JP4301896B2 - Signal analysis device, voice recognition device, program, recording medium, and electronic device - Google Patents

Abstract

A signal analysis device includes: frequency band division means for dividing an input signal into a signal of a plurality of frequency bands; band energy extraction means for extracting band energy in each band for the input signal divided into respective bands; and normalization means for normalizing the extracted band energy for each band and obtaining normalized band energy for each band. With this configuration, it is possible to obtain a stable speech recognition accuracy even under a high-noise environment. Moreover, it is possible to obtain a stable speech recognition accuracy even when speech having no noise section is input or when the noise level is gradually changed during vocalization.

Description

  The present invention relates to a signal analysis device that analyzes an input voice / acoustic signal, a signal processing device, and a voice recognition device using the signal analysis device.

  In the signal analysis means used for speech recognition, a signal is extracted while shifting an interval of about several tens of milliseconds from the input of the signal at an interval of about several milliseconds to several tens of milliseconds to obtain an analysis frame. Next, acoustic parameters are calculated from the waveform of the input signal in each analysis frame to obtain time-series acoustic parameters.

  In specific speaker speech recognition, the obtained time-series acoustic parameters are collated with a time-series pattern (standard pattern) of acoustic parameters registered in advance. The standard pattern most similar to the acoustic parameter of the input signal is taken as the recognition result.

  In the case of speaker-independent speaker recognition, acoustic parameters are calculated from a large amount of data in advance, and statistics of the acoustic parameters calculated for each speech unit are obtained to create a stochastic acoustic model. At the same time, the probabilistic acoustic model for each speech unit is connected to create a word model or sentence model. For the acoustic parameters analyzed from the input signal, the probability likelihood for the word model or sentence model is calculated and compared, and the word model or sentence model with the highest probability likelihood is used as the recognition result. As the speech unit here, for example, a unit such as a phoneme, a syllable, or a word is used.

  Thus, in order to perform speech recognition, the signal analysis technique for obtaining acoustic parameters is an important technique for determining recognition performance. For this reason, a signal analysis technique that is not easily affected by distortions such as noise and line characteristics is required. Non-Patent Document 1 discloses MFCC (Mel Frequency Cepstrum Coefficient), which is used as such a signal analysis technique and converts band energy into a cepstrum coefficient.

  FIG. 1 is a flowchart showing a procedure for obtaining the MFCC. Hereinafter, the MFCC analysis method will be described with reference to FIG. In MFCC, first, a speech waveform is input to the signal analyzer for each analysis frame (step S101). Next, a Hamming window function is applied so as not to cause a sudden change at both ends of the cutout section of the frame (step S102). Next, the energy on the linear frequency axis in each frame is obtained by using FFT (First Fourier Transform) (step S103). The energy on the linear frequency axis is collected for each band equally divided on the mel frequency axis and converted into band energy (step S104). The converted band energy is logarithmically converted for each band (step S105). Next, MFCC is obtained by cosine transforming the power in each band (step S106). The obtained MFCC is output from the signal analyzer (step S107). Thus, the cepstrum coefficient obtained by equally dividing on the mel frequency axis is referred to as a mel frequency cepstrum coefficient (MFCC). The mel frequency is a frequency unit according to human auditory characteristics, and the resolution at a low frequency is finer than the resolution at a high frequency. For this reason, it is known that if the cepstrum of the same order is used, the speech recognition performance is excellent as compared with the case where the linear frequency axis is used.

  In this specification, steps S101 to S105 may be referred to as a frequency analysis step (step S201), and steps S106 to S107 may be referred to as a parameter conversion step (step S202).

  However, speech recognition is actually performed by additive noise such as background noise in homes, offices, cars, outdoors, etc., noise from cooling fans generated by the device itself, and additive noise electromagnetically mixed into the waveform. Done in the presence. Furthermore, there are multiplicative distortions such as transfer characteristics such as a microphone or a voice transmission circuit, and a change in acoustic characteristics due to a difference in distance and angle when a speaker goes to the microphone. MFCC is susceptible to these noises and distortions. For this reason, in speech recognition using MFCC, there is a problem that the recognition performance is significantly lowered when used in a noisy environment or when using a method with large distortion of the line characteristics. In order to reduce the influence of such noise and distortion, a signal analysis method as described below is used.

[SS method and its simplification method]
Non-Patent Document 2 discloses a spectral subtraction (SS) method as a method for reducing the influence of additive noise. In this method, an input speech signal is subjected to frequency analysis to obtain an input amplitude spectrum or power (square) spectrum, an estimated noise spectrum estimated in a noise interval is multiplied by a predetermined coefficient α, and the spectrum subtraction coefficient α is multiplied. The noise component is suppressed by subtracting the estimated noise spectrum after the input from the input spectrum. In this method, in order to improve the voice quality, it is necessary to perform processing by dividing into a large number of bands.
Further, Patent Document 1 discloses a method of improving the SS method and reducing noise components by reducing band division.

[CMS method]
Non-Patent Document 1 discloses a CMS method (Cepstrum Mean Subtraction) as a method for reducing the influence of multiplicative distortion. This method is also called a CMN method (Cepstrum Mean Normalization). This is based on the assumption that the multiplicative distortion is obtained as a long-term average of the cepstrum of the spoken speech. Specifically, by subtracting the average value of the cepstrum coefficient of the input voice from the cepstrum coefficient, the influence of distortion based on the characteristics of the acoustic system such as a line or a microphone can be reduced. This is equivalent to subtracting the distortion of the logarithmic spectrum, which is the cepstrum cosine transform.

  However, in actual input speech, the sound source position of the sound is different from the sound source position of the noise. Since the line characteristics to be normalized are different between the voice section and the noise section, applying the average cepstrum of the voice section to the noise section makes the noise section cepstrum unstable and adversely affects recognition performance. . For this reason, an improved method of the CMS method that can be applied to the noise interval has been studied.

[E-CMN method]
Non-Patent Document 3 and Patent Document 2 propose an E-CMN method that is an improved method of the CMS method. In the E-CMN method, the cepstrum average of the speech segment and the cepstrum average of the non-speech segment are obtained separately, and normalization processing is performed separately for the speech segment and the non-speech segment. This method can reduce the misrecognition rate.

