JP6401126B2 - Feature amount vector calculation apparatus, feature amount vector calculation method, and feature amount vector calculation program. - Google Patents

Description

  The present invention relates to a feature vector calculation device, a speech recognition device, a feature vector calculation method, and a feature vector calculation program.

  In recent years, HMM acoustics based on DNN (Deep Neural Network), which have higher recognition accuracy than HMM (Hidden Markov Model) acoustic models (GMM-HMM acoustic models) based on GMM (Gaussian Mixture Model) as acoustic models in speech recognition technology. A model (DNN-HMM acoustic model) has been used (for example, see Non-Patent Documents 1 and 2). In the DNN-HMM acoustic model, the recognition accuracy of the input speech data affected by the speaker, noise, channel, etc. fluctuates. Therefore, adaptation of the DNN-HMM acoustic model to various fluctuation factors has been actively studied. (For example, see Non-Patent Documents 3 and 4). For example, the DNN-HMM acoustic model adaptation to speaker variation mainly based on a feature vector called i-vector that expresses speaker features with vectors of about tens to hundreds of dimensions is simple and highly accurate. It attracts attention as a technique (see, for example, Non-Patent Documents 4 and 5).

Geoffrey Hilton, Li Deng, Dong Yu, George E. Dahl, Abdel-rahman Mohamed, Navdeep Jaitly Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N. Sainath, and Brian Kingsbury, “Deep Neural Networks for Acoustic Modeling in Speech Recognition, ”[Online], SIGNAL PROCESSING MAGAZINE 2012, Volume: 29, Issue: 6, p.82-p.97, [Search June 29, 2015], Internet <http://www.isip.piconepress.com /courses/temple/ece_8527/lectures/2014_spring/lecture_38_spmag.pdf> T. Yoshioka, MJF Gales, “Environmentally robust ASR front-end for deep neural network acoustic models,” [online], Computer Speech & Language, Volume 31, Issue 1, May 2015, p.65-p.86, [Heisei Search on June 29, 2015], Internet <http://www.sciencedirect.com/science/article/pii/S0885230814001259> Hank Liao, “SPEAKER ADAPTATION OF CONTEXT DEPENDENT DEEP NEURAL NETWORKS,” [online], in Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference, 26-31 May 2013, p.7947-p.7951, [Heisei Searched on June 29, 2015], Internet <http://mazsola.iit.uni-miskolc.hu/~czap/letoltes/IS14/IS2014/PDF/AUTHOR/IS140624.PDF> Michael L. Seltzer, Dong Yu, Yongqiang Wang, “AN INVESTIGATION OF DEEP NEURAL NETWORKS FOR NOISE ROBUST SPEECH RECOGNITION,” [online], in Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference, 26-31 May 2013 , p.7398-p.7402, [Search June 29, 2015], Internet <http://research.microsoft.com/pubs/194344/0007398.pdf> George Saon, Hagen Soltau, David Nahamoo and Michel Picheny, “Speaker Adaptation of Neural Network Acoustic Models Using I-Vectors,” [online], in Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE Workshop, 8-12 Dec 2013, p.55-p.59, [Searched on June 29, 2015], Internet Mickael Rouvier, Benoit Favre, “Speaker adaptation of DNN-based ASR with i-vectors: Does it actually adapt models to speakers ?,” [online], in Proc. Of INTERSPEECH, [searched June 29, 2015], Internet <http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.481.855&rep=rep1&type=pdf> Tetsuji Ogawa, Sayaka Shioda, “Speaker recognition using i-vector”, Journal of the Acoustical Society of Japan, Vol.70, No.6, p.332-p.339, 2014-06-01, The Acoustical Society of Japan Yi Hu, Philipos C. Loizou, “Subjective comparison and evaluation of speech enhancement algorithms,” [online], in Acoustics, Speech and Signal Processing 2006, ICASSP 2006 Proceedings, 2006 IEEE International Conference (Volume: 1), [2015 Search June 29], Internet <http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2098693/> Yun Lei, Nicolas Scheffer, Luciana Ferrer, Mitchel MacLaren, “A NOVEL SCHEME FOR SPEAKER RECOGNITION USING A PHONETICALLY-AWARE DEEP NEURAL NETWORK,” [online], in Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference, 4 -9 May 2014, P.1714-1718, [Search June 29, 2015], Internet <http://www.sri.com/sites/default/files/publications/dnn.pdf> P. Kenny, V. Gupta, T. Stafylakis, P. Ouellet and J. Alam, “Deep Neural Networks for extracting Baum-Welch statistics for Speaker Recognition,” [online], in The Speaker and Language Recognition Workshop 16 -19 June 2014, [Search June 29, 2015], Internet <http://cs.uef.fi/odyssey2014/program/pdfs/28.pdf> Daniel Garcia-Romero, Xiaohui Zhang, Alan McCree, Daniel Povey, “IMPROVING SPEAKER RECOGNITION PERFORMANCE IN THE DOMAIN ADAPTATION CHALLENGE USING DEEP NEURAL NETWORKS,” [online], in Spoken Language Technology Workshop (SLT), 2014 IEEE, 7-10 Dec 2014, [Search June 29, 2015], Internet <http://www.danielpovey.com/files/2014_slt_dnn.pdf>

