JP3493849B2 - Voice recognition device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of urging a user to speak loudly in a voice recognition device according to the magnitude of background noise. Also,
The present invention relates to the field of speech recognition. The technology of voice recognition is used when a device is operated by a voice, when a voice is used as an input means in place of a keyboard or the like, and as a means for an interface between a robot or another system and a human.

[0002]

2. Description of the Related Art As a voice recognition technique, a pattern matching technique called DP matching or an HMM (Hid
The stochastic method called den Markov Model) has been widely used so far. Here, in DP matching, a standard pattern corresponding to a candidate word to be recognized is created in advance, and a feature amount pattern obtained by analyzing an input voice signal and all standard patterns are associated with each other on the time axis. While matching (calculating the distance), the most similar one is selected.

In the HMM method, a probabilistic model corresponding to a candidate word to be recognized is created in advance by using a Markov model having some states and symbol output probabilities at the time of transition of the states, and the input is input. This is a method in which the probability of occurrence of a feature amount sequence obtained by analyzing a voice signal is obtained from each model and the one with the highest occurrence probability is selected. In continuous speech recognition, the word sequence corresponding to the input signal is determined using grammar and semantic connections of words in addition to the DP matching and HMM methods.

[0004]

Now, such speech recognition may be used in various environments, and its background noise is often a factor of deteriorating recognition performance. In particular, when background noise is superimposed on the input speech, the feature amount obtained by acoustic analysis of the input signal changes significantly, and it becomes far from the standard pattern or probabilistic model, which causes a problem that it cannot be recognized. To do.

In order to avoid this, a noise removal method called spectrum subtraction and an adaptive processing method using a plurality of microphones have been proposed so far. These are intended to remove the noise contained in the input signal, but it is a very difficult problem to remove the noise from the speech signal already affected by the noise.

An object of the present invention is to observe the magnitude of background noise and urge the user to speak loudly according to the magnitude of the background noise, that is, urge the user to speak loudly. The purpose of this is to reduce the influence of background noise itself, and as a result, prevent deterioration of recognition performance.

[0007]

A speech recognition apparatus according to the present invention comprises a speech signal input section, an acoustic analysis section for analyzing the input signal and extracting a feature amount, and learning data in advance in the learning process. Based on the feature amount obtained by the analysis, the parameter to be used for recognition is obtained, and the parameter storage unit for storing this parameter and the feature amount obtained by analyzing the input signal are stored. By scoring based on the distance or the occurrence probability from the set parameters, the recognition unit that determines the word or the sequence of words corresponding to the input signal, the energy of the input voice signal is calculated, and the input voice signal is calculated. the energy of the signal is, of background noise d
A determination unit that determines whether the energy of the input voice signal is less than the threshold value, which is determined based on the average value of energy and the standard deviation, and when the energy of the input voice signal is determined to be less than the threshold value, a loud voice is input. It is characterized in that it comprises a warning unit for warning the user as described above. Further , the determination unit of the voice recognition device according to the present invention, when obtaining the average value and the standard deviation of the energy of the background noise, the input signal determined as the non-voice portion based on the discrimination result of voiced / unvoiced. make use of,
The feature is that it is updated sequentially. further,
The determination unit of the voice recognition device according to the present invention, as a condition for updating the average spectrum of the unvoiced portion, the average value of the energy and the standard deviation, etc. If it continues to be judged as voice,
It is characterized by forcibly updating. Further, the determination unit of the voice recognition device according to the present invention uses the energy of the input signal in order to discriminate between voiced and non-voiced voices, and if there is a predetermined threshold value or more, it is voiced, and otherwise it is voiceless. Is characterized by.

[0008]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described in detail below with reference to the drawings. First, the speech recognition will be briefly described, and then the system configuration and the operation of the present invention will be described in detail.

(A) Speech recognition Speech recognition is often composed of an input unit 11, an acoustic analysis unit 12, a recognition unit 13, a parameter storage unit 14, and an output unit 15 as shown in FIG. The input unit 1-1 is configured by a device such as a microphone that inputs a sound signal, an amplifier that amplifies the input signal, and an AD converter that converts the input signal into a digital signal. Then, after sampling the input signal (for example, at 12 kHz), the acoustic analysis unit 1
Send to 2.

