JP2001215992A - Voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition device which surely recognizes inputted voice under various environmental conditions. SOLUTION: Inputted voice is supplied to a spectrum subtracting section 20 through a filter 10 and a spectrum analysis section 12. In the section 20, noise is subtracted from the inputted voice and the result is supplied to a feature extracting section 22. A noise difference section 14 computes the difference between input noise and the noise generated while learning a voice dictionary 26. The section 20 cancels the difference between the input noise and the noise of the dictionary 26 by subtracting the difference from the inputted voice. A subtracting magnification used in the subtraction is determined based on the SNR of the inputted voice and the difference spectrum from the section 14. The magnification is also determined for every analysis frame.

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 technique for recognizing speech generated under noise.

[0002]

2. Description of the Related Art Conventionally, there has been known a technique of recognizing speech even under noise by comparing features of speech obtained by subtracting noise from input speech with standard speech previously obtained by learning.

For example, a noise suppression device disclosed in Japanese Patent Application Laid-Open No. H11-154000 and a speech recognition system using the device include a noise power spectrum calculated from a power spectrum calculated based on an input signal in a voice section. There is described a technique for performing speech recognition by subtracting the result of multiplication by a subtraction coefficient to eliminate the influence of noise.

[0004]

Generally, in a spectrum subtraction technique for subtracting a noise spectrum, noise is estimated by averaging several tens of frames in a noise section before occurrence, and the estimated noise is input to a speech section. Is subtracted in the frequency domain for each frame (analysis unit).

However, when comparing the input voice from which the influence of noise has been removed in this way with a standard pattern prepared in advance, a voice generated in an environment where some noise exists is used as a standard pattern (voice dictionary). In this case (even if the noise is not controlled, some noise remains because it cannot be completely removed), since the comparison target is sound with noise, a difference occurs between the two and the recognition rate may decrease. There is.

In the above prior art, the subtraction magnification is set to a value larger than 1. This is because the estimated noise is averaged, but the sound is adjusted to a sound in a section with a large power. In consideration of the fact that the recognition rate is improved as a whole, there is also a problem that if the subtraction magnification is increased even in a section where the power is small, distortion due to excessive noise is generated and the recognition rate is reduced.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned problems of the prior art, and has as its object to surely recognize an input voice even when a standard pattern to be compared is a pattern uttered under noise. It is another object of the present invention to provide an apparatus capable of suppressing a decrease in recognition rate even under various environments.

[0008]

In order to achieve the above object, the present invention provides a speech recognition apparatus for recognizing a feature of a speech obtained by subtracting noise from an input speech by comparing it with a standard speech obtained by learning. And calculating means for calculating the noise to be subtracted from the input voice based on a difference between the learning noise included in the standard voice and the input noise included in the input voice. By performing the difference calculation in consideration of the noise included in the learning, it is possible to effectively suppress a decrease in the recognition rate.

The present invention also relates to a speech recognition apparatus for recognizing a speech by comparing an input speech with a standard speech obtained by learning, wherein the learning noise contained in the standard speech and the input noise contained in the input speech are recognized. Calculating means for calculating the noise to be added to the standard sound based on the difference from the above. By performing the addition operation in consideration of the noise included in the learning, it is possible to effectively suppress a decrease in the recognition rate.

Here, it is preferable to further comprise means for determining a ratio to be subtracted or a ratio to be added according to the SNR of the input voice. As the noise level increases, the utterance level also increases in proportion to the noise level.
SNR considering noise level (noise power)
By determining the ratio to be subtracted and the addition ratio in step (1), the recognition rate can be improved irrespective of the magnitude of the audio power. Here, the SNR is defined by the ratio between the voice power and the noise power.

Preferably, the SNR of the input voice is calculated based on weighting in the frequency domain, and more specifically, it is desirable to perform a filtering process based on human auditory characteristics.

A ratio to be subtracted for each of the different spectral bands or according to the power variance of the input noise;
Alternatively, it is preferable to further include means for determining a ratio to be added. By changing the ratio for each spectrum band, the recognition rate can be improved in all the bands, and by determining the ratio in accordance with the power variance of the input noise, the recognition rate is improved using the Lambert effect. be able to.

It is also preferable that the ratio is determined based on the SNR or the power of the input noise for each voice analysis frame. Since the noise changes for each analysis unit (frame), the recognition can be performed with higher accuracy by changing the ratio for each analysis unit.