  However, in the above technique, DNN-HMM acoustic model adaptation based on i-vector is performed assuming clean input voice data that is not affected by noise or channel distortion. Alternatively, even if the input voice data is affected by noise and channel distortion, DNN-HMM acoustic model adaptation is performed without taking any measures against them.

  Here, since i-vectors are extracted based on the features of the input audio data, if noise or channel distortion is added to the input audio data, the extracted i-vector is also affected by noise or channel distortion. Receive. Therefore, when the input voice data is affected by noise or channel distortion, the effect of adapting the DNN-HMM acoustic model based on the i-vector is reduced.

  An example of the embodiment disclosed in the present application is to suppress, for example, a reduction in the effect of DNN-HMM acoustic model adaptation based on a feature vector.

  In an example of the embodiment of the present application, a first feature vector is extracted from input speech. A second feature amount vector is extracted from the speech in which noise or channel distortion reduction processing has been performed on the input speech. Then, based on the parameters of the mixed distribution model obtained by learning the voice that has been subjected to reduction processing on the voice including noise or distortion, the probability that the second feature vector corresponds to each distribution of the mixed distribution model is indicated. Calculate the posterior probability. Then, an average vector of each distribution in the mixed distribution model is calculated from the speech including noise or distortion and the posterior probability. Then, the 0th-order Baum-Welch statistic and the 1st-order Baum-Welch statistic for the input speech are calculated from the first feature vector, the posterior probability, and the average vector. Then, a feature vector is calculated from the zeroth-order Baum-Welch statistic and the first-order Baum-Welch statistic.

  According to an example of the embodiment disclosed in the present application, for example, it is possible to suppress a reduction in the effect of DNN-HMM acoustic model adaptation based on a feature vector.

FIG. 1 is a diagram illustrating an example of an outline of input of a basic feature vector into a DNN-HMM acoustic model according to the related art. FIG. 2 is a diagram illustrating an example of an outline of basic feature vector and i-vector input to a DNN-HMM acoustic model according to the related art. FIG. 3 is a diagram illustrating an example of an outline of an i-vector extraction procedure according to the related art. FIG. 4 is a diagram illustrating an example of the i-vector calculation apparatus according to the embodiment. FIG. 5 is a flowchart illustrating an example of the i-vector extraction process according to the embodiment. FIG. 6 is a diagram illustrating an example of a computer that realizes an i-vector calculation apparatus and a speech recognition apparatus including the i-vector calculation apparatus according to the embodiment by executing a program.

  Hereinafter, an exemplary embodiment related to the disclosed technology of the present application will be described with reference to the drawings. The disclosed technology of the present application is not limited by the following embodiments. Further, the following embodiments may be appropriately combined. Prior to the description of the embodiments disclosed in the present application, the premise prior art will be described, and then the embodiments disclosed in the present application will be described.

In addition, in the following description, when “^ A” is written with respect to the symbol A, it is equivalent to “a symbol with a ^ immediately above A” as shown in the following equation (1-1). Suppose there is. In addition, when “-A” is written for the symbol A, it is assumed to be equivalent to “a symbol with − immediately above A” as shown in the following equation (1-2). Also, when “{A} _α ^β ” is written for the symbol A, “α is written as a subscript to the right of {A}, as shown in the following equation (1-3): { It is assumed that β is equivalent to a symbol in which β is superscripted on the right side of A}. Further, “A” is described as “Vector A” when A is a vector, “Matrix A” when A is a matrix, and “Set A” when A is a set.

