JPH06110492A - Speech recognition device - Google Patents

Abstract

PURPOSE:To provide a means which effectively remove a noise component other than a speech that a user vocalizes. CONSTITUTION:This device is provided with a vowel detection processing part 4 which detects a vowel part in spectrum data in a certain word section and a noise decision processing part 5 which decides whether the word section includes a speech vocalized by the user or a background noise according to information regarding the presence/absence state or the rate of the vowel part in the spectrum data in the word section received from the vowel detection processing part 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for discriminating between a user's uttered voice and a sudden and loud noise that is difficult to distinguish from a voice uttered by a user in a voice input device.

In recent years, voice recognition devices have been gradually put into practical use, and can be used as auxiliary input means.
Also, due to FA and other factors, computers are being introduced into environments where conventional input means such as keyboards cannot be used, and a voice recognition device is being operated as a new input means to replace the conventional input means. Required to do so. Therefore, it is important to put into practical use a voice recognition device that enables input with high accuracy and a high recognition rate.

[0003]

2. Description of the Related Art In a conventional voice recognition apparatus, the presence or absence of an input sound or the volume of an input sound volume is mainly used as a reference for detecting an input voice. Therefore, the conventional voice recognition device operates only when some sound is detected or when a sound volume in a certain level range is detected.

However, as described above, the voice recognition device is required to operate in a high noise environment such as an FA environment. Therefore, the background of the voice recognition device is almost the same as a user's vocal sound such as an impact sound. There are many loud noises. Therefore, the voice recognition device often targets such noise as well.

[0005]

As described above, in the conventional voice recognition apparatus, the background noise is erroneously recognized as the input voice and the meaningless processing / execution result is output. Therefore, there has been a problem that the operation of a system having a voice recognition device is frequently hindered.

The present invention has been made in view of such conventional problems, and improves the performance of a voice recognition device by effectively eliminating input noise other than the voice uttered by the user. Then, it aims at the smooth operation of the system having the voice recognition device.

[0007]

According to the invention, the above mentioned objects are achieved by means of the patent claims.

That is, according to the first aspect of the present invention, a sampling processing unit for sampling input voice for each unit time to generate voice data, and a word section detection processing unit for detecting a valid word section as input data from the voice data. A vowel detection processing unit for detecting a vowel part from the spectrum data of a certain word section in a device having a spectrum analysis processing unit for analyzing frequency components of voice data and generating spectrum data, and the vowel detection processing. Based on the information received from the department regarding the presence or absence or the ratio of the vowel portion included in the spectrum data of a certain word section, whether the word section is the vocalized voice of the user,
Alternatively, it is a voice recognition device provided with a noise discrimination processing unit for discriminating whether or not it is caused by background noise.

Further, the invention of claim 2 is provided with a means for holding spectral data of each vowel in advance, and a mathematical distance scale and a distance are provided between the spectral data of each vowel and the input spectral data. It is a voice recognition device provided with a vowel detection processing section for detecting a vowel part from the inputted spectrum data by using a threshold value.

Further, the invention according to claim 3 is that a sampling processing unit for sampling input voice for each unit time to generate voice data, and a word section detection processing unit for detecting a word section effective as input data from the voice data. And a noise discrimination processing unit that determines whether the word segment is due to the voice of the user or the impact background noise by checking the steepness of the rising edge based on the magnitude of the volume change at the beginning of the word segment. It is a voice recognition device comprising.

According to a fourth aspect of the present invention, a power calculation section for generating a power value from the waveform amplitude amount of the input voice, a sampling processing section for sampling the input voice for each unit time to generate voice data, and a voice processing section A word section detection processing unit that detects a valid word section as input data from the data,
Noise that determines whether the word segment is due to the user's uttered voice or impulsive background noise by checking the steepness of the rising edge based on the magnitude of the difference between the power values of the input sounds at the beginning and immediately preceding part of the word segment. A voice recognition device including a discrimination processing unit.