[Voice detection]
The SS method, the simplification method thereof, and the E-CMN method described above require speech detection in order to distinguish between a speech segment and a non-speech segment. As a voice detection technique, a standard method is disclosed in a voice communication standard used for a mobile phone. In speech detection, generally, an input signal is temporally divided into a speech interval and a noise interval based on a change in energy of the input signal in a short time, a change in spectral shape, and the like.

JP 2001-228893 A (Claims, paragraphs 0027 to 0029) JP-A-10-254494 (Claims)

Edited by Kiyohiro Shikano et al., “Voice Recognition System”, first edition, Ohm Co., Ltd., May 15, 2001, p. 13-15 S. Boll, “Suppression of acoustic noise in speech using spectral subtraction”, IEEE Trans. ASSP, 1979, vol. ASSP-27, no. 2, pp. 113-120 Makoto Shokai and two others, “Model Adaptation Method E-CMN / PMC Based on Cepstrum Average Normalization Method and HMM Synthesis Method and Application to Speech Recognition in Cars”, IEICE Transactions, Electronic Information Communication Academic Society, 1997, J80-D-II, No. 10, p. 2636-2644

  However, the signal analysis methods described above and the speech recognition methods used for these analysis methods have the following problems.

[Problems with SS method and its simplification method]
In the SS method, the spectrum of the noise section is subtracted in the voice section, which is excellent when estimating the input voice in an environment with little noise. However, in the noise section, the noise power is small, but the noise spectral shape itself does not change. When performing speech recognition based on statistics that are currently mainstream, if noise having a spectrum shape different from the statistically learned noise is input, the matching accuracy of the noise section is lowered. For this reason, since the noise section is erroneously collated with the voice section, high recognition accuracy as a whole cannot be obtained. In order to prevent this reduction in recognition accuracy, some kind of adaptive spectral correction is required. One of the methods is a method of normalizing the noise spectrum as in the E-CMN method. However, the E-CMN method has the following problems.

[Problems with E-CMN method]
As described above, in the E-CMN method, since the cepstrum average is obtained independently in the voice interval and the noise interval, the line characteristics can be normalized more accurately. In particular, according to this method, since the spectrum shape of the noise section can be flattened, it is possible to improve the matching accuracy of the noise section that cannot be solved by the SS method.
However, the determination of the speech interval and the noise interval depends on the detection accuracy of the speech interval. For this reason, in a high noise environment, there is a problem that speech segmentation accuracy is reduced and normalization is performed based on an incorrect cepstrum average value, which adversely affects recognition accuracy. The following describes problems with voice detection.

[Noise interval length]
In the SS method and the E-CMN method, the noise spectrum is estimated for a section determined as a noise section when a speech section is detected. However, in an environment where the signal analyzer is actually used, if a voice is uttered immediately after the apparatus starts operation, a sufficient noise interval length for estimating the noise spectrum cannot be obtained. For this reason, there is a problem that noise cannot be adapted or wrong noise is adapted.

[Difficulty of detecting speech interval]
In a quiet environment, the speech section and the noise section can be easily separated. For this reason, it is relatively easy to detect the utterance section by using the time change of the voice power. However, when the S / N ratio is low in a noisy environment, the accuracy of the start and end points of the speech section is reduced. Even within the utterance section, the speech is excluded from some sections where the speech power is high. For example, most parts are buried in noise. For this reason, the difference between the speech segment and the non-speech segment becomes ambiguous, making it difficult to detect the speech segment.

  When a voice utterance section of such a low S / N ratio is subjected to frequency analysis and a logarithmically converted spectrum is observed, the voice spectrum is represented in a frequency band where the voice power is dominant. On the other hand, in the frequency band where the power of other noise is dominant, the spectrum of noise is expressed. That is, even in the voice utterance section, noise may be represented by the band. For this reason, it is difficult to accurately obtain the spectrum average of speech subjected to line distortion in the method of separating the speech section and the noise section collectively for all frequency bands as in the conventional method. On the contrary, when the voice section cannot be detected, the original voice section is also processed as a noise section. For this reason, even the spectral average of noise cannot be obtained accurately.

That is, an object of the present invention is to provide a signal analyzer that can reduce the adverse effects caused by the inability to detect a speech section even in a high noise environment.
Another object of the present invention is to provide a signal analyzer capable of increasing the accuracy of voice recognition even when a voice having no noise section is input or when the noise level gradually changes during utterance. There is to do.

In order to solve the above problems, the signal analysis apparatus of the present invention includes a frequency analysis unit that calculates band energy of a plurality of frequency bands from an input signal at each time, and a surrounding for each band of the plurality of frequency bands. The voice level is calculated for the band having higher energy than the noise of the other, and the level calculation means for calculating the noise level for the other band, and the energy higher than the surrounding noise for normalization of the band energy. Normalization means for calculating the normalized band energy using the calculated voice level for the band having the above, and normalizing means for calculating the normalized band energy using the calculated noise level for the other band. Have.
According to a preferred embodiment of the present invention, there is further provided update coefficient obtaining means for obtaining an update coefficient using a difference between the inputted band energy and the noise level calculated before that time, In the level calculation means, the noise level or the voice level is sequentially updated by controlling at least one adaptation amount of the noise level or the voice level according to the update coefficient .
More preferably, the level calculation means sets the input band energy and the minimum value of the band energy input before that time as a noise level and the maximum value as a voice level.
More preferably, it further comprises normalization coefficient acquisition means for obtaining a normalization coefficient using a difference between the input band energy and the noise level calculated before that time. By having means for determining the subtraction amount of the noise level according to the normalization coefficient, the band energy is adaptively normalized successively .

In here, Normalization refers noise type and size, characteristics of the line, caused by environmental factors, such as input sensitivity to the microphone, to reduce the amount of energy distribution moves the signal input To do. Specifically, normalization is performed by subtracting the average of the energy distribution of the signal to attenuate the DC component, subtracting the value of the environmental noise, or the like. In this specification, the normalization includes controlling how the energy distribution of the input level spreads.