The acoustic analysis unit 12 extracts a feature amount necessary for recognition from the input voice signal. For example, simple signal energy, number of zero crossings, pitch, etc. are extracted, and frequency analysis is performed by linear predictive analysis (LPC), fast Fourier transform (FFT), bandpass filter (BPF), and wavelet transform. Do it. Then, for example, the feature quantity is extracted as a vector time series having elements such as band-divided energy. Further, as the amount of change in the feature amount, for example, difference data may be simultaneously extracted as one of the feature amounts. KL conversion, or
In some cases, a suitable mapping such as a neural network may be applied to further convert to a feature amount having a high degree of separation. Further, the feature quantity vector may be compressed by vector quantization or the like and converted into a quantized feature quantity.
In any case, the acoustic analysis unit 12 extracts a feature amount necessary for recognition from the input voice signal, and the extracted feature amount is used by the recognition unit 1
Send to 3.

The recognition unit 13 performs recognition processing for unknown voice data by using the parameter 14 created in advance based on the feature amount obtained by acoustically analyzing the voice data for learning. Here, the recognition is to select a word corresponding to the input from the given recognition dictionary for the input voice signal. As a recognition method for that purpose, a method using DP matching, a neural network, an HMM, etc. is mainly used.

In DP matching, a standard pattern called a template is obtained in advance as a parameter from the feature amount obtained by analyzing each voice signal, and the one that is judged to be the closest to the feature amount of the unknown voice is found. Is the method. In order to absorb the fluctuation of speech rate,
A method called dynamic time warping is often used to expand or contract the time axis so as to minimize distortion with the template.

The neural network is intended to be recognized by a network model that mimics the structure of the human brain, and the weighting coefficient of the path is determined in advance as a parameter in the learning process, and the characteristics of unknown speech are added to the network. Based on the output obtained by inputting the quantity, the distance to each word in the dictionary is calculated and the recognition word is decided. In addition, the HMM attempts to perform recognition using a probabilistic model. For the state transition model, the transition probability and the output symbol probability are previously determined based on the learning data, and the feature amount of the unknown speech is calculated. In this method, the recognition word is decided based on the occurrence probability of each model.

As described above, generally, in the recognition process, the parameters (templates, weighting factors of the network model, statistical parameters of the probabilistic model, etc.) determined in advance from the learning data are used as the learning process.
Is calculated and stored in the parameter storage unit 14. Then, in the recognition process, after the acoustic analysis 12 of the input unknown voice signal, the parameter 14 obtained by learning is analyzed.
Is used to score each word in the given dictionary such as distance and occurrence probability according to its recognition method, and the one with the highest score (or
The top two or more) are selected as recognition results. Then, the obtained recognition result is transmitted to the output unit 15. The output unit 15 issues a command such as displaying the transmitted recognition result on the screen, outputting as a sound, and using the recognition result to operate another device.

The above is a brief description of word speech recognition, but in the case of continuous speech recognition, the parameter storage unit 14 also stores the grammar of the language, the semantic connection of words, and the like. The sequence of consecutive words corresponding to the input voice signal is recognized by the recognition unit 13 while using these restrictions.
The recognition result is transmitted to the output unit 15.

(B) System Configuration In the speech recognition processing as described above, what is output from the acoustic analysis unit 12 is a characteristic obtained by analyzing the input speech signal at every predetermined minute time interval. It is a series of quantities. This feature series is used both during learning and during recognition. The problem here is that the feature amount extracted by the acoustic analysis unit 12 varies greatly due to the influence of background noise. Therefore, a voice recognition device having a configuration as shown in FIG. 2 is realized. The input unit 21, the acoustic analysis unit 22, the recognition unit 23, the parameter storage unit 24, and the output unit 25 are the same as those in FIG. The background noise observation unit 26 and the warning unit 27 are incorporated in this.

The background noise observing section 26 calculates the average value μ and the standard deviation σ of the background noise energy. This calculation method will be described later. Then, the functional expression (1) for determining the threshold value r is defined in advance from the μ and σ.