[0014]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the overall configuration of this embodiment. The input voice input from the microphone is supplied to the spectrum analysis unit 12 via the filter 10. When no sound is input, noise is supplied to the spectrum analysis unit 12 via the filter 10. The filter 10 is a filter in consideration of human auditory characteristics, and specifically, is a filter that preferentially transmits a high frequency region. The filter 10 is not always required,
The voice or noise input from the microphone may be directly supplied to the spectrum analyzer 12.

The spectrum analyzer 12 analyzes the spectrum of the input voice or noise by FFT or the like, and calculates the power for each frequency. The calculated spectrum is smoothed and supplied to the noise difference unit 14.

The noise difference section 14 includes a spectrum analysis section 1
2, the input noise spectrum (a spectrum in a section in which no sound is input from the microphone and the noise is input, which is estimated to be noise included in the sound) is supplied, and the learning sound to be compared is provided. A learning noise spectrum is supplied from a database 18 that stores learning noise data included when the dictionary is uttered. The noise difference unit 14 calculates a difference between these two spectra, that is, the input noise spectrum and the learning noise spectrum, and sets the difference as an estimated noise spectrum. Specifically, the SNR (estimated noise) of the estimated noise spectrum is

SNR (estimated noise) = SNR2−SNR1 (1) Here, SNR1 is the SNR of the learning noise spectrum, and SNR2 is the SN of the input noise spectrum.
R. Here, the SNR is defined as a ratio (Speech to Noise Ratio) between the power of the voice section and the power of the noise section, and specifically,

SNR = 10Log (ΣP (Si) / Σ (i)) / (ΣP (Nj) / Σ (i)) (2) The SNR of the input noise spectrum can be calculated from the ratio of the noise power obtained by the analysis by the spectrum analysis unit 12 to the voice power obtained by the utterance experimental value. The same applies to the SNR of the learning noise spectrum.

As described above, when the difference between the input noise spectrum and the learning noise spectrum is calculated to calculate the difference between the two spectra, the calculation result is supplied to the spectrum subtraction unit 20.

In the spectral subtraction section 20,
The difference between the input voice pattern (input signal spectrum in the voice section) supplied through the filter 10 and the spectrum analysis unit 12 and the estimated noise supplied from the noise difference unit 14 is calculated, and the voice from which the influence of noise has been removed is calculated. The pattern is extracted and supplied to the feature extracting unit 22.

The feature extraction unit 22 extracts a feature portion from the input voice pattern from which the influence of noise has been removed, and supplies the extracted feature portion to the phoneme recognition unit 24. In the phoneme recognition unit 24, the features extracted based on the speech dictionary 26 prepared by learning in advance (noise at the time of learning is added to the speech pattern of the speech dictionary) and the acoustic model 28 are assigned to any phoneme. It verifies whether it is applicable and recognizes and outputs phonemes.

FIG. 2 schematically shows the difference calculation in the noise difference section 14. In the figure, (a) shows the SNR (SNR1) of the learning noise spectrum,
(B) shows the SNR (SNR2) of the input noise spectrum. In the noise difference unit 14, these two supplied
Based on the two SNRs, a difference amount to be spectrally subtracted is calculated based on the above equation (1).

FIG. 3 schematically shows the state of the difference in the spectral subtraction section 20. The difference between the input voice spectrum (solid line in the figure) supplied through the filter 10 and the spectrum analysis unit 12 and the estimated noise (SNR2−SNR1 in the figure, dashed line in the figure) supplied from the noise difference unit 14 is calculated. Thereby, the difference between the noise at the time of learning and the noise at the time of voice input is canceled, and it is possible to accurately collate with the voice data recorded in the voice dictionary 26.

It should be noted that the above-described processing may be considered to subtract the input noise from the input voice, add learning noise to the result obtained by subtracting the input noise, and collate with the noise-added voice data recorded in the voice dictionary 26. . That is, if the above-described processing is expressed by a mathematical formula, (input voice)-{(input noise)-(learning noise)} = (input voice)-(input noise)
+ (Learning noise), and it can be understood that even if the noise at the time of learning is added to the speech dictionary, it can be recognized without being affected by the noise at the time of learning.

On the other hand, the spectrum subtraction section 20
In the case where the estimated noise is subtracted from the input voice and the subtraction magnification, which is the subtraction magnification, is fixed, it is difficult to stably improve the recognition rate under various environments as described above. More specifically, the subtraction magnification becomes too large in a section where the power is small, and distortion is caused by excessive noise, resulting in a reduction in the recognition rate.