  In the signal analysis apparatus of the present invention, the normalization means may be configured by a low-frequency cutoff filter that attenuates a DC component from the band energy extracted for each band. According to this configuration, since the low frequency component including the DC component of the input signal is attenuated, normalization can be performed with the simplest configuration.

  The noise level is the average of the set of energy distributions that are assumed to represent noise in the input band energy and the energy distribution is relatively low, and the voice level also represents voice. The average of the set whose energy distribution is estimated to be relatively high is shown. In the present specification, the average of energy divided under such conditions as the level of energy, such as the noise level or the voice level, may be referred to as “conditional average”.

  In the signal analysis apparatus of the present invention, when conditions such as noise and speech are adopted, a relatively low energy set and a high energy set in the energy distribution are not divided discretely, but intermediately. It is desirable to handle a continuous input range continuously. That is, it is desirable to determine the likelihood of speech and noise from continuous values based on the magnitude of the band energy for each frame without detecting a speech section.

  The speech recognition apparatus may perform normalization of each band energy in a process of extracting an MFCC (Mel Frequency Cepstrum Coefficients) parameter. According to this configuration, the normalized cepstrum coefficient can be obtained by performing linear conversion from the normalized band energy to the cepstrum parameter.

It may be a signal analysis program for causing the computer to execute the signal analysis device, or a signal analysis program for causing the computer to execute the signal processing device. Moreover, the speech recognition program for making a computer perform the said speech recognition apparatus may be sufficient.
Moreover, the computer-readable recording medium which recorded these programs may be sufficient.

  In addition, the electronic apparatus may include the voice recognition device, and may select and execute a function based on a result of recognition of the voice signal included in the input signal by the voice recognition device. The electronic apparatus of the present invention is not easily affected by distortion due to noise or line characteristics. As a result, it is preferable to use as an electronic device such as a voice recognition type remote controller used at home or a mobile phone.

The signal analyzer of the present invention has the following effects.
(1) In the present invention, the band energy from the input signal at each time obtained for each frequency band is detected without detecting the voice section in the entire frequency band, and the conditional band energy is separately determined for each band. Calculate levels and normalize each band energy. That is, even in the voice utterance section, a band having a dominant noise energy is processed as a noise section, and only a band having a dominant voice energy is processed as a voice section. As a result, the line characteristics of the input signal can be normalized more accurately.

  Even in the signal analysis apparatus of the present invention, when voice and noise are mixed, the determination of voice and noise for each band may be wrong. However, since the voice and the noise are distinguished separately for each band, the influence can be reduced as compared with the case where the voice and the noise are distinguished for the entire band. That is, according to the configuration of the present invention, it is possible to reduce the section detection error between the speech section and the noise section.

  (2) In the signal analyzer of the present invention, when normalizing band energy for each band, normalization processing is performed in the logarithmic spectrum region. When normalization processing is performed in the logarithmic spectrum region, an effect similar to that of the CMN method in which correction is performed using a cepstrum coefficient that is linear transformation (cosine transformation) of a logarithmic spectrum is obtained. As a result, it is possible to normalize the distortion of the line characteristics due to the influence of the microphone or the like. That is, even if the noise spectrum at the time of speech recognition is different from the noise spectrum at the time of learning, it is possible to prevent the recognition accuracy from being greatly deteriorated. Although the subtraction can be performed by the amplitude level as in the SS method, the distortion of the line characteristics cannot be normalized.

  Further, in the signal analyzing apparatus of the present invention, the noise level and the voice level are separately obtained for each band, and the noise level for each band or the voice level for each band is normalized, so that the same effect as the E-CMN method is obtained. Is obtained.

  (3) In the signal analysis device of the present invention, speech and noise are determined from the extracted band energy sequence. That is, even in the utterance section, a part of the band is determined as noise. For this reason, if the utterance is composed of phonemes having different spectrum shapes, the estimation of the noise level in almost all frequency bands is completed within the speech utterance section. That is, in the signal analysis device of the present invention, it is possible to estimate the noise level even when there is no noise section.

  The signal analysis apparatus of the present invention is particularly preferably used for a portable device based on battery driving. In a portable device driven by a battery, an input signal is analyzed only when speaking in order to reduce battery consumption. That is, even in a usage mode in which no noise section exists, the noise spectrum can be estimated equivalently, so that not only the distortion of the line characteristic but also the distortion such as noise can be normalized.

  The best mode for carrying out the present invention will be described below with reference to the drawings. In addition, this invention is not limited by these. The best mode for carrying out the present invention will be described by comparing the MFCC method and the E-CMN method.

[Configuration of MFCC method]
FIG. 2 is a block diagram showing the configuration of a conventional signal analyzer for performing MFCC analysis. In the figure, reference numeral 101 denotes frequency analysis means, and 102 denotes parameter conversion means. The frequency analysis unit 101 performs the frequency analysis step (step S201) in FIG. 1, and the parameter conversion unit 102 performs the parameter conversion step (step S202) in FIG.

[Configuration of E-CMN method]
FIG. 3 is a block diagram showing a configuration of a signal analyzer that performs analysis by the conventional E-CMN method using the MFCC method. In this figure, in addition to the configuration of the signal analysis apparatus that performs the MFCC analysis of FIG. 3, a voice section detection unit 203 that detects a voice section from an input signal, an average update unit 201, and a subtraction processing unit 202 are added. It has a configuration. The input speech is processed by the frequency analysis means 101 and the parameter calculation means 102, whereby the MFCC is obtained. At the same time, the voice segment is detected by the voice segment detector 203 from the input voice. The average updating unit updates the average cepstrum obtained from the parameter calculation unit 102 using the voice section information obtained by the voice section detecting unit 203. Specifically, the average cepstrum of the voice is updated in the case of the voice section, and the average cepstrum of the noise is updated in the case of the non-voice section. The subtraction processing unit 202 subtracts the average cepstrum of speech from the current cepstrum output from the parameter calculation unit 102 if it is a speech segment, using the speech segment information obtained by the speech segment detection unit 203, and performs non-speech If it is an interval, the average cepstrum of noise is subtracted from the current cepstrum output from the parameter calculation means 102.

