JP2013122508A - Voice recognition device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition device which generates a discriminative feature quantity conversion optimized by voice recognition by using a voice signal for learning and correct answer class information of the voice signal for learning, and suppresses an influence of acoustic noise by applying the feature quantity conversion to a recognition voice signal, and by extracting and using a sound feature quantity.SOLUTION: A learning unit 10 includes: a class identification conversion construction unit 11 which calculates a class identification conversion from correct answer class information of vector series and the voice signal for learning by using the vector series to which a voice input unit 1 converts the voice signal for learning and correct answer class information of the voice signal for learning as inputs; and a dimensional compression and linear orthogonal transform construction unit 12 which calculates a dimensional compression and linear orthogonal transform from the vector series and the correct answer class information of the voice signal for learning or the class identification conversion. An application unit 20 includes: a feature quantity extraction unit 21 which extracts a sound feature quantity from the vector series and the class identification conversion or the dimensional compression and linear orthogonal transform by using the vector series to which the voice input unit 1 converts a recognition voice signal; and a voice recognition unit 2 which performs voice recognition by using the sound feature quantity and outputs a recognition result.

Description

  The present invention relates to a speech recognition apparatus that converts an acoustic feature amount into text and a program thereof.

  Speech recognition is a technique for extracting an acoustic feature from an input speech signal and converting it into text by pattern matching or a hidden Markov model. Among them, the acoustic feature extraction is a technique for converting an input voice signal into a time-series vector, that is, a feature using signal processing / signal analysis. In practical speech recognition, an acoustic feature called Mel-Frequency Cepstrum Coefficient extracted based on acoustic analysis is often used.

  An existing problem of speech recognition is that the recognition performance under noise such as in the real environment is low. As summarized in Non-Patent Document 1, the techniques often used to improve the recognition performance include spectral subtraction and cepstrum average normalization as hidden Markov models as techniques for reducing noise during the acoustic feature extraction process. As a method for adapting to noise, there are a posteriori probability maximization method and a maximum likelihood linear regression method.

  In order to improve the recognition performance, there is a method of using acoustic feature amount extraction by a discriminative method instead of a method mainly using acoustic analysis and signal processing. As an example, in Non-Patent Document 2 and Non-Patent Document 3, the extraction of the acoustic feature amount from the speech signal is obtained by optimization based on the maximum mutual information criterion and phoneme error minimization criterion.

  Genetic algorithms, on the other hand, represent optimal solution candidates as genes, generate multiple individuals with genes, and search for the optimal solution by performing operations such as selection, crossover, and mutation according to fitness. It is. In Non-Patent Document 4, the extraction method of the Mel scale cepstrum coefficient is improved by a genetic algorithm using the output likelihood of the hidden Markov model.

  In addition to the prior case of using a genetic algorithm for speech processing, Patent Document 1 discloses a method of speech analysis, in which an acoustic tube cross-sectional area required for generation is generated for the purpose of mechanically and electrically generating speech. Control parameters such as current values are obtained using a genetic algorithm.

Japanese Patent Application Laid-Open No. 07-114397

“Patent Application Technology Trend Survey on Speech Recognition Technology,” JPO, 2003 D.Povey et al. "FMPE: Discriminatively trained features for speech recognition" International Conference ICASSP2005, pages 961-964, 2005 D.Povey et al. “Boosted MMI for model and feature-space discriminative training” International Conference ICASSP2008, 4057-4060, 2008 Z.Behzad et al. `` Discriminative transformation for speech features based on genetic algorithm and HMM likelihoods '' IEICE Electronics Express, Vol. 7, No. 4, pp. 247-253, 2010

  Conventional speech recognition has a problem in that the recognition performance is significantly deteriorated in a noisy environment. One of the factors is that the mel scale cepstrum coefficient extraction process, which is an acoustic feature quantity, is a technique based on signal processing, and is not necessarily the optimum feature quantity extraction method for speech recognition. Various attempts have been made to improve the mel scale cepstrum coefficient extraction method for this problem, but the problem is that the speech recognition performance obtained is still insufficient.

