the conventional voice recognition system, usually, in addition to the voice data to be recognized, the ambient noise is included in the voice data detected by the microphone, etc. This affects the voice recognition process and causes a reduction in the voice recognition rate. In particular, when the environmental noise is noise having a frequency spectrum similar to that of a user's voice such as television or radio, the voice recognition rate is remarkably reduced, and the interface becomes inconvenient for the user.

FIG. 16 shows a conventional voice recognition system. In this voice recognition system, the voice capturing unit 101
The voice and environmental noise captured by the signal amplification unit 10
In step 2, the signal is amplified to an appropriate level, sampled and quantized by the A / D converter 103, and then directly output to the voice recognition unit 104. When the environmental noise signal resembles a human voice, such as news on television or radio, it is sent to the voice recognition unit 104 with the frequency spectrum as it is and the temporal shift of the frequency spectrum. Then, the voice recognition unit 10
In 4, the Fourier transform processing by the Fourier transform unit 111 is performed on the voice data including the environmental noise, and the feature amount extraction unit 1
The feature extraction is performed by 12, and the likelihood calculation unit 115 using the acoustic model of the acoustic model storage unit 113 and the language model of the language model storage unit 114 performs likelihood calculation and speech recognition. As a result, the voice recognition of the voice recognition unit 104 is adversely affected, resulting in erroneous recognition.

Therefore, the present invention has been proposed in view of the above-mentioned circumstances, and a speech recognition apparatus and method capable of preventing a reduction in speech recognition rate even in a noise environment having a frequency spectrum similar to speech. The purpose is to provide.

[0007]

In order to solve the above-mentioned problems, in a voice recognition apparatus according to the present invention, a voice capturing means for capturing at least a voice uttered by a user to generate a voice signal, and the above voice capturing. A noise superimposing means for superimposing a noise signal having a slope of a signal level with respect to a change in frequency on a voice signal taken in by the means to reduce a signal peak of a voice other than a voice emitted from a user; And a voice recognition unit for performing voice recognition processing on the voice signal on which the noise signal is superimposed by the superimposing unit.

Further, in the voice recognition method according to the present invention, at least a voice uttered by a user is taken in to generate a voice signal, and a noise signal having a signal level inclination with respect to a change in frequency is added to the voice signal. A noise signal that superimposes and flattens the frequency spectrum other than the voice emitted from the user is superimposed, and voice recognition processing is performed on the voice signal on which the noise signal is superimposed.

[0009]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings.

The present invention is applied to a voice recognition device configured as shown below.

[Basic Configuration Example of Voice Recognition Device] As shown in FIG. 1, this voice recognition device includes a voice capturing unit 1, a noise superimposing unit 2, and a superposed noise data output unit 3 each including, for example, a microphone exposed to the outside. The voice recognition unit 4 is provided.

The voice capturing unit 1 is composed of, for example, a condenser microphone, and captures, for example, a user voice uttered by a user located immediately before the voice recognition device and also captures environmental noise around the user. This environmental noise has a level lower than that of the user's voice to be recognized, but is a component not necessary for voice recognition, and includes, for example, a television around the user or a voice emitted from a person outside the user. The voice capturing unit 1 generates voice data including user voice and environmental noise and outputs the voice data to the noise superimposing unit 2.

The superimposed noise data output unit 3 outputs the prepared superimposed noise data or the successively generated superimposed noise data to the noise superimposing unit 2. The superposed noise data is, for example, data stored in a storage mechanism such as a ROM (Read Only Memory) or data obtained by arithmetically processing the stored data. This superposed noise data has a frequency spectrum that is flatter than changes in the frequency spectrum of the user voice and changes in the frequency spectrum of the environmental noise and has a small temporal change. For example, noise whose sound pressure level changes according to the frequency band. , It has a steady spectrum such as pink noise in which the sound pressure level decreases as the frequency band increases.