  As a speech section detection method, generally, a speech section is detected using a short-time signal power for each frame or a spectrum outline for each frame. For these methods, standard methods used for voice calls such as cellular phones are used.

[Configuration of the present invention]
FIG. 4 is a diagram showing the configuration of the signal analysis apparatus of the present invention. In the present invention, an update coefficient acquisition unit 301, a level calculation unit 302, and a normalization unit are provided for each band between the frequency analysis unit 101 and the parameter calculation unit 102 of the signal analysis apparatus used for the MFCC analysis of FIG. 303 and a normalization coefficient acquisition unit 304 are provided. The frequency band analyzing means 101 of the present invention includes a frequency band dividing means 305 that divides an input signal into signals of a plurality of frequency bands, and band energy for each band for the input signal divided into these bands. And band energy extraction means 306 for extracting.

  The update coefficient acquisition unit 301 compares the noise level and voice level calculated before the time when the signal is input with the band energy for each current band obtained by the frequency analysis unit 101, and the level calculation unit 302. To obtain the update coefficient used to update the level.

  The level calculation means 302 uses the update coefficient obtained by the update coefficient acquisition means 301 to update and store the noise level and voice level based on the difference between the input energy and the noise level and voice level. . A specific method will be described later.

  In the normalization coefficient acquisition unit 304, as in the update coefficient acquisition unit 301, the noise level and voice level calculated before the time when the signal is input, and the current band obtained by the frequency analysis unit 101. The normalization coefficient used by the normalization means 303 is calculated.

  The normalization unit 303 uses the normalization coefficient obtained by the normalization coefficient acquisition unit 304 and the voice level and noise level obtained by the level calculation unit 302 to obtain the current band energy obtained by the frequency analysis unit 101. Is normalized and output.

  In this embodiment, the update coefficient acquisition unit 301 and the normalization coefficient acquisition unit 304 have different configurations. However, since both perform similar processing, the update coefficient acquisition unit 301 and the normalization coefficient acquisition unit 304 are The same configuration may be used.

  In the implementation in the literature of CMS method and E-CMN method, each time a word or sentence is uttered, the voice of each time is stored, and the level of the voice level is obtained for normalization or recognition, or the previous utterance It is described to normalize using levels. However, it is not realistic to perform the recognition process after the utterance is completed because the response is slow. Further, in a portable device or the like, there is no guarantee that the previous use environment is the same as the current use environment. Therefore, in the present embodiment, the noise level is updated in real time, not the level for each utterance or the level of the previous utterance, and the band energy is updated in real time using the updated noise level. A mode of normalization will be described. Note that the signal analysis apparatus of the present invention may normalize the band energy using the level for each utterance or the level of the previous utterance.

[Operation]
FIG. 5 is a flowchart showing the flow of signal analysis processing according to the present invention. FIG. 6 is a diagram showing a correspondence relationship between the input band energy obtained by the signal analysis processing according to the present invention, the update coefficient, the normalization coefficient, and the normalized band energy. Hereinafter, the signal analysis processing of the present invention will be described in detail with reference to FIGS. Note that an update coefficient and a normalization coefficient are applied using the fact that noise is low energy and voice is high energy.

  The processing performed in the frequency analysis step (step S201) and the parameter analysis step (step S202) is the same as the processing in FIG.

  The signal analyzer of the present invention operates in principle if it can be divided into two or more bands. In addition to the Mel frequency axis, the frequency axis operates even if it is a park frequency axis or a linear frequency axis. When signal analysis is performed for the purpose of speech recognition, the number of band divisions and frequency scale in frequency analysis may be values compliant with MFCC, and the number of divisions is about 10 to 30. For example, it has been confirmed by experiments that an audio sampled at 11 kHz is effective if it is divided into 24 bands on the mel frequency axis and converted into a 12-dimensional cepstrum. By diverting the frequency analysis step and parameter conversion step processing in the conventional MFCC processing to the signal analysis apparatus of the present invention, it is possible to suppress an increase in processing amount.

  The processing performed in steps S203 to S205 is performed independently for each band. In the coefficient acquisition step (step S203), the update coefficient and the normalization coefficient are calculated based on the difference between the band energy obtained in the frequency analysis step (step S201) and the noise level for each band obtained before the input time of the input signal. Ask for. Next, the noise level of the band is updated based on the update coefficient obtained for each band (step S204). Next, the band energy of the band is normalized using the noise level of the band and the normalization coefficient updated in step S204 (step S205). The normalized band energy is converted into a cepstrum coefficient (step S202).

  In the present embodiment, the cepstrum coefficients generally used for speech recognition are converted. However, it is not always necessary to convert to a cepstrum coefficient, and the normalized energy may be output as it is.

[Update criteria]
Next, a method for updating the noise level in the band energy using the update coefficient will be described with reference to FIG. FIG. 6C is a diagram showing the relationship between the input band energy and the update coefficient.
As a method for estimating the noise level, there is a method in which a histogram of the input energy is calculated and the lower few percent is adopted as the noise level. However, a certain amount of input is required to calculate the histogram. For this reason, it is preferable to perform the process which updates a noise level sequentially in real time demonstrated below.

  If the noise level at time t is N (t), the input band energy is E (t), and the update coefficient is α (t), the noise level N (t) is updated as follows, for example. The unit of noise level and input band energy is considered to be decibel (dB).

N (t) = (1−α (t)) * N (t−1) + α (t) * E (t) (1)

α (t) = 0 (when N (t−1) + R <E (t))

α (t) = A * (1- (E (t) −N (t−1)) / R)
(When N (t−1) <E (t) ≦ N (t−1) + R)

α (t) = A (when E (t) ≦ N (t−1))

  Here, A represents the maximum update coefficient, and is a value of 0 or more and 1 or less. A is set to a value of about 0.02, for example. R indicates a boundary range between noise and voice energy, and is, for example, about 2 dB. Since this update coefficient can determine an average value when particularly low energy is distributed in the time series of input band energy, the noise level can be updated.