  An object of the present invention is to improve the performance of speech recognition under noise by extracting discriminative acoustic features that are optimized for speech recognition and that take noise environment into consideration, which is different from the Mel scale cepstrum coefficient. is there.

  As another object of the present invention, a computer can be optimized for speech recognition, extract a discriminative acoustic feature amount in consideration of a noise environment, and can be used as a speech recognition device using the acoustic feature amount. To provide a program.

  The present invention is characterized by providing a speech recognition apparatus that includes extracting an acoustic feature amount optimized for speech recognition in speech recognition.

  In order to achieve the above object, the invention according to claim 1 is characterized in that a voice input unit that inputs a voice signal of a speaker, converts the voice signal into a digital signal, converts the digital signal into a vector sequence, and outputs the vector signal; A class identification conversion construction unit that inputs a vector sequence of learning speech signals output by the unit and correct class information of the learning speech signals and outputs class identification conversion; and the learning speech signal output by the speech input unit A dimensional compression / linear orthogonal transformation construction unit that inputs a vector sequence of the above, correct class information of the learning speech signal and the class identification transformation and outputs dimensional compression / linear orthogonal transformation, and recognition output by the speech input unit A feature quantity extraction unit for inputting a vector sequence of the speech signal, the class identification transform, and the dimension compression / linear orthogonal transform to extract an acoustic feature, and the recognized speech signal as the feature Based on the acoustic features out portion is generated in which the gist of the speech recognition apparatus characterized by comprising a voice recognition unit for performing voice recognition.

  According to a second aspect of the present invention, there is provided a speech input unit that inputs a speech signal of a speaker, converts the speech signal into a digital signal, converts the digital signal into a vector series, and outputs the speech signal, and learning speech output from the speech input unit A feature amount conversion construction unit that inputs a vector sequence of signals and correct class information of a learning speech signal and outputs a feature amount conversion by a genetic algorithm; a vector sequence of recognized speech signals output by the speech input unit; A feature amount extraction unit that inputs a feature amount conversion output by the feature amount conversion construction unit and extracts an acoustic feature amount; and speech recognition based on the acoustic feature amount generated by the feature amount extraction unit. The gist of the present invention is a speech recognition device including a speech recognition unit for performing the speech recognition.

  The invention according to claim 3 is a program for causing a computer to perform voice recognition, and is characterized by realizing the voice recognition device according to claim 1 or claim 2 by a computer. It is.

  According to the first aspect of the present invention, it is possible to extract a discriminative acoustic feature quantity by performing feature quantity conversion optimized for speech recognition by acoustic feature quantity extraction having a learning mechanism. In the first aspect of the present invention, first, a class identification transformation is generated that returns an intermediate vector composed of information about whether or not a vector extracted from a vector sequence belongs to a class to be identified (phoneme or the like). Secondly, a dimensional compression / linear orthogonal transformation for performing dimensional compression and linear orthogonalization on the intermediate vector obtained by the class identification transformation is generated. As described above, according to the first aspect of the present invention, it is possible to extract a feature amount having high recognition performance.

  According to the second aspect of the present invention, it is possible to extract a discriminative acoustic feature quantity by performing feature quantity conversion optimized for speech recognition by acoustic feature quantity extraction having a learning mechanism. According to the second aspect of the present invention, the conversion from the vector sequence to the acoustic feature amount is generated using a genetic algorithm. That is, according to the second aspect of the present invention, it is possible to extract a feature quantity with high recognition performance.

  According to the invention of claim 3, by executing the program, the computer can be easily realized as the speech recognition apparatus according to claim 1 or claim 2.

  As described above, the speech recognition apparatus according to the present invention extracts acoustic feature amounts that are optimized for speech recognition and are robust even under noisy environments by means of acoustic feature amount extraction having a learning mechanism. Realizes high voice recognition performance even under.