[0018]

[Equation 1] Then, it is determined whether the energy e of the input signal satisfies the expression (2).

[0019]

[Equation 2] The determination result is sent to the warning unit 27. For example, as in the equation (3), a value obtained by adding the standard energy σ to the average energy μ of the background noise times c is set as the threshold r.

[0020]

[Equation 3] However, as in equation (4), the upper limit value r _max and the lower limit value r
Set _min .

[Equation 4] Then, it is determined whether or not the voice is input so as to satisfy the expression (2), and the result is transmitted to the warning unit 27.

In the warning unit 27, based on this judgment result,
The user is warned so that the energy e of the input signal satisfies the expression (2). For example, the LED is installed so that the user can understand, and only when the formula (2) is satisfied, LE is set.
D is turned on, or a display 31 as shown in FIG. 3 is prepared, and the level meter 33 is displayed so as to change according to the size of e.
A warning line 32 is displayed in accordance with. That is, the user is required to speak so that the LED lights up or the level meter 33 crosses the warning line 32.

Alternatively, for a utterance that does not satisfy the determination formula (2), a warning sound may be emitted to the user, or a warning message may be displayed or responded. For example, "Please speak a little louder. 』And so on to send a message to the user.

(C) Background Noise Observing Section FIG. 4 is a detailed description of the operation of the background noise observing section 26. First, the input voice signal is frequency-analyzed (4-1) at every predetermined minute time interval, and a feature quantity vector (hereinafter, referred to as an input spectrum) having band-divided energy as an element is obtained. For example, by using a bandpass filter or calculating the amplitude characteristic by FFT, the feature quantity vector having the energy of each band as an element can be obtained at constant time intervals. This is represented as Xt. Xt is an n-dimensional vector observed at time t, as shown in equation (5).

[0024]

[Equation 5] Further, since each element of X _t represents energy, Expression (6) is satisfied.

[Equation 6] Here, T represents transposition. The frequency analysis may use the analysis result of the acoustic analysis unit 22. Then calculate the energy of the input signal according to equation (7),
It is determined whether the expression (2) is satisfied (4-2).

[0025]

[Equation 7] The result is transmitted to the warning unit 27. Then, the average spectrum V of the background noise, and the average value μ of energy and the standard deviation σ are updated (43). Here, V is represented as an n-dimensional vector expression (8).

[0026]

[Equation 8] The method of updating these parameters will be described later in detail. Then, finally, the threshold value r is calculated by the equation (2) according to μ and σ (4-
4). At the time t, the threshold value r used for the determination of the equation (2) is the one obtained at the time t-1.

(D) Method of Updating Background Noise Next, the method of obtaining the average spectrum V of background noise, the average value μ of energy, and the standard deviation σ will be described. Basically, the input signal is discriminated between voiced and non-voiced, and only when it is determined that there is no voiced portion, the update is performed according to the equation (9).

[0028]

[Equation 9] Here, V on the right side is a value before updating, and V on the left side is a value after updating. Further, α ₁ and β ₁ are weighting factors, and
0).

[Equation 10] For example, α ₁ = 0.95, β ₁ = 0.05, etc. are set. It is also possible to reflect new X _t in V more temporally by increasing β. Using the updated V, the energy average value of the background noise is expressed by equation (11).

[0029]

[Equation 11] Also, if the variance is updated as in equation (12), the standard deviation
You get a sigma.

[Equation 12] Here, sigma ² on the right side value before updating, the left side of the sigma ² is the value after the update, e, mu, respectively (7), and that determined by (11). Further, α ₂ and β ₂ are weighting factors and are set so as to satisfy the equation (13).

[0030]

[Equation 13] In addition to the updating method described above, it is also possible to always store m input spectra X _t at the time determined to be updated, starting from the newest time, and calculate the average value as a new V. . Of course, the energy average value μ and the standard deviation σ can also be easily obtained.

By the way, with the updating method as described above, for example, although no voice is input,
There is a risk that the average spectrum of the background noise may remain unupdated if the voice is mistakenly determined to be voiced because the environment has changed.