In the above formula, (α (t) / A) can be considered as an index indicating the likelihood of noise.
It has been experimentally confirmed that if the initial value N (0) of the noise level is N (0) = E (1), a quick rise characteristic can be obtained and a good recognition result can be obtained.

If the noise level is updated using the above equation, the lowering of the noise will follow relatively fast, and the upward speed will gradually slow down to the upward change of 1 frame. If there is an increase in noise larger than R (dB) in the meantime, no tracking is performed. Such tracking of the change rate of noise can be controlled by the parameters of A and R. When A is 0.02, a noise change slower than about 0.5 Hz follows upward. In general, the change rate of the sound has many components of several Hz to several tens Hz because several to dozens of phonemes are exchanged per second. On the other hand, background noise is often slower than that. However, if the noise energy may suddenly rise due to sudden noise, it will not be possible to follow up after that. For this reason, by setting the lower limit of α (t) to a minute value, for example, about 0.001, instead of 0, it can be set to follow after a few seconds.
In confirming the operation of these parameters, the tracking speed can be confirmed by inputting artificial data in which the energy change speed for each band is adjusted.

  If the voice level at time t is S (t), the input band energy is E (t), and the update coefficient is γ (t), the voice level S (t) is updated as follows, for example.

S (t) = (1−γ (t)) * N (t−1) + γ (t) * E (t) Equation 2

γ (t) = C (when N (t−1) + R <E (t))

γ (t) = C * (E (t) −N (t−1) / R)
(When N (t−1) <E (t) ≦ N (t−1) + R)

γ (t) = 0 (when E (t) ≦ N (t−1))

Here, C represents the maximum update coefficient and is a value of 1 or less. C is set to a value of about 0.02, for example, as in A above. R represents the boundary range between noise and voice energy, and may be the same value as or different from the boundary region at the noise level. Since this update coefficient can determine an average value when particularly high energy is distributed in the time series of input band energy, the voice level can be updated.
It is preferable to use the last S (t) in the previous utterance as the initial value S (0). When the final S (t) cannot be obtained, it has been experimentally confirmed that good results can be obtained by using S (0) = 0. Here, (γ (t) / C) can be considered as an index indicating the likelihood of speech.

  As for the sound level, a fixed value can be used without updating. In this case, it is effective to calculate and use an average sound level from a large amount of sound data.

  In order to update the noise level and the speech level, it is necessary to determine whether it is a noise interval or a speech interval from the time series of the input signal. In Expression 1 and Expression 2, detection of both sections is not obtained in a binary manner, but can take an intermediate value (FIG. 6C). The determination of the noise section and the voice section is performed for each frequency band. For this reason, the noise section and the voice section determined in each band are different from the noise section and the voice section determined in other bands. Furthermore, the noise section and the voice section determined in each band are different from the actual speaker's utterance section.

The update coefficient need not be common in all frequency bands. An update coefficient that is different for each band is stored in advance, so that an optimum update coefficient can be adapted for each band.
For example, when the noise in the low frequency band fluctuates relatively slowly, the risk that the maximum value A of the update coefficient in the low frequency band is reduced and the voice input is erroneously followed and updated. Less.

[Simplified acquisition of noise level and voice level]
As a method for obtaining the noise level, the minimum value of the input energy can be set as the noise level until the input time. This can be considered as a case where the energy boundary range is 0 dB and the maximum update coefficient is 1 in the above formula (1). As for the sound level, the minimum value of the input energy can be set as the sound level until the input time. This utilizes the fact that noise is low energy and voice is high energy.
The method for obtaining the noise level and the voice level is not limited to this example, and the noise level and the voice level are obtained as long as the method can obtain a low value and a high value within the energy distribution range. be able to.

[Normalization criteria]
Next, a method for normalizing band energy using a normalization coefficient will be described with reference to FIG. FIG. 6B is a diagram illustrating the relationship between the input band energy and the normalization coefficient.

When the noise level at time t is N (t), the input band energy is E (t), and the normalization coefficient is β (t), the update of the normalized band energy E ′ (t) is obtained, for example, by the following equation. It is done. The unit of noise level and energy is decibel (dB).

E ′ (t) = E (t) −β (t) * N (t) —Equation 3

β (t) = 0 (when N (t−1) + R <E (t))

β (t) = B * (1- (E (t) −N (t−1)) / R)
(When N (t−1) <E (t) ≦ N (t−1) + R)

β (t) = B (when E (t) ≦ N (t−1))

  Here, B indicates the maximum amount of subtraction and is a value of 1 or less. B is set to a value of about 0.5, for example. R indicates a boundary range between noise and voice energy, and is, for example, about 2 dB. It may be the same value as the boundary region with the update coefficient, or may be a different value.

Next, a method for normalizing band energy using the voice level S (t) in addition to the noise level will be described. Normalization of the band energy using the voice level can be calculated by the following equation, for example.

E ′ (t) = E (t) −β (t) * N (t) −δ (t) * S (t)

δ (t) = D (when N (t−1) + R <E (t))

δ (t) = D * (E (t) −N (t−1)) / R
(When N (t−1) <E (t) ≦ N (t−1) + R)

δ (t) = 0 (when E (t) ≦ N (t−1))

  Here, D indicates the maximum subtraction amount and is a value of 1 or less. C is set to a value of about 0.5, for example, as in B above. R represents the boundary range between noise and voice energy, and may be the same value as or different from the boundary region at the noise level.

  In this example, the sound level normalization coefficient is obtained using the difference between the input energy E (t) and the noise level N (t). By using this method, it is possible to reduce a decrease in recognition accuracy due to characteristics of speakers and lines. Note that the normalization coefficient of the sound level can also be obtained by using the difference between the input energy E (t) and the sound level S (t).

  In this example, in order to normalize the band energy, a method of subtracting the voice level or noise level by multiplying by a coefficient obtained by a predetermined calculation is shown. However, in order to perform normalization, the method is not limited to this method. For example, a normalization effect may be obtained even by a division method such as dividing input energy by a voice level. A normalization method may be appropriately changed according to the input dynamic range and the magnitude of environmental change.