The noise superimposing unit 2 performs a process of superimposing the superposed noise data from the superposed noise data output unit 3 on the voice data from the voice capturing unit 1, and the voice data on which the superposed noise data is superposed is recognized by the voice recognition unit 4. Output to.

The noise superimposing section 2 synthesizes the voice data of the temporal sound pressure level change shown in FIG. 2A and the superposed noise data of the temporal sound pressure level change shown in FIG. 2B. At this time, the two time-series data are sequentially compared, and the one having the larger absolute value is selected as the superimposed audio data. As a result, the noise superimposing unit 2, as shown in FIG.
At time t2, t4, and t6, voice data is used, and at time t
At 1, t3 and t5, the superimposed noise data is used to create voice data to be output to the voice recognition unit 4. Note that, in the case shown in FIG. 2, an example in which the voice data and the superimposed noise data are consequently alternately selected and the voice data to be output to the voice recognition unit 4 is created is shown, but the present invention is not limited to this.

Here, the noise superimposing unit 2 receives, for example, for each voice frame, superimposed noise data at the same time as voice data to be recognized. The noise superimposing unit 2 compares the absolute value of the sound pressure level of the voice data with the absolute value of the sound pressure level of the superimposed noise data, and outputs the one with the higher sound pressure level as the voice data. The output voice data does not become larger than the input voice data, and even if the input voice data is close to the maximum value of the input range of the voice recognition unit 4, the digit of the data increases due to the superimposition processing. Therefore, the feature amount of voice data used for voice recognition is not lost.

Further, the noise superimposing unit 2 weights the voice data and the superposed noise data not only in the case of superimposing as described with reference to FIG.
You may superimpose so that it may become the audio data output to. That is, the noise superimposing unit 2 sets the voice data as “x1”,
When the superimposed noise data is “x2”, the value y of the voice data to be output to the voice recognition unit 4 is determined by performing the calculation y = (ax1 + bx2). Here, a in the above equation is a weighting coefficient for audio data, and b is a weighting coefficient for superimposed noise data.

The voice recognition section 4 performs a voice recognition process using the voice data from the noise superposition section 2 to generate a voice recognition result. At this time, the voice recognition unit 4 performs a short-time Fourier transform for each voice frame to extract the peak of the frequency spectrum and its temporal change as a feature amount of voice recognition,
Performs voice recognition processing and likelihood calculation processing using an acoustic model and a language model.

As shown in FIG. 3, when the voice capturing unit 1 captures a voice, the voice recognition apparatus thus configured generates voice data and outputs it to the noise superimposing unit 2 (step S1). The superposition unit 2 superimposes the superposition noise data from the superposition noise data output unit 3 on the voice data (step S2), and then the voice recognition unit 4 performs a voice recognition process using the voice data on which the superposition noise data is superposed. And outputs the voice recognition result.

Specifically, a voice recognition apparatus receives a voice including a voice signal component A and an environmental noise component B of a temporal change in sound pressure level as shown in FIG. As shown in FIG. 5, the frequency spectrum resulting from the Fourier transform of such an audio signal has a frequency spectrum peak A1 and a frequency spectrum peak A2 due to the user audio component A and a frequency spectrum due to the environmental noise component B in a certain period. Peak B1 of

Next, FIG. 6 shows a temporal change in the sound pressure level when the noise superimposing unit 2 superimposes the superimposed noise data on the voice data captured by the voice capturing unit 1.
As a result, the user voice component A, the environmental noise component B, and the superimposed noise component C are superimposed on the voice data. As shown in FIG. 7, the frequency spectrum peaks A1 and A2 of the frequency spectrum due to the user voice component A remain in the frequency spectrum resulting from the Fourier transform of such audio data, but the peak B1 of the frequency spectrum due to the environmental noise component B remains. Can be seen to disappear due to the frequency spectrum C1 due to the superimposed noise component C.

That is, the noise superimposing section 2 superimposes on the voice data the superposed noise data whose sound pressure level is lower than that of the user voice component A and whose frequency spectrum is flatter than that of the environmental noise component B. The frequency spectrum has a shape such that environmental noise is buried in superposed noise.