The functional block diagram of one Embodiment of the speech recognition apparatus of Claim 1. The functional block diagram of one Embodiment of the speech recognition apparatus of Claim 2. Explanatory drawing of the conversion from the digitized audio | voice signal to a frequency vector series. The performance comparison figure according to noise kind of one Embodiment of the speech recognition apparatus of Claim 1 and the conventional speech recognition apparatus. The performance comparison figure according to noise intensity with one Embodiment of the speech recognition apparatus of Claim 1, and the conventional speech recognition apparatus.

  Hereinafter, an embodiment of a speech recognition apparatus embodying the present invention will be described with reference to FIGS.

FIG. 1 is an embodiment of the invention described in claim 1.
The speech input unit 1 inputs the speech of the speaker, samples the input speech signal by the sampling theorem, quantizes it at an appropriate quantization step, and converts it into a digital signal. Subsequently, as shown in FIG. 3, the digital signal is cut out at a certain time length every certain time to extract a voice frame. For each voice frame, frequency information is obtained by using Fourier transform, and the frequency information is collected and vectorized by a mel filter bank to be converted into a frequency vector. A vector sequence (hereinafter referred to as a frequency vector sequence) is generated by arranging the frequency vectors in time series with the entire digitized signal.

  The learning unit 10 uses the frequency vector sequence obtained by inputting the learning speech signal to the speech input unit 1 and the correct class information of the learning speech signal to calculate class identification transformation and dimensional compression / linear orthogonal transformation. The application unit 20 converts the frequency vector sequence obtained by inputting the recognized speech signal to the speech input unit 1 using the class identification transformation and dimensional compression / linear orthogonal transformation obtained by the learning unit 10 into an acoustic feature amount, A recognition result is output by performing voice recognition.

  First, the class identification conversion construction unit 11 configures a classifier that determines whether or not a frequency vector belongs to the class for each class. As an example, “a discriminator for determining whether or not a given frequency vector is for phoneme a” and “a discriminator for determining whether or not a given frequency vector is for phoneme b” are constructed.

A configuration example of a classifier using a genetic algorithm in the class identification conversion construction unit 11 will be described below.
A classifier f indicating whether or not the N-dimensional frequency vector x belongs to c for class c is expressed by the following equation. The discriminator returns a positive value if the frequency vector x belongs to c, and returns a negative value if it does not belong.

  The discriminator parameter a (c) of the discriminator f is determined using the learning speech signal and correct class information of the learning speech signal. That is, a (c) is determined so as to return a correct identification result for as many vectors as possible in the frequency vector sequence obtained by inputting the learning speech signal to the speech input unit 1.

  Specifically, a method for determining the discriminator parameter a (c) by a genetic algorithm will be described. A plurality of individuals using the discriminator parameter a (c) having N + 1 coefficients as genes are generated to constitute the initial generation. Each individual is initialized randomly.

  The fitness E (v) is calculated with respect to the individual v of the current generation by the following formula. Where a is a discriminator parameter obtained from the gene of the individual v, rk is a kth frequency vector of a frequency vector sequence obtained by inputting a learning speech signal to the speech input unit 1, and lk is a kth frequency vector. The correct class information is 1 when rk belongs to class c and -1 when it does not belong. K is the total number of frequency vectors obtained by inputting the learning speech signal to the speech input unit 1. Sgn is a sign function.

  Based on the fitness, the basic operation of the genetic algorithm (elite selection / inheritance / crossover / mutation) is performed to create the next generation from the current generation (hereinafter, the generation of the next generation from the current generation is referred to as generation change). ).

  The generation change is repeated a certain number of times from the initial generation to generate the final generation. An individual with the highest fitness is taken out from the final generation, and the gene of the individual is analyzed to obtain an optimized discriminator parameter a (c).

  Note that the present invention does not prevent the use of a support vector machine or the like for the determination of the discriminator parameter a (c). Although a linear classifier is shown as an example of the classifier, the present invention does not prevent the use of a nonlinear classifier.