  FIG. 6A is a diagram illustrating the relationship of the normalized band energy with respect to the input band energy using Expression 3 and Expression 4. The case where both the noise level and the voice level are used is shown in the graph of E ′ = E−β * N−δ * S. A case where only noise energy is used when D = 0 is shown in a graph of E ′ = E−β * N. The band energy can be normalized by appropriately changing the normalization coefficient to be applied according to the noise level and the sound level of each band. In addition to this, when the environment to be used is a place where the noise is very stable, it is not necessary to apply noise. Therefore, by setting B = 0, only the voice level can be used.

[When using a low-frequency cutoff filter]
When the width R (dB) of the boundary region between noise and speech is set to a very large value, the noise level is updated with a constant update coefficient A in almost all sections, and the input band energy in almost all sections is updated. Subtraction is performed by a constant normalization factor B. In order to easily realize this, as shown in FIG. 15, a low-frequency cutoff filter may be provided in each band. FIG. 15 is a diagram showing the configuration of the signal analyzing means of the present invention using the low-frequency cutoff filter 307 as the normalizing means. That is, after performing frequency analysis on the input signal, low-frequency cutoff is performed on the energy of each band, and then converted into speech parameters, thereby obtaining a certain degree of normalization effect. In this case, the low-frequency cutoff filter preferably has a frequency lower than about 1 Hz to about 10 Hz, which is the rate of change of the spectrum due to voice, that is, 1 Hz or less.

As an example of the realization of this configuration, t is a frame, x (t) is an input to the low-frequency cutoff filter, and y (t) is an output.

y (t) = x (t) -z (t)
z (t) = x (t) * 0.02 + z (t) * 0.98

When the frame period is 10 ms, the attenuation characteristic is -5.7 dB at 0.5 Hz, -1 dB, 0.24 Hz, approximately -3 dB, 0 Hz. The implementation of this configuration is not limited to the above formula.
In addition, in environments where the frequency of noise changes between the low and high frequency bands, it is possible to improve performance more suited to the usage environment by configuring with low-frequency cutoff filters with different characteristics for each band. be able to.

[Operation of this embodiment from the viewpoint of spectrum]
Next, an embodiment of the present invention will be described from the viewpoint of spectrum based on the drawings.

  FIG. 7 is a diagram illustrating an example of an acoustic signal including a speech waveform. In the figure, the horizontal axis represents time, and the vertical axis represents amplitude. In this figure, the section from time t1 to t2 represents the utterance section, and indicates that the entire displayed time includes noise.

  FIG. 8 is a simplified diagram of the spectrogram of an acoustic signal including the speech waveform shown in FIG. In the figure, the horizontal axis represents time and the vertical axis represents frequency. The section from time t1 to t2 represents the utterance section. The actual spectrum is a continuous value, but in the example of this figure, for the sake of simplicity, a region where the energy is relatively higher than other portions is shown by a closed curve and shaded.

  FIG. 9 is a diagram illustrating a range in which normalization processing is performed when the spectrum illustrated in FIG. 8 is normalized using the conventional E-CMN method. In the figure, the horizontal axis indicates time, and each segment on the horizontal axis indicates an analysis frame. The vertical axis indicates the frequency, and each division on the vertical axis indicates the frequency band. A region where the energy is relatively higher than other portions is closed by a closed curve. The shaded part is the application range as a speech section, and the other part is the application range as a noise section.

  When the E-CMN method is used, when normalizing the cepstrum coefficient, it is not possible to select and normalize the frequency band. For this reason, the speech cepstrum coefficient is updated in the interval from the time t1 to the time t2 determined as the speech interval, and the noise cepstrum coefficient is updated in the other intervals. The updated cepstrum coefficient is used to normalize the cepstrum for each interval. For this reason, when noise is included in the speech section, the cepstrum coefficient is erroneously updated.

  FIG. 10 is a diagram showing a range in which normalization processing is performed when the spectrum shown in FIG. 8 is normalized using the signal analyzer of the present invention. In the figure, the horizontal axis indicates time, and each segment on the horizontal axis indicates an analysis frame. The vertical axis indicates the frequency, and each division on the vertical axis indicates the frequency band. The shaded part is the application range as a speech section, and the other part is the application range as a noise section.

  In the case of the present invention, a speech section and a noise section are determined for each band. As a result, regardless of the voice utterance period (t1 to t2), the voice level of the band is updated and updated in the band and frame (shaded portion in the figure) having higher energy than the surrounding noise. Band energy is normalized using the voice level. In other parts, even in the voice utterance section (t1 to t2), the energy is low and it is determined as the noise section. Therefore, the noise level of the band is updated, and the updated noise level is used. Band energy is normalized.

  That is, according to the present invention, when a speech segment is determined, a clear non-speech segment may not exist as in the prior art. If the speech includes a plurality of types of phonemes, the noise level can be updated in all the bands included in the speech section.

  In FIG. 10, for the sake of simplicity, a shaded portion is provided in the drawing, and the speech section and the noise section are illustrated so that there is a clear boundary. However, as can be seen from Equations 1 to 4, there is no clear boundary between the speech interval and the noise interval. Actually, the processing is performed so that the band energy is gradually switched in the range of the noise level N (t) to N (t) + R. Only when R = 0, a clear boundary as shown in FIG. 10 is formed.

[Example of input without noise interval]
FIG. 11 is a diagram illustrating a situation where adaptation of a noise spectrum by the E-CMN method proceeds when an acoustic signal including the speech waveform illustrated in FIG. 7 is input. FIG. 12 is a diagram illustrating a state in which the adaptation of the noise spectrum by the signal analysis apparatus of the present invention proceeds when an acoustic signal including the speech waveform illustrated in FIG. 7 is input. In the figure, shaded portions indicate bands and frames in which noise is correctly estimated.

  Here, a case where input to the apparatus starts from the same time t1 as the utterance start time is shown as an example of input with no noise section. In a battery-driven portable device, voice input starts with an instruction such as a button. In such a case, it is considered that an input without a noise interval is performed.

  Except for special cases such as an anechoic room, in an actual environment system in which the level of input noise is not known in advance, it is impossible to cut out speech for input speech that does not have a noise section. In this case, whether the first input frame is interpreted as speech or noise depends on the system design.