According to a fifth aspect of the present invention, a power calculation section for generating a power value from the waveform amplitude amount of the input voice, a sampling processing section for sampling the input voice for each unit time to generate voice data, and a voice processing section A word section detection processing unit that detects a valid word section as input data from the data,
From the magnitude of the difference in the power value of the input sound between the beginning of the word section and the first maximum power value section, see the steepness of the rising edge to determine whether the word section is due to the user's vocal sound or impact background noise. A speech recognition apparatus including a noise discrimination processing unit for discriminating.

Further, according to the invention of claim 6, a sampling processing unit for sampling the input voice for each unit time to generate voice data, and a word section detection processing unit for detecting a word section effective as the input data from the voice data. , And the frequency component of the voice data is analyzed as a vector value with an appropriate degree in the frequency direction, and a spectrum analysis processing unit that generates spectrum data and a difference in the vector length of the spectrum data at the beginning and immediately preceding portion of the word section A speech recognition apparatus comprising a noise discrimination processing unit for judging whether the word segment is due to a user's uttered voice or due to a shocking background noise by observing the steepness of rising from large and small.

According to a seventh aspect of the present invention, a power calculation section for generating a power value from the waveform amplitude amount of the input voice, a sampling processing section for sampling the input voice for each unit time to generate voice data, and a voice processing section A word section detection processing unit that detects a valid word section as input data from the data,
The frequency component of the voice data is analyzed as a vector value that takes an appropriate degree in the frequency direction, and the spectrum analysis processing unit that generates the spectrum data and the vector length of the spectrum data at the beginning of the word section and the first maximum power value part The speech recognition device includes a noise discrimination processing unit that determines whether the word segment is due to the user's uttered voice or impactive background noise by observing the steepness of rising from the difference.

According to the invention of claim 8, a sampling processing section for sampling the input voice for each unit time to generate voice data, and a word section detection processing section for detecting a valid word section as input data from the voice data. , The spectrum analysis processing unit that analyzes the frequency components of the voice data and generates the spectrum data, and the sharpness of the rising edge based on the magnitude of the spectrum change at the beginning of the word section, and that word section is the uttered voice of the user. And a noise discrimination processing unit for discriminating whether it is due to impact background noise.

According to the invention of claim 9, a power calculation section for generating a power value from the waveform amplitude amount of the input voice, a sampling processing section for sampling the input voice for each unit time to generate voice data, A word section detection processing unit that detects a valid word section as input data from the data,
A spectrum analysis processing unit that analyzes the frequency components of the voice data and generates spectrum data, and the word section is used based on the sharpness of the rising from the magnitude of the spectrum change at the beginning and the first maximum power value part of the word section. A voice recognition device comprising: a noise discrimination processing unit for discriminating whether the voice is a person's uttered voice or a shocking background noise.

According to a tenth aspect of the present invention, a sampling processing unit for sampling input voice for each unit time to generate voice data, and a word section detection processing unit for detecting a valid word section as input data from the voice data. , And the frequency component of the voice data is analyzed as a vector value with an appropriate degree in the frequency direction, and a spectrum analysis processing unit that generates spectrum data, and the distance between vectors of the spectrum data at the beginning and immediately preceding portion of the word section Big and small
The speech recognition apparatus includes a noise discrimination processing unit that determines whether the word segment is due to the voiced voice of the user or the impact background noise in view of the steepness of rising.

According to the invention of claim 11, a power calculation unit for generating a power value from the waveform amplitude amount of the input voice, a sampling processing unit for sampling the input voice at every unit time to generate voice data, and a voice processing unit A word segment detection processing unit that detects a valid word segment as input data from the data, and a spectrum analysis processing unit that analyzes the frequency component of the voice data as a vector value having an appropriate order in the frequency direction and generates spectrum data. And, from the size of the distance between the vectors of the spectrum data at the beginning of the word section and the first maximum power value part, see the steepness of the rising edge,
A speech recognition apparatus comprising: a noise discrimination processing unit that discriminates whether the word section is due to a voiced voice of a user or due to impact background noise.

FIG. 1 is a diagram for explaining the principle of the present invention. In the figure, the sampling processing unit 1 generates voice data by converting an analog signal of input voice into a digital signal for each unit time. The word section detection processing unit 2 detects a word section effective as input data in the voice data from the volume information of the voice data. The spectrum analysis processing unit 3
The frequency component of the voice data is analyzed and the spectrum data is generated.