Therefore, in the present embodiment, a subtraction magnification setting unit 30 is further provided, and the estimated noise output from the noise difference unit 14 is multiplied by the subtraction magnification α and supplied to the spectrum subtraction unit 20.

The subtraction magnification setting unit 30 basically changes the subtraction magnification dynamically in accordance with the power of the input voice. Generally, as shown in FIG. 4, when the noise level increases, the utterance level increases. Also, since there is a so-called Lambert effect, which increases almost in proportion to the noise level, it is difficult to set an optimum subtraction magnification. Therefore, in the present embodiment, as shown in FIG.
R is calculated by the SNR calculation unit 34, and the calculated SNR is supplied to the subtraction ratio setting unit 30, and the subtraction ratio setting unit 30 sets the subtraction ratio based on the SNR of the input voice. Specifically, the subtraction magnification is set to increase as the SNR of the input sound increases. Rather than simply changing the subtraction magnification according to the power of the input voice, by changing the subtraction magnification according to the SNR of the input voice, high-precision voice recognition in consideration of the Lambert effect becomes possible. Excessive pulling in the section where the power is small can be reliably prevented.

It is known that, even when noise is included, there are a band where the recognition rate is greatly reduced and a band where the degree of deterioration is relatively small. That is, there are a band that is strong against noise and a band that is weak against noise. For example, the applicant of the present application
It has been confirmed that when the noise spectrum exists in the range of Hz to 3 kHz, the recognition rate is greatly reduced compared to the case where the noise spectrum exists in another band. Therefore, by using a high-pass filter or a low-pass filter or the like to extract a signal of only a specific band from an input voice pattern and performing voice recognition, voice recognition can be performed with high accuracy even in a noisy environment. However, since the spectrum and power of noise change in various ways, the configuration of speech recognition using a fixed band-pass filter or the like cannot flexibly cope with environmental changes. There is a risk of lowering.

Therefore, in this embodiment, the subtraction magnification is changed for each band to flexibly cope with various driving environments. Therefore, as shown in FIG. 1, the estimated noise output from the noise difference unit 14 is supplied to the subtraction magnification setting unit 30 and the subtraction magnification setting unit 30
Here, the subtraction magnification is determined for each spectral band of the estimated noise based on the noise pattern / magnification conversion table 36, and the difference to be subtracted by the spectrum subtraction unit 20 is determined. The noise pattern / magnification conversion table 36 is for determining the noise pattern and the subtraction magnification for each band at that time in advance and holding the same in a table format.
For example, the subtraction magnification at 1 kHz to 3 kHz is set to be larger than that of the other bands.

FIG. 5 schematically shows the processing in the subtraction magnification setting section 30. (A) and (c)
7A is an example of the spectrum of the estimated noise output from the noise difference section 14, FIG. 7A is an example of a spectrum that is relatively flat, and FIG. 7C is an example of a spectrum in which much power exists on the low frequency side.
(B) is a subtraction magnification determined for each band when (a) is input, and (d) is a subtraction magnification determined for each band when (c) is input.
Basically, the subtraction magnification is changed according to the power of the estimated noise (that is, the subtraction magnification is increased as the power increases), but the subtraction magnification is further reduced for a band where the recognition rate of noise is relatively low. Is increasing. As described above, by determining the subtraction magnification for each of the estimated noises, that is, the spectrum band of the difference between the input noise spectrum and the learning noise, any driving environment,
That is, any noise pattern can be recognized with high accuracy.

Note that the subtraction magnification αi for each band is
In particular

Α i = β i · P i (3) Here, βi is a coefficient of band i experimentally obtained, Pi is an estimated noise power of band i, and i is a frequency band.

Further, in the present embodiment, as shown in FIG. 1, the average power and the variance (or deviation) of the input noise emphasized in the high frequency range by the filter 10 are calculated by the power calculation unit 32, and the subtraction magnification setting unit is used. 30. In the subtraction magnification setting unit 30,
The subtraction magnification is determined from the deviation of the power peak value from the average value, that is, the power variance value. The larger the variance, the larger the subtraction magnification is set, and the smaller the variance, the smaller the subtraction magnification.

FIG. 6 shows the relationship between the power spectrum of the input noise and the deviation. In the figure, the dotted line indicates the average value of the input noise power over time, and σ1 and σ2 indicate the deviation of the peak value from the average value. σ1> σ2, and
The subtraction magnification in the case of the deviation σ1 is set to be larger than the subtraction magnification in the case of the deviation σ2. This allows
When the input noise power is low, even if the Lambert effect with a low utterance level occurs, the subtraction magnification is not unnecessarily increased and distortion due to excessive noise is not generated, and the recognition rate can be improved.