  In the case of a system that interprets the first input frame as noise, if the E-CMN method is applied, the utterance interval from time t1 to time t2 in FIG. 11 is all recognized as a noise interval. For this reason, an incorrect noise spectrum is estimated. As a result, the first utterance is flattened as noise and cannot be recognized. Then, the original noise level is estimated after t2 when the voice input ends.

  On the other hand, even in the case of a system that interprets the first input frame as speech, if the E-CMN method is applied, the speech interval noise spectrum from time t1 to time t2 in FIG. 11 cannot be estimated. For this reason, a problem arises in the recognition accuracy of the ending in the vicinity of t2.

  In the case of a system that interprets the first input frame as noise, if the signal analyzer of the present invention is used, an error in estimation of the noise spectrum occurs in the dominant band of speech. However, when the spectrum shape changes and the voice power of the band becomes equal to or lower than the noise power, the noise power of the band is estimated. In the example of FIG. 12, the estimation of the noise power ends at time t3 earlier than time t2 when utterance ends. Therefore, the noise spectrum can be correctly normalized between the times t3 and t2. That is, when the signal analyzer of the present invention is used, correct normalization can be performed earlier than when the E-CMN method is used. Also, in a band where noise estimation is completed, if the power of the band increases after the completion time, the voice power is estimated, so that accurate voice recognition can be performed in the middle of utterance.

  As described above, when the signal analysis device of the present invention is used, even a device that instructs the start of voice input with a button or the like can perform highly accurate voice recognition from the first utterance.

  As described above, the signal analysis apparatus of the present invention can estimate the noise spectrum even during voice utterance. As a result, even when the noise spectrum gradually changes during sound generation, normalization can be performed by adapting to the noise change as long as the change proceeds slowly. Therefore, when the signal analysis device of the present invention is applied to a speech recognition device, a speech recognition device capable of performing more stable speech recognition can be obtained.

[Voice recognition device and its recognition method]
FIG. 13 is a block diagram showing an example of a speech recognition system using the speech recognition apparatus of the present invention. The speech recognition system generally includes an acoustic model learning device 401 and a speech recognition device 402.

  The voice database 403 is for learning an acoustic model. It is mainly stored in a personal computer or a fixed disk of a workstation. Reference numeral 404 denotes signal analysis means using the signal analysis apparatus of the present invention. Actually, the configuration shown in FIG. 6 or the configuration shown in FIG. 6 is used with a configuration in which a portion for obtaining a temporal change amount of the acoustic parameter is added. Reference numeral 405 denotes an acoustic model learning unit that statistically obtains statistics for each speech unit such as each phoneme or each syllable from the language database 406 that records the utterance content of the speech database and the output of the signal analysis unit 404. The model is generally a hidden Markov model.

  Reference numeral 407 denotes an acoustic model obtained by the acoustic model learning unit 405. Reference numeral 408 denotes a language dictionary created separately. The language dictionary 408 includes a word dictionary in which words are represented by phoneme strings, and grammar data that defines connection restrictions between words. The creation of the language dictionary 408 may be performed manually, or the connection probability between words may be obtained by statistics from sentences included in the language database 406.

  Reference numeral 409 denotes signal analysis means for performing signal analysis similar to the signal analysis device 404. Reference numeral 410 denotes likelihood calculation means, which obtains the likelihood of each speech unit for the input signal at each time from each statistic of the acoustic model 407 and the acoustic parameters obtained by the signal analysis means 409. Reference numeral 411 denotes a matching unit that calculates likelihoods of likely linguistic hypotheses from the obtained time series of likelihoods of respective speech units, and outputs candidates in descending order of likelihood. As a speech recognition method, there may be an implementation that does not clearly separate likelihood calculation and collation means.

[Programs, electronic devices]
Speech recognition is generally performed by digital processing using a general-purpose or signal processing CPU. FIG. 14 is a diagram illustrating a configuration of a voice recognition device of the present invention and an electronic device including the voice recognition device of the present invention.

  Reference numeral 501 denotes a data and address bus in a digital device such as a personal computer. Each processing means is connected to this bus and performs each processing. Reference numeral 502 denotes a plurality of input means such as buttons, a keyboard, and a microphone. Audio input is not limited to input from a microphone, but may be input via a communication line after being converted into an electrical signal by another device. Reference numeral 503 denotes a CPU that controls the device in accordance with an instruction from the input unit 502 and recognizes the input voice. Reference numeral 504 denotes a work memory for processing by the CPU and a program memory including a voice recognition program. Reference numeral 505 denotes an output device such as a display, a buzzer, a speaker, or a lamp. The result of recognizing the voice may be displayed as a candidate, some processing may be performed in response to the recognition result, and further, the processed result may be displayed. When this electronic device is a mobile phone, wireless communication means (not shown) is added to these processing blocks. In personal computers and portable information devices, communication means and external storage devices are added.

  Examples of selecting and executing a function based on the result recognized by the voice recognition device include, for example, switching a television channel, playing or stopping a video device, setting the temperature of an air conditioner, and the like. In the case of an information terminal, communication control, program execution control, character input, and the like can be given.

A control program for these apparatuses including a signal analysis program or a voice recognition program is realized by an information processing program recorded on a program recording medium. The program recording medium in the above embodiment is a program medium composed of a ROM (Read Only Memory) provided separately from a RAM (Random Accelerator Memory). Alternatively, it may be a program medium that is loaded into an external auxiliary recording device and read out. In any case, the program reading means for reading the information processing program from the program medium may have a configuration for directly accessing and reading the program medium, or a program storage area ( (Not shown) may be downloaded, and the program storage area may be accessed and read. It is assumed that a download program for downloading from the program medium to the program recording area of the RAM is stored in advance in the main unit.
Here, the program medium is configured to be separable from the main body side, such as a tape system such as a magnetic tape or a cassette tape, a magnetic disk such as a floppy disk or a hard disk, a CD (compact disk) -ROM, a MO (magneto-optical). ) Discs, optical discs such as discs, MD (mini discs), DVDs (digital versatile discs), card systems such as IC (integrated circuit) cards and optical cards, mask ROMs, EPROMs (ultraviolet erasable ROMs), EEPROMs ( It is a medium that carries a fixed program, including a semiconductor memory system such as an electrically erasable ROM) and a flash ROM.
In addition, the speech recognition apparatus or electronic device in the above embodiment includes a modem and can be connected to a communication network including the Internet. In this case, the program medium may be a medium that fluidly carries the program by downloading from a communication network or the like. In this case, it is assumed that a download program for downloading from the communication network is stored in the main device in advance. Or it shall be installed from another recording medium.