Further, the vowel detection processing section 4 detects a vowel part from the spectrum data by judging the similarity with the vowel spectrum data prepared in advance. The noise discrimination processing unit 5 determines whether the word segment detected by the word segment detection processing unit 2 is due to the input voice of the utterer or is due to background noise depending on the presence or absence of a vowel part included in the word segment. Determine if it is a thing.

FIG. 3 also illustrates the principle of the present invention. In the figure, a sampling processing unit 21 and a word section detection processing unit 2
2 and the spectrum analysis processing unit 23 are the same as those shown in FIG. The noise discrimination processing unit 25 discriminates, for example, whether the detected word segment is the input voice of the speaker or the shocking background noise based on the magnitude of the volume change or the spectrum change at the beginning of the word segment.

[0022]

As shown in FIG. 1, according to the present invention, the spectrum data in the word section is monitored by the vowel detection processing section 4, and the spectrum data which can be judged as a vowel is added with the information that it is a vowel part.

It is generally known that in many languages including Japanese, a word is surely accompanied by a vowel phoneme, but the noise discrimination processing section 5 of FIG. 1 utilizes this property. That is, it is checked whether or not there is a portion to which information indicating that it is a vowel portion is added in the word section, or if it is extremely small, and if it is, this is discriminated as noise.

On the other hand, if there is additional information for a certain vowel part or more, it is judged that the voice is uttered by the user and the operation is started. As described above, the voice recognition device according to the present invention excludes the background noise when it is input, and outputs the recognition result only when the user's uttered voice is input.

Further, as shown in FIG. 3, in the present invention, the word section detection processing section 22 detects the word section based on the voice data generated by the sampling processing section 21, and the spectrum analysis section 23 detects the spectrum data. Is output. The noise discrimination processing unit 25 uses the voice data, the word section,
With reference to the spectrum data, the power change amount and the spectrum change amount are calculated from the voice data and the spectrum data, which are the first element of the word section and the element immediately before it.

Then, the likelihood determination with the impact background noise is performed for each variation. It is generally known that impact background noise shows a relatively sharp power or spectrum change with respect to voice (human voice). Therefore,
In this judgment, if it is possible to evaluate the likelihood of noise to be large, the word section is rejected as if there was input of impulsive background noise, and if the likelihood is small, the word section is Part of the voice data is processed by the voice recognition processing unit 26 of the next stage.
Output to.

As described above, in the voice recognition device according to the present invention, when the impact noise is input, it is surely eliminated, so that the recognition result can be output only for the voice uttered by the user. become.

[0028]

FIG. 2 is a block diagram of an embodiment of the present invention. In the figure, the vowel detection processing unit 14 has vowel spectrum data 17 for various vowels in advance, and calculates the similarity with each unit portion of the input spectrum data 16 obtained from the input voice. As a result, it is determined whether or not a unit portion of the input spectrum data 16 is a vowel, and the vowel portion is detected.

The display of the detection result is performed, for example, by adding vowel information 18 indicating that a certain unit part of the input spectrum data 16 is a vowel part to the input spectrum data 16.

The noise discrimination processing section 15 refers to the spectrum data of the range included in the word section detected by the word section detection processing section 12, and regards this as the portion to which the vowel information 18 regarding the presence or absence of the vowel is added. , And the part that was not added. Then, for example, if the portion to which the vowel information 18 is added exceeds a certain threshold value, the data is regarded as a significant input voice, and if it is within the threshold value, it is determined to be background noise.

In the above embodiment, the noise discrimination processing section 15
In the above, the sum of the parts to which the vowel information 18 indicating the vowel parts is added is used as the criterion for discriminating the noise, but other than this, for example, the ratio occupying the entire word section may be used as the discrimination criterion. The sampling processing unit 11 and the spectrum analysis processing unit 13 operate in the same manner as those having the same names shown in FIG.

FIG. 4 is a block diagram of another embodiment of the present invention. In FIG. 4, the word section detected by the word section detection processing unit (not shown) and the spectrum data output by the spectrum analysis unit are'spectral data with word section information '.
Is passed to the noise discrimination processing section 35.