Therefore, it can be said that it is necessary to make the average spectrum V of the background noise follow the environment in response to the environmental changes. One way to do this is to estimate the maximum length of the input voice and forcibly update the vector V if the time to be judged as a voiced part is too long. Can be considered. That is, the time d that is continuously determined to be voice is counted,
If this exceeds a certain threshold value dmax, and if the equation (14) is satisfied, V, μ, σ may be forcibly updated even if it is determined that there is voice. .

[0033]

[Equation 14] The updating method at this time is as described above. FIG. 5 is a summary of the background noise updating method described above. To explain again, first, based on the judgment 51 of voiced / unvoiced, if it is judged that there is no voiced, V is updated (5-4). In other cases, if it is determined by the determination formula (14) that the length of the continuous voiced portion is long 5-2, V is updated (5-
4). By updating V in this way, a newer average spectrum V of background noise, and a new average value μ and standard deviation σ of energy can be obtained while following changes in the environment to some extent.

(E) Discrimination Method of Voiced Part and Non-voiced Part Now, a voiced / unvoiced identification method, which is important in updating the background noise spectrum and energy, will be briefly described. Basically, attention is paid to the difference between the frequency characteristics (spectrum) of background noise and voice, and the presence or absence of voice is discriminated according to whether it is far or near from the average spectrum of background noise. First, the difference vector V−X _t is obtained from the input spectrum X _t and the average spectrum V of the background noise. And
If the magnitude ε of the vector V−X _t is large, it is determined that there is voice, and if it is small, it is determined that there is no voice. That is, with respect to a certain threshold value r ₁ , it is determined that the voice is present if the equation (15) is satisfied, and the voice is not provided if the equation (15) is not satisfied.

[0035]

[Equation 15]

Here, as ε, equations (16), (1
7) etc. can be used.

[Equation 16]

[Equation 17] Also, Mahalanobis that takes dispersion into account
Distance or the like may be used. Threshold value r of judgment formula (15)
_It is desirable to set 1 to a value that seems to be appropriate through experiments. Alternatively, r ₁ may be varied according to the average value μ of the background noise energy and the standard deviation σ like r used in the determination formula (2). In any case, if X _t is far away from V, voice, X
_{If t} is in the vicinity of V, it may be determined that there is no voice.

As another method for distinguishing between voiced and unvoiced voices, a method of simply determining the voiced voice portion and the voiceless voice portion according to the energy of the input signal can be considered. Specifically, the energy e of the input signal is compared with a certain threshold value r ₂ , and it is determined that the voice is present if the equation (18) is satisfied, and is not voiced if the equation (18) is not satisfied.

[0038]

[Equation 18] If is established, it is judged that there is voice, and if not, it is judged that there is no voice. Also in this case, it is desirable to set the threshold value r ₂ to a value that is considered appropriate through experiments and the like. Alternatively, similarly to the threshold value r, r ₂ may be changed according to the average value μ and the standard deviation σ of the background noise energy. As yet another method, the number of times z that adjacent samples of the input signal take different positive and negative sign values per unit time
(This is called the zero-crossing velocity) and a certain threshold r
_If Expression (19) is satisfied as compared with ₃ , it is possible to determine that there is voice and the others are silent.

[0039]

[Formula 19] In addition, it is also possible to identify by combining the above determination expressions (15), (18), and (19). Of course, it is also possible to use a voiced / unvoiced determination method other than the above.

[0040]

As described above, the background noise is observed,
Depending on the average value and standard deviation of the background noise energy , if the user is warned to speak loudly, the user will be vocalized accordingly, and as a result, the influence of background noise will be reduced. Therefore, it is possible to prevent the deterioration of the recognition performance. It should be noted that the present invention is aimed at urging the user to utter a loud voice, and is not intended to extract a voiced section from a voiced / unvoiced discrimination result. This is very different from the voice section detection that is performed.

[Brief description of drawings]

FIG. 1 is a block diagram showing a general configuration of a voice recognition device.

FIG. 2 is a block diagram showing a configuration of a voice recognition device of the present invention.