In the above embodiment, the case where the subtraction magnification is determined over the entire utterance section has been described. However, it is also preferable to determine the subtraction magnification for each analysis frame for speech recognition. For example, a microphone has two inputs, and an SNR for each analysis frame is calculated using a signal from one input. Then, the subtraction magnification is determined based on the SNR in frame units. As a result, the subtraction magnification can be set for each analysis unit, and the speech recognition rate can be further improved. Of course, when the subtraction magnification is determined for each analysis frame, it is also preferable to calculate the difference between the input noise and the learning noise for each analysis frame and determine the difference based on the SNR. It is also preferable to change the magnification based on the power of each analysis frame instead of the SNR.

As described above, in the embodiment of the present invention, a case has been described where the characteristics of a voice obtained by subtracting noise from an input voice are compared with a voice dictionary. It is also possible to add to the data in 26 and compare with the input voice, and both are technically equivalent. The magnification when adding the different data to the voice dictionary 26 is also the same as the subtraction magnification.
And power can be determined.

FIG. 7 is a block diagram showing the configuration in this case. The difference from FIG. 1 is that the estimated noise calculated by the noise difference unit 14 is supplied to the spectrum addition unit 21, and the estimated noise, that is, the input noise is added to the learning speech data stored in the speech dictionary 26 by the spectrum addition unit 21. And the difference between learning noise. In addition, the magnification at the time of adding to the audio data of the voice dictionary 26, that is, the addition magnification is determined by the addition magnification setting unit 31 (corresponding to the subtraction magnification setting unit 30 in FIG. 1). , Specifically, S of the input voice
The magnification is determined for each NR, power variance, or spectrum band of the estimated noise.

[0036]

As described above, according to the present invention, even when standard speech is learned in a noisy environment, input speech can be reliably recognized. Further, under an arbitrary traveling environment in which noise changes variously, it is possible to suppress a decrease in the traveling recognition rate.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram of an embodiment.

FIG. 2 is an explanatory diagram of noise difference processing.

FIG. 3 is an explanatory diagram of spectrum subtraction.

FIG. 4 is an explanatory diagram showing the Lambert effect.

FIG. 5 is an explanatory diagram for determining a subtraction magnification for each spectrum band.

FIG. 6 is a graph showing the variance of input audio power.

FIG. 7 is a configuration block diagram of another embodiment.

[Explanation of symbols]

Reference Signs List 10 filter, 12 spectrum analysis unit, 14 noise difference unit, 18 learning noise database, 20 spectrum subtraction unit, 22 feature extraction unit, 24 phoneme recognition unit, 26 speech dictionary, 28 acoustic model database, 30 subtraction magnification setting unit, 32 power Calculator, 34 SNR calculator, 36 Noise pattern / magnification conversion table.

Claims (8)

[Claims] 1. A speech recognition apparatus for recognizing a feature of a speech obtained by subtracting noise from an input speech by comparing the feature with a standard speech obtained by learning, wherein a learning noise included in the standard speech and the input speech are recognized. And a calculating means for calculating the noise to be subtracted from the input voice based on a difference from the input noise included in the voice recognition. 2. A speech recognition apparatus for recognizing an input voice by comparing an input voice with a standard voice obtained by learning, wherein a difference between learning noise included in the standard voice and input noise included in the input voice is determined. And a calculating means for calculating a noise to be added to the standard voice based on the standard voice. 3. The apparatus according to claim 1, further comprising: means for determining a ratio to be subtracted or a ratio to be added in accordance with the SNR of the input voice. Voice recognition device. 4. The apparatus according to claim 3, wherein the SNR of the input speech is calculated based on weighting in a frequency domain. 5. The apparatus according to claim 1, further comprising: means for determining a ratio to be subtracted or a ratio to be added for each of the different spectral bands. Voice recognition device. 6. The apparatus according to claim 1, further comprising: a ratio to be subtracted according to a power variance of the input noise;
Alternatively, a means for determining a ratio to be added is provided. 7. The apparatus according to claim 1, wherein a ratio to be subtracted or a ratio to be added is determined based on an SNR of the input noise for each audio analysis frame. Recognition device. 8. The apparatus according to claim 1, wherein a ratio to be subtracted or a ratio to be added is determined based on the power of the input noise for each voice analysis frame. Recognition device.