  Second, the class identification conversion construction unit 11 configures the classifier for all classes. When class c discriminator parameter a (c) is expressed as a row vector, all the row vectors are put together to form a class discrimination conversion matrix. As an output of the class identification conversion construction unit 11, a class identification conversion matrix of C rows and N + 1 columns is obtained. Where C is the number of identification classes.

  The dimension compression / linear orthogonal transformation construction unit 12 performs dimension compression / linear orthogonalization of the class identification transformation obtained by the class identification transformation construction unit 11.

A configuration example of dimensional compression / linear orthogonalization using a genetic algorithm in the dimensional compression / linear orthogonal transformation construction unit 12 will be described below.
In the frequency vector sequence obtained by inputting the learning speech signal to the speech input unit 1, each frequency vector of the frequency vector sequence is multiplied by the class identification transformation matrix A by the extended frequency vector x ′ as shown in the following equation. A C-dimensional vector y (hereinafter referred to as an intermediate vector) is obtained.

  The calculation of the first dimension g1 of dimensional compression / linear orthogonal transformation is performed on the intermediate vector y by the following equation.

  The parameter b (1) of g1 is determined using the learning speech signal and the correct class information of the learning speech signal. That is, in the frequency vector sequence obtained by inputting the learning speech signal to the speech input unit 1, the average vector of the intermediate vector y is calculated for each class, the value of the average vector g1 is obtained, and the variance is maximum. B (1) is determined so that

  Specifically, a method for determining the parameter b (1) by a genetic algorithm will be described. A plurality of individuals having genes of the parameter b (1) having C coefficients are generated, and the initial generation is configured. Each individual is initialized randomly.

  The fitness E (v) is calculated with respect to the individual v of the current generation by the following formula. Where b is a parameter obtained from the gene of the individual v. Also, var is a function for obtaining the variance.

  Based on the fitness, the basic operation of the genetic algorithm is performed to create the next generation from the current generation (generation change).

  The generation change is repeated a certain number of times from the initial generation to generate the final generation. An optimized parameter b (1) is obtained by taking out the individual with the highest fitness from the final generation and analyzing the gene of the individual.

  The calculation of the second dimension g2 of the dimension compression / linear orthogonal transformation is performed by the following equation.

  The parameter b (2) of g2 is determined by a genetic algorithm in the same manner as the parameter b (1) of g1. However, b (2) is obtained under the constraint that satisfies the following expression so that the inner product with b (1) becomes zero.

  Similarly, the m-th order gm of the dimension compression / linear orthogonal transformation is the parameters b (m) and b (1), b (m) and b (2),..., B (m) and b (m− Under the constraint that the inner product of 1) is all 0, the gm value of the average vector is obtained, and b (m) is determined so that the variance is maximized.

  Note that the present invention does not prevent the use of principal component analysis or the like in determining the parameters b (1) and b (2) to b (m).

  As an output of the dimension compression / linear orthogonal transformation construction unit 12, a dimension compression / linear orthogonal transformation matrix of M rows and C columns is obtained. However, M is the number of output dimensions of dimension compression / linear orthogonal transformation.

  When performing speech recognition, a recognized speech signal is input to the speech input unit 1 and converted into a frequency vector sequence.

  The feature quantity extraction unit 21 uses the class identification transformation obtained by the class identification transformation construction unit 11 and the dimensional compression / linear orthogonal transformation obtained by the dimensional compression / linear orthogonal transformation construction unit 12 to convert the frequency vector series into acoustic features. Convert to quantity. In the frequency vector sequence obtained by inputting the recognized speech signal to the speech input unit 1, for each frequency vector of the frequency vector sequence, the class identification transformation matrix A and the dimension compression are expanded into the extended frequency vector x ′ as in the following equation: Multiply the linear orthogonal transformation matrix B to obtain an M-dimensional acoustic feature quantity z.