By superimposing the superposed noise data on the voice data in this way, the environmental noise component is flattened in frequency, and the superposed noise data is superposed continuously on the input voice data to obtain the time. The environmental noise component is flattened and output to the voice recognition unit 4. It should be noted that the superimposed noise data only needs to be able to flatten the frequency spectrum of environmental noise.

That is, as shown in FIG. 7, the signal has a slope with respect to a change in frequency such that the sound pressure level gradually decreases in the high frequency band and the sound pressure level is inversely proportional to the frequency. May be.

As described above, according to the speech recognition apparatus in the basic configuration example, the peak of the frequency spectrum of the environmental noise component can be flattened by superimposing the noise superposition, and the environmental noise component can be recognized by the speech recognition. The voice recognition process can be performed by the unit 4 without extracting the feature of the user voice. Therefore, according to this voice recognition device, erroneous recognition of voice recognition is reduced even if, for example, a television or the like exists around the user and the voice capturing unit 1 captures the voice emitted from the television or the like. be able to.

Further, according to this speech recognition apparatus, since the speech data on which the superimposed noise data is superimposed is input using the input channel of the speech recognition unit 4, for example, the acoustic model of the speech recognition unit 4 or its learning algorithm is used. Does not affect, and has an advantage that it can be applied by using various voice recognition methods in addition to the effect of reducing erroneous recognition of voice recognition regardless of the voice recognition algorithm of the voice recognition unit 4.

[Second Configuration Example of Voice Recognition Device] Next, a voice recognition device according to a second configuration example to which the present invention is applied will be described with reference to FIG. The same parts as those of the basic configuration example described above are designated by the same reference numerals, and detailed description thereof will be omitted.

The speech recognition apparatus according to the second configuration example differs from the basic configuration example in that it includes a superposed noise storage unit 11 and a superposed noise selection unit 12 instead of the superposed noise data output unit 3.

The superposed noise storage unit 11 stores a plurality of superposed noise data. The superposed noise storage unit 11 stores a plurality of superposed noise data as a data pattern in which a plurality of frequency spectra and a plurality of sound pressure levels are combined.

The superposed noise selection unit 12 selects a data pattern of superposed noise data to be superposed on voice data in the noise superposed unit 2 and outputs at least one superposed noise data to the noise superposed unit 2.

According to the speech recognition apparatus according to the second configuration example, a plurality of types of superposed noise data are created and stored in the superposed noise storage unit 11 in advance, and the superposed noise selection unit 12 is provided.
Since optimum superposed noise data can be selected and output to the noise superimposing unit 2 by means of the above, in addition to the effect of the basic configuration example, environmental noise can be further reduced to reduce false recognition of the voice recognition unit 4. You can

[Third Configuration Example of Voice Recognition Device] Next, a voice recognition device according to a third configuration example to which the present invention is applied will be described with reference to FIG. The same parts as those in the above-described configuration example are designated by the same reference numerals, and detailed description thereof will be omitted.

The speech recognition apparatus according to the third configuration example includes an environmental noise capturing unit 2 that captures environmental noise around the speech recognition apparatus.
1 and an environmental noise measuring unit 22 that measures an environmental noise level around the voice capturing unit 1, which is different from the second configuration example.

The environmental noise capturing unit 21 captures environmental noise around the voice recognition device during the period in which the voice recognition device is not performing voice recognition processing, creates environmental noise data, and outputs the environmental noise data to the environmental noise measurement unit 22.

The environmental noise measuring section 22 measures the average sound pressure level of the environmental noise captured by the environmental noise capturing section 21 and outputs it to the superimposed noise selecting section 12 as average noise information.

The superposed noise selection unit 12 selects superposed noise data having an appropriate sound pressure level based on the average noise information and reads it from the superposed noise storage unit 11 when the voice recognition device recognizes the voice data. . Specifically, the superposed noise selection unit 12 selects superposed noise data that flattens the peak of the frequency spectrum of the environmental noise, based on the average sound pressure level of the environmental noise indicated by the average noise information. It is output to the noise superimposing unit 2.