This "spectral data with word section information" is represented by a vector of strength in each dimension, which has an appropriate dimension on the frequency axis. When the dimension of this vector is infinite, the sum of the values of each element becomes equal to the integrated value of the original spectrum data, so the sum of each element of the vector is used as the vector length, and this is used as the pseudo power.

The noise discrimination processing section 35 refers to the data at the beginning of the word section and the data immediately before it, and calculates the difference in vector length as the power change amount and the inter-vector distance as the spectrum change amount. For each of the power and spectrum, a power threshold value and a spectrum threshold value are set in advance so that the distributions of both voice and impulsive background noise can be separated.

Then, the respective threshold values are compared and evaluated with the power change amount and the spectrum change amount. For example, when both are evaluated as noise, it is assumed that the impact background noise is input and the word section is determined. Reject. If not rejected, the spectrum data of the word section is passed to the voice recognition processing unit (not shown) as'voice discriminated data '.

In the above embodiment, the noise discrimination processing section 35.
Although the difference in vector length is used as the amount of power change calculated by, the volume of the original input voice, that is, the difference in power value obtained from the amount of waveform amplitude may be used instead. The power value may be calculated by providing a power calculation processing unit.

Further, as the spectrum data obtained by the spectrum analysis unit, a vector of strength in each dimension having an appropriate dimension on the frequency axis is taken, and as a spectrum change amount obtained by the noise discrimination processing unit, a vector distance is calculated. However, these may use other spectral expression parameters, such as cepstrum and the distance between the parameters.

Further, the noise discrimination processing unit 35 refers to the data at the beginning of the word section and the data immediately before it as the object of measuring the variation amount. This is the first maximum power value data at the beginning of the word section. And the data immediately before the word section may be referred to. In addition, the noise discrimination processing unit 3
Although the logical product of both the power and spectrum determination results was used as the noise determination criterion in 5, the logical sum may be used if necessary.

[0039]

As described above, according to the present invention,
Even if background noise is mixed in the input voice,
Since only background noise can be excluded, unnecessary recognition processing will not be erroneously executed. Therefore, there is an effect that it is possible to avoid outputting meaningless recognition results, and it largely contributes to improvement of operation efficiency of the system using the voice recognition device.

[Brief description of drawings]

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

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

FIG. 3 is a diagram illustrating the principle of the present invention.

FIG. 4 is a configuration diagram of another embodiment of the present invention.

[Explanation of symbols]

 1,11,21 Sampling processing unit 2,12,22 Word section detection processing unit 3,13,23 Spectral analysis processing unit 4,14 Vowel detection processing unit 5,15,25,35 Noise discrimination processing unit 16 Input spectrum data 17 Vowel spectrum data 18 Vowel information 26 Speech recognition processing unit

Claims (11)