[Input of basic feature vector to DNN-HMM acoustic model according to prior art]
FIG. 1 is a diagram illustrating an example of an outline of input of a basic feature vector into a DNN-HMM acoustic model according to the related art. As shown in FIG. 1, in general, in speech recognition, input speech data is acoustically analyzed in units of a frame length of about 30 msec and a frame shift of about 10 msec, and about 40-dimensional MFCC (Mel-Frequency Cepstral Coefficient) or FBANK A basic feature vector such as (log-mel Filter BANK) is extracted for each frame.

  As shown in FIG. 1, when a basic feature vector of one frame is given, the DNN-HMM acoustic model outputs a posterior probability vector of the HMM state of the frame. More specifically, the DNN-HMM acoustic model is provided with basic feature vectors having a total dimension of about several hundred to several hundreds, for example, in which the frame and feature vectors for five frames before and after the frame are connected. In response, the posterior probability vector of the HMM state of the frame is output. This basic framework for speech recognition is described in detail in Non-Patent Documents 1 and 2, for example.

[Input of basic feature vector and i-vector to DNN-HMM acoustic model according to prior art]
FIG. 2 is a diagram illustrating an example of an outline of basic feature vector and i-vector input to a DNN-HMM acoustic model according to the related art. As shown in FIG. 2, apart from basic feature vectors such as MFCC and FBANK, feature values called i-vectors that represent speaker features contained in input speech data with vectors of about several tens to several hundreds of dimensions A vector is extracted from the input speech data. Then, an extended feature vector obtained by concatenating the basic feature vector and the i-vector is given to the DNN-HMM acoustic model and used mainly for speech recognition adapted to speaker variation. The effectiveness of this method is described in detail in Non-Patent Documents 4 and 5, for example.

  Here, when i-vector is used for speech recognition of input speech data to which noise and channel distortion are added, the DNN-HMM acoustic model is applied to noise and channel variations in addition to speaker variations. Therefore, it is desirable that the i-vector includes noise and channel distortion information in addition to speaker characteristics. i-vector was originally developed in the field of speaker recognition.

[Existing procedure of i-vector according to the prior art]
FIG. 3 is a diagram illustrating an example of an outline of an i-vector extraction procedure according to the related art. The i-vector extraction procedure will be described below. Hereinafter, only the part related to the disclosed technology in the i-vector extraction procedure will be described. The details of the appearance of i-vector and the extraction procedure are described in detail in Non-Patent Document 7, for example.

  A standard i-vector extraction method in speaker recognition of the prior art is a GMM (Gaussian Mixture Model) -UBM (Universal Background Model) (GMM-UBM) approach. UBM is a kind of GMM. In the GMM-UBM approach, a “voice-like” model (UBM) is learned using a large amount of speech data for UBM learning of a large number of unspecified speakers, and a new speaker model (GMM) This is a method of adaptively obtaining UBM using a small amount of voice data of a speaker.

  On the other hand, in recent speaker recognition, a framework using a GMM super vector obtained by connecting GMM average vectors by the number of mixtures as a feature vector has become mainstream. The GMM super vector represents audio data as time series data as one point on a vector space. i-vector is also based on this GMM supervector.

Here, a D-dimensional L frame feature vector sequence X _u obtained from the input speech data u is defined as in the following equation (2). The feature vector x _t (t = 1, 2,..., L) is, for example, MFCC, and its dimension number D is, for example, 40.

In addition, c = 1, 2,..., C is a subscript (for example, C = 2048) representing a Gaussian distribution of UBM (GMM), and the mixing weight π _{c of} the c-th Gaussian distribution and the c-th average vector m _c, when the diagonal covariance matrix sigma _c, the parameter set Ω of UBM, represented by (3) below.

At this time, the likelihood p (x _t | Ω) of the UBM with respect to the feature quantity vector x _t is given by the following equation (4).

  The speaker-independent CD (C × D) -dimensional GMM supervector m obtained from the UBM is expressed by the following equation (5). However, T on the right side of the equation is a transposition symbol of a matrix or a vector.

The CD-dimensional GMM super vector M _u of the input audio data u is assumed to be obtained by the following equation (6).

Here, the matrix T in the above equation (6) is a CD dimension × M dimension rectangular matrix (M << CD) called a total variation matrix, and the vector w _u is an M dimension i− with respect to the input speech data u. vector. In other words, the i-vector represents the characteristics of the speakers included in each input speech data u as a “difference” (dimensionally compressed) from the average speaker (average of UBM) in the GMM super vector space. It can be said that.