According to such a voice recognition device, even if the average sound pressure level of the environmental noise captured by the voice capturing unit 1 is high, the superimposed noise data corresponding to the environmental noise can be selected and superimposed. . That is, according to this speech recognition device, the superimposed noise data used by the superimposed noise selecting unit 12 based on the correlation between the average noise information obtained by the environmental noise measuring unit 22 and the superimposed noise data stored in the superimposed noise storage unit 11. Select. Therefore, according to this speech recognition device, it is possible to select superposed noise data having a sound pressure level substantially the same as the peak of the frequency spectrum of the environmental noise component, and to reliably flatten the peak of the frequency spectrum of the environmental noise component. Therefore, the erroneous recognition of the voice recognition unit 4 can be surely reduced.

Here, the effect of erroneous recognition by superimposing the superimposed noise data on the environmental noise differs depending on the sound pressure level of the environmental noise and the average sound pressure level of the superimposed noise data. That is, if the average sound pressure level of the superimposed noise data is too small compared to the environmental noise level, the peak of the frequency spectrum of the environmental noise component cannot be flattened, and the average sound pressure level of the superimposed noise data is If it is too large compared to the noise level, the peak height of the frequency spectrum of the user voice component becomes small. Therefore, by selecting the superimposed noise data in accordance with the environmental noise level measured by the environmental noise measuring unit 22, it is possible to further reduce erroneous recognition.

Further, according to this speech recognition apparatus, since the environmental noise capturing section 21 is provided separately from the speech capturing section 1, the average sound pressure level of the environmental noise regardless of whether the speech is recognized or not. Can be measured.

[Fourth Configuration Example of Voice Recognition Device] Next, a voice recognition device according to a fourth configuration example to which the present invention is applied will be described with reference to FIG. The same parts as those in the above-described configuration example are designated by the same reference numerals, and detailed description thereof will be omitted.

The voice recognition apparatus according to the fourth configuration example differs from the second configuration example in that the voice recognition unit 4 and the superimposed noise selection unit 12 are connected.

As described above, the voice recognition unit 4 calculates the likelihood of the voice recognition result together with the voice recognition result of the user voice. In the fourth configuration example, the voice recognition unit 4 outputs likelihood information to the superposed noise selection unit 12.

The superposed noise selection unit 12 receives the likelihood information from the voice recognition unit 4 and selects the superposed noise data stored in the superposed noise storage unit 11 based on the likelihood of voice recognition.
The superposed noise selection unit 12 does not change the superposed noise data used in the noise superposition unit 2 when it is determined that the likelihood is high and the voice recognition is accurately performed, for example. On the other hand, for example, the superposed noise selection unit 12 selects superposed noise data different from the superposed noise data used in the noise superposition unit 2 when it is determined that the likelihood is low and the voice recognition is not accurately performed. And outputs it to the noise superimposing unit 2.

According to such a voice recognition device, since the superposed noise selecting unit 12 selects the superposed noise data having the higher likelihood, the likelihood of the voice recognition by the voice recognizing unit 4 is further increased, False recognition of the voice recognition unit 4 can be further reduced.

[Fifth Configuration Example of Voice Recognition Device] Next, a voice recognition device according to a fifth configuration example to which the present invention is applied will be described with reference to FIG. The same parts as those in the above-described configuration example are designated by the same reference numerals, and detailed description thereof will be omitted.

The speech recognition apparatus according to the fifth configuration example has a signal amplification section 31, A between the speech acquisition section 1 and the noise superposition section 2.
It is different from the speech recognition apparatus according to the first configuration example in that the / D conversion unit 32 and the noise removal unit 33 are provided. In another configuration example, if the signal amplification unit 31 and the A / D conversion unit 32 are not shown, either digital audio data or analog audio signal may be used. And

In this voice recognition device, the voice capturing unit 1
The audio signal generated by the above is amplified by the signal amplifier 31 to a predetermined sound pressure level, converted into digital audio data by the A / D converter 32, and output to the noise eliminator 33.