  The speech recognition unit 2 performs speech recognition using the acoustic feature amount output from the feature amount extraction unit 21. A hidden Markov model is used as a model, and acoustic features are matched with the model using a Viterbi algorithm, and the most likely word hypothesis candidate is output as a recognition result.

FIG. 2 shows an embodiment of the invention described in claim 2.
The speech input unit 1 inputs the speech of the speaker, samples the input speech signal by the sampling theorem, quantizes it at an appropriate quantization step, and converts it into a digital signal. Subsequently, as shown in FIG. 3, the digital signal is cut out at a certain time length every certain time to extract a voice frame. For each voice frame, frequency information is obtained by using Fourier transform, and the frequency information is collected and vectorized by a mel filter bank to be converted into a frequency vector. A vector sequence (hereinafter referred to as a frequency vector sequence) is generated by arranging the frequency vectors in time series with the entire digitized signal.

  The learning unit 10 calculates the feature amount conversion using the frequency vector series obtained by inputting the learning speech signal to the speech input unit 1 and the correct class information of the learning speech signal. The application unit 20 uses the feature amount conversion obtained by the learning unit 10 to convert a frequency vector sequence obtained by inputting the recognized speech signal to the speech input unit 1 into an acoustic feature amount, and performs speech recognition. Output the recognition result.

  The feature quantity conversion construction unit 16 generates a feature quantity conversion for converting a frequency vector series into an acoustic feature quantity by a genetic algorithm.

A configuration example of feature amount conversion using a genetic algorithm in the feature amount conversion construction unit 16 will be described below.
In the frequency vector sequence obtained by inputting the learning speech signal to the speech input unit 1, the calculation of the first dimension h1 of the feature amount conversion is performed for each frequency vector x (N dimension) of the frequency vector sequence by the following equation: To do.

  The parameter d (1) of the h1 is determined using the learning speech signal and the correct answer class information of the learning speech signal. That is, in the frequency vector sequence obtained by inputting the learning speech signal to the speech input unit 1, an average vector is calculated for each class, the h1 value of the average vector is obtained, and the variance is maximized. , D (1) is determined.

  Specifically, a method for determining the parameter d (1) by a genetic algorithm will be described. A plurality of individuals having a parameter d (1) having N coefficients as genes are generated to form an initial generation. Each individual is initialized randomly.

  The fitness E (v) is calculated with respect to the individual v of the current generation by the following formula. Where d is a parameter obtained from the gene of the individual v. Also, var is a function for obtaining the variance.

  Based on the fitness, the basic operation of the genetic algorithm is performed to create the next generation from the current generation (generation change).

  The generation change is repeated a certain number of times from the initial generation to generate the final generation. The optimized parameter d (1) is obtained by taking out the individual with the highest fitness from the final generation and analyzing the gene of the individual.

  The calculation of the second dimension h2 of the feature amount conversion is performed by the following equation.

  The parameter d (2) of h2 is determined by a genetic algorithm in the same manner as the parameter d (1) of h1. However, d (2) is obtained under the constraint that satisfies the following expression so that the inner product with d (1) becomes zero.

  In the same manner, the m-th order hm of the feature amount conversion is parameters d (m) and d (1), d (m) and d (2),..., D (m) and d (m−1). Under the constraint that the inner products are all 0, the hm value of the average vector is obtained, and d (m) is determined so that the variance becomes maximum.

  As an output of the feature quantity conversion construction unit 16, a feature quantity conversion matrix of M rows and N columns is obtained. However, M is the number of output dimensions of feature quantity conversion.

  When performing speech recognition, a recognized speech signal is input to the speech input unit 1 and converted into a frequency vector sequence.

  The feature quantity extraction unit 26 converts the frequency vector series into an acoustic feature quantity using the feature quantity conversion obtained by the feature quantity conversion construction unit 16. In the frequency vector sequence obtained by inputting the recognized speech signal to the speech input unit 1, each frequency vector x of the frequency vector sequence is multiplied by a feature amount transformation matrix D as shown in the following equation to obtain an M-dimensional sound. A feature value z is obtained.