FIG. 3 is a diagram showing a display example in a warning unit.

FIG. 4 is a diagram for explaining an operation of a background noise observation unit.

FIG. 5 is a diagram for explaining a background noise updating method.

[Explanation of symbols]

21 input unit, 22 acoustic analysis unit, 23 recognition unit,
24 parameter storage unit, 25 output unit, 26 background noise observation unit, 27 warning unit

─────────────────────────────────────────────────── ─── Continuation of the front page (72) Hiroshi Tsunoda, 6-735 Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Corporation (72) Hiroaki Ogawa 6-35, Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Corporation (72) Inventor List Masanori 6-735 Kitashinagawa, Shinagawa-ku, Tokyo Sony Corporation (72) Inventor Honda, etc. 6-735 Kitashinagawa, Shinagawa-ku, Tokyo Incorporated (72) Inventor Satoshi Fujimura Sat. 6-735 Kitashinagawa, Shinagawa-ku, Tokyo Sony Corporation (56) References JP-A-63-223795 (JP, A) JP-A-7-64595 ( JP, A) JP 58-85498 (JP, A) JP 1-502858 (JP, A) JP 1-302298 (JP, A) JP 6-75588 (JP, A) JP Flat 8-23756 (JP, B2) Patent 2589468 (JP, B ) Sieve, digital sound processing, Japan, Tokai University Press, September 25, 1985, p. 153, the first 10 to 16 rows (58) investigated the field (Int.Cl. ^7, DB name) G10L 11 / 00-11/02 G10L 15/00-15/28

Claims (7)

(57) [Claims] 1. An input unit for a voice signal, an acoustic analysis unit for analyzing the input signal to extract a feature amount, and a learning process,
Based on the feature amount obtained by analyzing the learning data in advance, the parameter used for recognition is obtained, and the parameter storage unit for storing this parameter and the input signal are obtained. A speech composed of a recognition unit that determines a word corresponding to an input signal or a sequence of words by performing scoring based on the distance or the occurrence probability from the stored characteristic amount and the stored parameters. In the recognition device, the energy of the input voice signal is calculated, and the energy of the input voice signal is the average value of the energy of background noise.
A determination unit that determines whether the energy of the input voice signal is less than the threshold value determined according to the standard deviation, and if the user determines that the energy of the input voice signal is less than the threshold value, the user inputs a loud voice. A voice recognition device comprising: a warning unit for warning the user. 2. The determination unit uses an input signal determined to be a non-voice part based on a discrimination result of voiced / unvoiced when obtaining an average value and a standard deviation of energy of the background noise. , claim 1, characterized in that is renewed sequentially
The voice recognition device described in. 3. The determination unit uses, as a condition for updating an average spectrum of a voiceless portion, an average value of energy and a standard deviation, voiced voice for a certain period of time or more based on a voiced / unvoiced discrimination result. The voice recognition device according to claim 2 , wherein the voice recognition device is forcibly updated when the determination is continued. 4. The determination unit uses the energy of an input signal to distinguish between voiced and non-voiced voices, and determines voiced voices if the energy is larger than a predetermined threshold value and no voices otherwise. The voice recognition device according to claim 2 . 5. The determination unit uses the zero-crossing speed of the input signal to distinguish between voiced and non-voiced voices, and determines that the voiced voice is present when the input signal is larger than a predetermined threshold value, and the voice is not voiced otherwise. The voice recognition device according to claim 2 , characterized in that 6. The determination unit uses a spectrum obtained by frequency-analyzing an input signal in order to discriminate between voiced and unvoiced speech, and a difference from an average spectrum of a non-voiced portion exceeds a predetermined threshold value. The voice recognition device according to claim 2 , wherein the voice recognition device has a voice if the voice becomes louder and a voice is not voiced otherwise. 7. The determination unit distinguishes between the energy of an input signal, a zero-crossing speed, and a spectrum obtained by frequency-analyzing the input signal and an average spectrum of a non-voiced portion in order to distinguish between voiced and unvoiced speech. The voice recognition device according to claim 2 , wherein the voice recognition and the non-voice recognition are performed by combining three feature amounts of difference.