The noise removing unit 33 performs noise removal processing such as spectral subtraction on the voice data from the A / D conversion unit 32 to remove the environmental noise component from the voice data, and the noise superimposing unit 2 receives the noise. Output. Then, the noise superimposing unit 2 superimposes the superimposed noise data on the voice data that has been subjected to the noise removing process by the noise removing unit 33.

According to such a voice recognition device, the environmental noise can be effectively reduced by the synergistic effect of the effect of the first configuration example and the noise removal effect of the noise removal unit 33, and further the voice recognition. False recognition of the unit 4 can be reduced. That is, before the noise superimposing unit 2 superimposes the superimposed noise data, the noise removing unit 33 removes the environmental noise, and the noise removing unit 33 removes the environmental noise and superimposes it on the voice data including the remaining environmental noise. Therefore, the noise superimposing unit 2
Is to be superimposed on the voice data whose environmental noise component level is lower than that of the voice data from which the ambient noise has not been removed by the noise removing unit 33, and the sound pressure level of the superposed noise data to be used is reduced. can do.

Therefore, according to this voice recognition device,
As a result of being able to reduce the influence on the user voice component due to the superposition of the superimposed noise data on the voice data, the peak of the frequency spectrum of the user voice can be reliably extracted by the voice recognition unit 4, and the voice recognition can be further performed. False recognition of the unit 4 can be reduced.

[Sixth Configuration Example of Voice Recognition Device] Next, a voice recognition device according to a sixth configuration example to which the present invention is applied will be described with reference to FIG. The same parts as those in the above-described configuration example are designated by the same reference numerals, and detailed description thereof will be omitted.

The speech recognition apparatus according to the sixth configuration example differs from the above-described configuration example in that it includes a superposed noise data generation unit 41 for generating superposed noise data, instead of the superposed noise selection unit 12.

In the sixth configuration example, the superposed noise storage unit 11 stores superposed noise data used as a template in the superposed noise data generation unit 41. The superposed noise data stored in the superposed noise storage unit 11 may be prepared in plural as described above, or may be single.

The superposed noise data generation unit 41 uses the superposed noise data stored in the superposed noise storage unit 11 as a template, modifies the superposed noise data read from the superposed noise storage unit 11 and outputs the modified superposed noise data to the noise superposition unit 2. Specifically, the superimposed noise data generator 41 includes an amplifier 42 and a filter 43,
The control unit (not shown) sets the amplification factor of the amplifier 42 and the pass frequency band of the filter 43.

In the superposed noise data generator 41 as described above, in a state where the amplification factor of the amplifier 42 and the pass frequency band of the filter 43 are set, the sound of the entire superposed noise data from the superposed noise storage unit 11 is output by the amplifier 42. The pressure level is converted, the sound pressure level is converted by the filter 43 for each frequency band of the superimposed noise data, and the superimposed noise data is transformed.

According to such a voice recognition device, the superimposed noise data stored in the superimposed noise storage unit 11 is used as a template,
The overall sound pressure level and frequency spectrum component can be changed, and various types of superposed noise data can be generated without storing other types of superposed noise data in the superposed noise storage unit 11. The capacity of the superposed noise storage unit 11 can be reduced.

Further, according to this voice recognition device, the amplification factor of the amplifier 42 and the pass frequency band of the filter 43 are controlled according to the voice recognition result of the voice recognition unit 4, the sound pressure level of environmental noise, and the frequency spectrum. By adjusting the sound pressure level in all frequency bands, it is possible to control the slope of the sound pressure level with respect to the frequency, and generate the optimum superimposed noise data to reduce the environmental noise for further speech recognition. False recognition of the unit 4 can be reduced.

[Seventh Configuration Example of Voice Recognition Device] Next, a voice recognition device according to a seventh configuration example to which the present invention is applied will be described with reference to FIG. The same parts as those in the above-described configuration example are designated by the same reference numerals, and detailed description thereof will be omitted.