[Claims] 1. A sampling processing unit (1) for sampling input voice every unit time to generate voice data.
And a word section detection processing unit (2) that detects a valid word section as input data from the voice data, and a spectrum analysis processing unit (3) that analyzes frequency components of the voice data and generates spectrum data. In vowel detection, a vowel detection processing unit (4) for detecting a vowel part from the spectrum data of a certain word section, and the presence or absence of a vowel part included in the spectrum data of a certain word section received from this vowel detection processing section (4). Alternatively, a noise discrimination processing unit (5) for deciding whether the word section is caused by a voice uttered by the user or caused by background noise based on the information about the ratio is provided. Voice recognition device. 2. A method for holding spectral data of each vowel in advance, and using the mathematical distance scale and the distance threshold between the spectral data of each vowel and the input spectral data, the input The voice recognition device according to claim 1, further comprising a vowel detection processing unit that detects a vowel portion from the spectrum data. 3. A sampling processing section for sampling input speech at every unit time to generate speech data, a word section detection processing section for detecting a word section effective as input data from the speech data, and a beginning section of the word section. And a noise discriminating processing unit for discriminating whether the word segment is due to the user's uttered voice or due to impact background noise, by observing the steepness of the rise from the magnitude of the volume change in Voice recognition device. 4. A power calculation unit that generates a power value from the waveform amplitude amount of the input voice, and a sampling processing unit that samples the input voice for each unit time to generate voice data.
The word section detection processing unit that detects a valid word section as input data from the voice data, and the word section is used based on the sharpness of the rising edge based on the difference in the power value of the input sound at the beginning and immediately preceding section of the word section. A speech recognition apparatus, comprising: a noise discrimination processing unit that discriminates whether the speech is generated by a person's voice or an impact background noise. 5. A power calculation unit for generating a power value from the waveform amplitude amount of the input voice, a sampling processing unit for sampling the input voice for each unit time, and generating voice data.
The word section detection processing unit that detects a valid word section as input data from the voice data, and the difference in the power value of the input sound between the beginning of the word section and the first maximum power value part, the sharpness of the rising edge A speech recognition apparatus, comprising: a noise discrimination processing unit that discriminates whether a word segment is generated by a user's voice or impact background noise. 6. A sampling processing section for sampling input voice for each unit time to generate voice data, a word section detection processing section for detecting a word section effective as input data from the voice data, and a frequency component of the voice data. Is analyzed as a vector value with an appropriate order in the frequency direction, and the spectrum analysis processing unit that generates spectrum data and the difference in the vector length of the spectrum data at the beginning and immediately preceding part of the word section The speech recognition apparatus, further comprising: a noise discrimination processing unit that discriminates whether the word section is due to the voiced voice of the user or due to impulsive background noise. 7. A power calculation unit for generating a power value from the waveform amplitude amount of the input voice, a sampling processing unit for sampling the input voice every unit time to generate voice data,
A word section detection processing unit that detects a valid word section as input data from voice data, and a spectrum analysis process that analyzes the frequency component of the voice data as a vector value with an appropriate order in the frequency direction and generates spectrum data. Part and the difference between the vector lengths of the spectrum data at the beginning of the word section and the first maximum power value section, the sharpness of the rising edge is checked to see if the word section is due to the user's uttered voice or the impactive background noise. A voice recognition device, comprising: a noise discrimination processing unit for discriminating whether the noise is due to 8. A sampling processing section for sampling input voice for each unit time to generate voice data, a word section detection processing section for detecting a word section effective as input data from the voice data, and a frequency component of the voice data. Parses
A spectrum analysis processing unit for generating spectrum data,
It is equipped with a noise discrimination processing unit that judges from the magnitude of the spectral change at the beginning of the word section, the sharpness of the rising edge, to determine whether the word section is due to the voice of the user or due to impulsive background noise. A voice recognition device characterized by: 9. A power calculation unit for generating a power value from the waveform amplitude amount of the input voice, a sampling processing unit for sampling the input voice for each unit time to generate voice data,
A word section detection processing unit that detects a valid word section as input data from the voice data, a spectrum analysis processing unit that analyzes the frequency component of the voice data and generates spectrum data, the beginning and first maximum power values of the word section. From the magnitude of the spectrum change in the part, see the steepness of the rise, and whether the word section is due to the voice of the user,
Alternatively, the speech recognition apparatus is provided with a noise discrimination processing unit for discriminating whether the noise is due to impact background noise. 10. A sampling processing unit for sampling input voice for each unit time to generate voice data, a word section detection processing unit for detecting a word section effective as input data from the voice data, and a frequency component of the voice data. Is analyzed as a vector value with an appropriate order in the frequency direction,
A spectrum analysis processing unit for generating spectrum data,
Noise that determines whether the word segment is due to the user's uttered voice or impulsive background noise by checking the steepness of the rising edge based on the magnitude of the vector distance of the spectrum data at the beginning and immediately preceding portion of the word segment. A voice recognition device comprising a discrimination processing unit. 11. A power calculation unit for generating a power value from a waveform amplitude amount of input voice, a sampling processing unit for sampling the input voice for each unit time to generate voice data, and a voice data effective as input data. A word section detection processing unit that detects a word section, a frequency component of voice data, which is analyzed as a vector value having an appropriate order in the frequency direction, and a spectrum analysis processing section that generates spectrum data. From the size of the vector distance of the spectrum data in the first maximum power value part,
A speech recognition apparatus, comprising: a noise discrimination processing unit for discriminating whether or not the word segment is due to a user's uttered voice or due to a shocking background noise in view of the steepness of rising.