Hereinafter, a specific series of extraction procedures of the vector w _u which is an i-vector will be described. First, let γ _t (c) be the posterior probability that x _t is generated from the Gaussian distribution that is the c-th in the UBM. The posterior probability γ _t (c) is obtained by the following equation (7).

Using the posterior probability γ _t (c), the 0th-order and 1st-order Baum-Welch statistics N _{u, c} and the vector F _{u, c} for the input speech data u using UBM are expressed by the following equation (8) and Each can be written as shown in equation (9). However, the vector _{Fu, c} is a D-dimensional vector.

Further, using the above equations (8) and (9), a matrix N _u that is a 0th-order and first-order Baum-Welch statistic, a vector, as in the following equations (10) and (11): Define _Fu . However, the matrix N _u is a CD dimension × CD dimension matrix, and the vector F _u is a D dimension vector.

Here, the matrix I _D that appears in the diagonal component of the above equation (10) is a D-dimensional × D-dimensional unit matrix. Further, the matrix Σ is a D-dimensional × D-dimensional diagonal matrix that models residual fluctuation components that cannot be expressed by the total fluctuation matrix T. Although the calculation procedure of the matrix T and the matrix Σ is omitted, the i-vectorw _u can be calculated as in the following equation (12) using the above. Note that the matrix I _M in the following equation (12) is an M-dimensional × M-dimensional unit matrix.

Specific series of extraction steps of i-vectorw _u shown in the above (7) to (12) can be divided roughly into two steps. The <first procedure> corresponds to the above equation (7), and the feature quantity x _t (t = 1, _t) of each frame of the L frame feature quantity vector sequence X _u obtained from the input speech data u. 2,..., L) is a procedure for calculating the posterior probability γ _t (c) generated from the c-th Gaussian distribution of UBM. <Second procedure> is a procedure for calculating i-vectorw _u according to the above equations (8) to (12) using the posterior probability γ _t (c) calculated in the above equation (7). It is.

Each Gaussian distribution in the UBM ideally corresponds to each phoneme context including phoneme information including the dependency of several phonemes before and after. In the <first procedure> of i-vector extraction, the posterior probability γ _t (c) is calculated, which corresponds to probabilistic estimation of the phoneme context of the vector x _t . Possible to accurately calculate the posterior probability γ _t (c) is a second step in the i-vector extraction, the i-vectorw _u representing the characteristics of the speaker, the speech contents of the phoneme contexts or input audio data u It is indispensable to calculate with high accuracy without depending on.

When speech recognition is performed in an actual environment, noise and channel distortion are often added to the input speech data u. In this case, it becomes difficult to accurately calculate the posterior probability γ _t (c) by the <first procedure> of i-vector extraction. As a result, the <second procedure> of i-vector extraction This makes it difficult to calculate i-vectors with high accuracy. In order to solve this problem, i-vector extraction represented by the above equations (7) to (12) is performed after reducing noise and channel distortion from the input speech data u using, for example, some speech enhancement technique. A method of performing a series of procedures is possible.

According to this method, the posterior probability γ _t (c) can be accurately calculated by the <first procedure> of i-vector extraction, but the <second procedure of i-vector extraction Since the processing target in <> is information with reduced noise and channel distortion, noise and channel distortion information will be lost from the actually calculated i-vector, and in addition to speaker characteristics, noise And channel distortion information is also included in the i-vector, which is inconvenient if you want to use it actively in speech recognition.

[I-vector extraction according to the embodiment]
As described above, in the i-vector extraction procedure, the first embodiment reduces the noise and channel distortion included in the input audio data u in the <first procedure> of i-vector extraction. The posterior probability γ _t (c) is calculated with high accuracy. (Second requirement) In the <second procedure> of i-vector extraction, in addition to speaker characteristics, information including noise and channel distortion is included. That is, i-vector is calculated using input speech data u including noise and channel distortion.