In the voice recognition apparatus according to the seventh configuration example, the voice signal captured by the voice capturing unit 1 is used as the signal amplifying unit 31.
Also, it is converted into voice data by the A / D converter 32 and input to the noise superimposing unit 2. Further, this speech recognition apparatus measures the environmental noise at the time of non-speech recognition by the environmental noise measuring unit 22 and supplies the average noise information to the superimposed noise selecting unit 12,
The superposed noise selection unit 12 selects superposed noise data based on the average noise information and reads it from the superposed noise storage unit 11,
The noise is generated by the superposed noise data generation unit 41 and input to the noise superposed unit 2. At this time, the amplifier 42 is operated by the voice recognition unit 4
The likelihood information from is input and the amplification factor is set according to the likelihood. The noise superimposing unit 2 superimposes the superposed noise data from the superposed noise data generating unit 41 on the voice data from the A / D conversion unit 32, and supplies the superposed noise data to the voice recognition unit 4.

The voice recognition unit 4 receives the voice data from the noise superposition unit 2 and generates a voice recognition result and likelihood information. The Fourier transform unit 51, the feature amount extraction unit 52, and the acoustic model storage unit 53. , A language model storage unit 54, and a likelihood calculation unit 55.

When the voice data is input from the noise superposition unit 2, the voice recognition unit 4 performs a Fourier transform process by the Fourier transform unit 51 to create the frequency spectrum of the voice data shown in FIG. 7, and the feature quantity extraction unit. The peak of the frequency spectrum is detected by 52, and the frequency band corresponding to the peak is extracted as a feature amount. When the feature amount is input from the feature amount extraction unit 52, the likelihood calculation unit 55 compares the feature amount with the acoustic model stored in the acoustic model storage unit 53 and selects the most approximate acoustic model. The speech recognition unit 4 selects the acoustic model in time series by performing such processing for each speech frame, and compares the acoustic model in time series with the language model stored in the language model storage unit 54. Create a speech recognition result. The likelihood calculator 55 also creates likelihood information indicating the likelihood of the speech recognition result based on the comparison result between the acoustic model and the feature amount and the comparison result between the time-series acoustic model and the language model. Output to the amplifier 42.

According to such a voice recognition device, the superposed noise data stored in the superposed noise storage unit 11 is selected based on the average noise information measured by the environmental noise measuring unit 22, and the selected superposed noise data is selected. Is transformed according to the likelihood, it is possible to further reduce the environmental noise and reduce the erroneous recognition of the voice recognition unit 4.

[Eighth Configuration Example of Speech Recognition Device] Next, a speech recognition device according to an eighth configuration example to which the present invention is applied will be described with reference to FIG. The same parts as those in the above-described configuration example are designated by the same reference numerals, and detailed description thereof will be omitted.

The speech recognition apparatus according to the eighth configuration example is a superposed noise signal generation circuit 61 for creating an analog superposed noise signal, and an adder 62 for adding the voice signal from the signal amplifier 31 and the superposed noise signal. The configuration is different from the above-described configuration example.

The superposed noise signal generation circuit 61 is connected to the noise generation transistor 71 connected to the drive circuit, the D / A conversion circuit 72 connected to the voice recognition unit 4, the noise generation transistor 71 and the D / A conversion circuit 72. The signal amplification circuit 73 is provided.

The noise generating transistor 71 is composed of, for example, a small signal transistor, and generates a thermal noise signal when a reverse bias is applied between the base and the emitter by the driving circuit. The thermal noise signal generated by the noise generating transistor 71 is transferred to the noise signal amplifier 73 via the circuit connection.
Entered in.

The D / A conversion circuit 72 inputs data having different levels according to the likelihood from the voice recognition unit 4, performs D / A conversion processing, and outputs an analog signal according to the likelihood to the signal amplification circuit 73. Output to.