FIG. 1 is a flowchart showing a procedure for obtaining the MFCC. FIG. 2 is a block diagram showing the configuration of a conventional signal analyzer for performing MFCC analysis. FIG. 3 is a block diagram showing a configuration of a signal analyzer that performs analysis by the conventional E-CMN method using the MFCC method. FIG. 4 is a diagram showing the configuration of the signal analysis apparatus of the present invention. FIG. 5 is a flowchart showing the flow of signal analysis processing according to the present invention. FIG. 6 is a diagram showing a correspondence relationship between the input band energy obtained by the signal analysis processing according to the present invention, the update coefficient, the normalization coefficient, and the normalized band energy. FIG. 6A is a diagram illustrating the relationship of the normalized band energy with respect to the input band energy using Expression 3 and Expression 4. FIG. 6B is a diagram illustrating the relationship between the input band energy and the normalization coefficient. FIG. 6C is a diagram showing the relationship between the input band energy and the update coefficient. FIG. 7 is a diagram illustrating an example of an acoustic signal including a speech waveform. FIG. 8 is a simplified diagram of the spectrogram of an acoustic signal including the speech waveform shown in FIG. FIG. 9 is a diagram illustrating a range in which normalization processing is performed when the spectrum illustrated in FIG. 8 is normalized using the conventional E-CMN method. FIG. 10 is a diagram showing a range in which normalization processing is performed when the spectrum shown in FIG. 8 is normalized using the signal analyzer of the present invention. FIG. 11 is a diagram illustrating a situation where adaptation of a noise spectrum by the E-CMN method proceeds when an acoustic signal including the speech waveform illustrated in FIG. 7 is input. FIG. 12 is a diagram illustrating a state in which the adaptation of the noise spectrum by the signal analysis apparatus of the present invention proceeds when an acoustic signal including the speech waveform illustrated in FIG. 7 is input. FIG. 13 is a block diagram showing an example of a speech recognition system using the speech recognition apparatus of the present invention. FIG. 14 is a diagram illustrating a configuration of a voice recognition device of the present invention and an electronic device including the voice recognition device of the present invention. FIG. 15 is a diagram showing the configuration of the signal analysis means of the present invention using a low-frequency cutoff filter as normalization means.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Frequency analysis means 102 Parameter conversion means 201 Average update means 202 Subtraction processing means 203 Speech section detection means 301 Update coefficient acquisition means 302 Level calculation means 303 Normalization means 304 Normalization coefficient acquisition means 305 Frequency band division means 306 Band energy extraction means 307 Low-frequency cutoff filter 401 Acoustic model learning device 402 Speech recognition device 403 Speech database 404 Signal analysis means 405 Acoustic model learning means 406 Language database 407 Acoustic model 408 Language dictionary 409 Signal analysis means 410 Likelihood calculation means 411 Collation means 501 Bus 502 Input means 503 CPU
504 Program memory 505 Output means

Claims (10)

A frequency analysis means for calculating band energy of a plurality of frequency bands from an input signal at each time;
For each band of the plurality of frequency bands, a level calculation unit that calculates a voice level for a band having energy higher than ambient noise, and calculates a noise level for other bands;
For normalization of band energy, normalization band energy is calculated using the calculated voice level for a band having energy higher than the ambient noise, and the calculation is performed for other bands. Analysis device having normalization means for calculating normalized band energy using the measured noise level . Update coefficient obtaining means for obtaining an update coefficient using a difference between the input band energy and the noise level calculated before that time;
2. The signal analysis according to claim 1 , wherein the level calculation means sequentially updates the noise level or the voice level by controlling at least one adaptation amount of the noise level or the voice level according to the update coefficient. apparatus. In the level calculation means,
2. The signal analyzer according to claim 1 , wherein the input band energy and the minimum value of the band energy input before that time are set as a noise level, and the maximum value is set as a voice level. A normalization coefficient obtaining unit for obtaining a normalization coefficient by using a difference between the input band energy and the noise level calculated before that time;
2. The signal analyzing apparatus according to claim 1 , wherein the normalizing means includes means for determining a subtraction amount of the noise level in accordance with the normalization coefficient so as to sequentially and adaptively normalize band energy . 5. A speech recognition apparatus comprising the signal analysis apparatus according to claim 1, wherein speech included in the input signal is recognized using an acoustic parameter obtained from the input signal by the signal analysis apparatus. A frequency analysis step of calculating band energy of a plurality of frequency bands from an input signal at each time;
For each of the plurality of frequency bands, a level calculation step of calculating a voice level for a band having higher energy than ambient noise and calculating a noise level for the other bands;
For band energy having a higher energy than the ambient noise, normalization band energy is calculated using the calculated voice level, and for the other bands, the calculated noise is calculated. A program for causing a computer to execute a normalization step of calculating normalized band energy using a level . A frequency analysis step for calculating band energy of a plurality of frequency bands from an input signal at each time;
For each of the plurality of frequency bands, a level calculation step of calculating a voice level for a band having higher energy than ambient noise and calculating a noise level for the other bands;
For band energy having a higher energy than the ambient noise, normalization band energy is calculated using the calculated voice level, and for the other bands, the calculated noise is calculated. A normalization step of calculating a normalized band energy using the level;
A program for causing a computer to execute a step of obtaining an acoustic parameter based on the calculated normalized band energy and recognizing a voice included in the input signal .   A computer-readable recording medium on which the program according to claim 6 is recorded.   A computer-readable recording medium on which the program according to claim 7 is recorded.   An electronic apparatus comprising the voice recognition device according to claim 5, wherein a function is selected and executed using a result of recognition of voice included in an input signal by the voice recognition device.

Images