  The speech recognition unit 2 performs speech recognition using the acoustic feature amount output from the feature amount extraction unit 26. A hidden Markov model is used as a model, and acoustic features are matched with the model using a Viterbi algorithm, and the most likely word hypothesis candidate is output as a recognition result.

  In addition, embodiment of this invention is not limited to the said embodiment, You may change the said embodiment in the range which does not deviate from the meaning of this invention.

  FIG. 4 shows the speech recognition accuracy of the speech recognition apparatus according to the embodiment of the invention described in claim 1 shown in FIG. 1 and the feature extraction for generating the speech input unit 1 and the Mel scale cepstrum coefficient. The speech recognition accuracy of the conventional speech recognition apparatus including the speech recognition unit 2 and the speech recognition unit 2 is evaluated with 8 different noise types using the speech recognition common evaluation infrastructure corpus CENSREC-1 under a noisy environment.

  FIG. 5 shows the speech recognition accuracy of the speech recognition apparatus according to an embodiment of the invention described in claim 1 shown in FIG. 1 and the feature extraction for generating the speech input unit 1 and the mel scale cepstrum coefficient. The speech recognition accuracy of the conventional speech recognition apparatus including the speech recognition unit 2 and the speech recognition unit 2 is evaluated with seven different noise intensities using the speech recognition common evaluation infrastructure corpus CENSREC-1 in a noisy environment.

  The speech recognition experiments shown in FIGS. 4 and 5 were performed under the same conditions as the reference method (baseline) attached to the speech recognition common evaluation infrastructure corpus CENSRECEC-1 in a noisy environment.

  From FIG.4 and FIG.5, the speech recognition apparatus by one Embodiment of invention of Claim 1 shown in FIG. 1 compared with the conventional speech recognition apparatus, the speech recognition performance improved significantly, It can be seen that the performance of speech recognition under noise, which is the object of the present invention, has been achieved.

DESCRIPTION OF SYMBOLS 1 ... Speech input part 2 ... Speech recognition part 10 ... Learning part 11 ... Class identification transformation construction part 12 ... Dimension compression / linear orthogonal transformation construction part 16 ... Feature quantity transformation construction part 20 ... Application part 21 ... Feature quantity extraction part 26 ... Feature extraction unit

Claims (3)

A voice input unit that inputs a voice signal of a speaker and converts the voice signal into a digital signal; converts the digital signal into a vector sequence;
A class identification conversion construction unit that inputs a vector sequence of learning speech signals output by the speech input unit and correct class information of the learning speech signals and outputs class identification conversion;
Dimensional compression / linear that inputs a vector sequence of the learning speech signal output from the speech input unit, correct class information of the learning speech signal, and the class identification transformation and outputs dimensional compression / linear orthogonal transformation An orthogonal transform construction unit;
A feature amount extraction unit that inputs a vector sequence of recognized speech signals output from the speech input unit, the class identification transformation, and the dimension compression / linear orthogonal transformation, and extracts an acoustic feature amount;
A speech recognition unit that performs speech recognition based on the acoustic feature amount generated by the feature amount extraction unit;
A speech recognition apparatus comprising: A voice input unit that inputs a voice signal of a speaker and converts the voice signal into a digital signal; converts the digital signal into a vector sequence;
A feature amount conversion construction unit that inputs a vector sequence of learning speech signals output by the speech input unit and correct class information of the learning speech signals and outputs feature amount conversion by a genetic algorithm;
A feature quantity extraction unit that inputs a vector sequence of recognized speech signals output by the voice input unit and a feature quantity conversion output by the feature quantity conversion construction unit, and extracts an acoustic feature quantity;
A speech recognition unit that performs speech recognition based on the acoustic feature amount generated by the feature amount extraction unit;
A speech recognition apparatus comprising:   A program for causing a computer to perform voice recognition, wherein the voice recognition device according to claim 1 or 2 is realized by a computer.

Images