The signal amplification circuit 73 amplifies the thermal noise signal from the noise generation transistor 71 according to the analog signal level from the D / A conversion circuit 72 and outputs it as a superimposed noise signal to the adder 62. In this way, the signal amplification circuit 73
Generates an overlapping noise signal having a sound pressure level according to the likelihood by changing the amplification factor according to the likelihood.

The adder 62 creates a voice signal as shown in FIG. 6 by adding the voice signal from the signal amplifier 31 and the superposed noise signal from the superposed noise signal generation circuit 61, Output to the / D converter 32. This adder 6
2 corresponds to the noise superimposing unit 2 described above, and functions as noise superimposing means. The audio signal created by the adder 62 is
The voice data is converted into voice data by the A / D conversion unit 32 and becomes the target of voice recognition by the voice recognition unit 4.

According to the voice recognition device having such a configuration, the RO for storing the superposed noise data created in advance is stored.
M, superimposing the superposed noise on the audio signal with an inexpensive configuration without requiring a signal processing unit for superimposing the noise such as a DSP (Digital Signal Processor) that creates superposed noise data or superimposes the digital data. You can

Further, in the eighth configuration example, when a superposed noise signal having different sound pressure levels depending on the frequency band is created, as shown in FIG. 15, an analog signal is provided between the signal amplification circuit 73 and the adder 62. A filter 74 is provided to control the pass frequency band of the analog filter 74. As a result, it is possible to create a superposed noise signal whose sound pressure level is inclined in the high frequency band.

The above-described embodiment is an example of the present invention. For this reason, the present invention is not limited to the above-described embodiment, and other than this embodiment, as long as it does not deviate from the technical idea of the present invention, various types according to the design etc. Of course, it is possible to change.

[0073]

According to the speech recognition apparatus and method of the present invention, at least speech uttered by a user is captured to generate a speech signal, and the speech signal has a signal level gradient with respect to a change in frequency. A noise signal is superimposed to flatten the frequency spectrum other than the voice emitted by the user. A noise signal is superimposed on the speech signal, and speech recognition processing is performed on the speech signal. It is possible to prevent the voice recognition rate from decreasing even in the presence of environmental noise having a spectrum.

[Brief description of drawings]

FIG. 1 is a block diagram showing a basic configuration of a voice recognition device to which the present invention is applied.

2A and 2B are diagrams for explaining processing when a noise superimposing unit superimposes voice data and superposed noise data, where FIG. 2A is a temporal change of the voice data, and FIG. (C) is a temporal change after superimposing the voice data and the superimposed noise data.

FIG. 3 is a flowchart showing a basic processing procedure of a voice recognition device to which the present invention is applied.

FIG. 4 is a diagram showing a relationship between a sound pressure level and time in audio data captured by an audio capturing unit.

FIG. 5 is a diagram showing a relationship between a sound pressure level and a frequency in audio data captured by an audio capturing unit.

FIG. 6 is a diagram showing the relationship between the sound pressure level and time after superimposing the superimposed noise data on the voice data by the noise superimposing unit.

FIG. 7 is a diagram for explaining another example of the relationship between the sound pressure level and the frequency when the superimposed noise data is superimposed on the audio data.

FIG. 8 is a block diagram showing a second configuration example of a voice recognition device to which the present invention has been applied.

FIG. 9 is a block diagram showing a third configuration example of a voice recognition device to which the present invention has been applied.

FIG. 10 is a block diagram showing a fourth configuration example of a voice recognition device to which the present invention has been applied.

FIG. 11 is a block diagram showing a fifth configuration example of a voice recognition device to which the present invention has been applied.

FIG. 12 is a block diagram showing a sixth configuration example of a voice recognition device to which the present invention has been applied.

FIG. 13 is a block diagram showing a seventh configuration example of a voice recognition device to which the present invention has been applied.

FIG. 14 is a block diagram showing an eighth configuration example of a voice recognition device to which the present invention has been applied.

FIG. 15 is a block diagram showing another circuit of the eighth configuration example of the voice recognition device to which the present invention is applied.

FIG. 16 is a block diagram showing a configuration of a conventional voice recognition system.

[Explanation of symbols]

1 Audio capture unit 2 Noise superimposition section 3 Superposed noise data output section 4 Speech recognition section 11 Superposed noise memory 12 Superposed noise selector 21 Environmental noise capture unit 22 Environmental noise measurement section 31 signal amplifier 32 A / D converter 33 Noise removal unit 41 Superposed noise data generator 42 amp 43 filters 51 Fourier transform unit 52 Feature Extraction Unit 53 Acoustic model storage 54 Language Model Storage 55 Likelihood calculator 61 Superposed noise signal generation circuit 62 adder 71 Noise generation transistor 72 D / A conversion circuit 73 Signal amplification circuit

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Masaru Nakamori             1048, Kadoma, Kadoma-shi, Osaka Matsushita Electric Works             Within the corporation (72) Inventor Haruhiro Kuboyama             1048, Kadoma, Kadoma-shi, Osaka Matsushita Electric Works             Within the corporation (72) Inventor Toshitaka Fujii             1048, Kadoma, Kadoma-shi, Osaka Matsushita Electric Works             Within the corporation F-term (reference) 5D015 EE05

Claims (11)

[Claims] 1. A voice capturing means for capturing at least a voice uttered by a user to generate a voice signal, and a voice signal captured by the voice capturing means,
A noise superimposing unit that superimposes a noise signal having a slope of a signal level with respect to a change in frequency to reduce a signal peak of a voice other than a voice uttered by a user, and the noise signal is superposed by the noise superimposing unit. A voice recognition device comprising: a voice recognition means for performing voice recognition processing on a voice signal. 2. The voice recognition device according to claim 1, wherein the noise superimposing means superimposes a noise signal whose sound pressure level gradually decreases in a high frequency band on a voice signal. 3. A noise storage means for storing a plurality of noise patterns is further provided, and the noise superimposing means selects a noise pattern stored in the noise storage means and superimposes a noise signal on a voice signal. The voice recognition device according to claim 1. 4. The apparatus further comprises noise measuring means for measuring the environmental noise level around the user from the audio signal generated by the audio capturing means, wherein the noise superimposing means sets the environmental noise level measured by the noise measuring means. 4. The voice recognition device according to claim 3, wherein the noise pattern stored in the noise storage means is selected in response to the noise pattern. 5. The voice recognition device according to claim 3, wherein the noise superimposing means selects a noise pattern stored in the noise storing means based on a recognition result of the voice recognizing means. 6. The apparatus further comprises noise creating means for modifying a template noise signal to create a plurality of types of noise signals, wherein the noise superimposing means superimposes the noise signal created by the noise creating means on a voice signal. The voice recognition device according to claim 3, wherein 7. The apparatus further comprises noise measuring means for measuring the ambient noise level around the user from the voice signal generated by the voice capturing means, and the noise creating means sets the ambient noise level measured by the noise measuring means. 7. The voice recognition device according to claim 6, wherein the template noise signal is transformed in accordance with the template noise signal. 8. The voice recognition device according to claim 6, wherein the noise generation unit transforms the model noise signal based on the recognition result of the voice recognition unit. 9. A noise removing means for removing environmental noise of the audio signal generated by the audio capturing means, wherein the noise superimposing means adds a noise signal to the audio signal from which the environmental noise is removed by the noise removing means. The voice recognition device according to claim 1, wherein the voice recognition device is superimposed. 10. The noise superimposing means compares the audio signal level from the audio capturing means with the noise signal level to be superimposed, and selects a signal having a large absolute value of the signal level as the superimposed audio signal. The voice recognition device according to claim 1, wherein 11. A voice signal generated by capturing at least a voice uttered by a user, generating a voice signal, superimposing a noise signal having a signal level inclination with respect to a change in frequency on the voice signal, A voice recognition method comprising: superimposing a noise signal for flattening a frequency spectrum other than voice, and performing voice recognition processing on the voice signal on which the noise signal is superposed.