FIG. 4 is a diagram illustrating an example of the i-vector calculation apparatus according to the embodiment. i-vector calculating unit 10, a first basic feature amount extracting section 11A, second basic feature amount extracting unit 11B, ^ γ _t (c) calculating unit 12, -m _c calculating unit 13, ^ N _{u, c} , ^ F _{u, c} calculation unit 14 and i-vector calculation unit 15. The first basic feature amount extracting section 11A, second basic feature amount extracting unit 11B, ^ γ _t (c) calculating unit 12, -m _c calculating unit _{13, ^ N u, c,} ^ F u, c The calculation unit 14 and the i-vector calculation unit 15 are processing units that perform processing in cooperation with a processing device such as a CPU (Central Processing Unit) and a temporary storage device such as a RAM (Random Access Memory). May be.

  In the embodiment, the D-dimensional and Q-frame feature vector time series O extracted from a large amount of speech data for UBM learning of a large number of unspecified speakers to which noise and channel distortion are added is represented by the following (13): Define it like an expression. The feature vector time series O is stored in the noise distortion speech feature storage unit 100A.

  In addition, a large amount of voice data for UBM learning of a large number of unspecified speakers to which noise and channel distortion are added is extracted from voice data obtained by reducing noise and channel distortion using a predetermined voice enhancement technique. A D-dimensional, Q-frame feature vector time series ^ O to be defined is defined as in the following equation (14). The feature vector time series ^ O is stored in the noise distortion reduced speech feature storage unit 100B.

A D-dimensional and L-frame feature vector sequence X _u extracted from input speech data u to which noise and channel distortion have been added is defined as in the following equation (15). The first basic feature quantity extraction unit 11A extracts a feature quantity vector series X _u from the input speech data u by the following equation (15).

For the input speech data u, the D-dimensional and L-frame feature vector time series { _circumflex over (X) _} u obtained from the input speech data {circumflex over (u)} using the above-described predetermined speech enhancement technique to reduce noise and channel distortion (16). The second basic feature quantity extraction unit 11B extracts the feature quantity vector series ^ X _u from the input speech data u by the following equation (16). Then, each likelihood p (^ x _t | ^ Ω) for ^ x _t (t = 1, 2,..., L) of the UBM (hereinafter referred to as ^ UBM) learned using the vector sequence ^ O. ) Can be written as the following equation (17).

Here, c = 1, 2,..., C is a subscript representing the UBM Gaussian distribution (for example, C = 2048), and the c-th Gaussian distribution weight ^ π _c and the c-th average vector ^ Assuming that m _{c is} a diagonal covariance matrix ^ Σ _c , the parameter set ^ Ω of ^ UBM is expressed by the following equation (18). The UBM parameter set ^ Ω is calculated from the feature vector time series ^ O by the UBM learning device 200 and stored in the ^ UBM storage unit 300.

^ Γ _t (c) Using the UBM parameter set ^ Ω, the calculation unit 12 obtains the feature vector ^ x _t (t = 1, 2,..., L) of each frame of ^ X _u The posterior probability ^ γ _t (c) generated from the c-th Gaussian distribution of UBM is calculated as in the following equation (19).

The posterior probability ^ γ _t (c) is calculated using ^ UBM with reduced noise and channel distortion and the vector sequence ^ X _u with reduced noise and channel distortion. This is the <first procedure> of the i-vector extraction procedure that satisfies (1). Subsequently, the ^ N _{u, c} , _{Fu, c} calculation unit 14 uses the posterior probability ^ γ _t (c) to calculate the 0th-order and 1st-order Baum-Welch statistics ^ N _u for the input speech data _{u. , c} and vector { _{circumflex over} (F) _{}, c} , respectively, as shown in the following equations (20) and (21). However, the vector ^ F _{u, c} is a D-dimensional vector.

It should be noted here that in the above equation (21), the feature vector x _t (t = 1, 2,) of the input speech data u in which noise and channel distortion are added to the calculation of the vector F _{u, c} . ..., L). By calculating the vector ＦF _{u, c} in this way, the i-vector that is finally extracted retains noise and channel distortion information in addition to the speaker characteristics. In the above procedure, the above (second requirement) is satisfied.

Incidentally, -m _c calculating unit 13, the -m _c in the above equation (21), a posterior probability ^ gamma _t (c), derived from the sound data for UBM learning represented by the formula (13) Using the feature vector time series O of the D dimension and Q frame, calculation is performed as shown in the following equation (22).

Even if the UBM is learned using the feature vector time series O, it is impossible to take correspondence between the UBM Gaussian distribution number and the ^ UBM Gaussian distribution number. it can not be used an average vector m _c of the Gaussian distribution of the number c. That is, UBM and ^ UBM are different, and since there is no relationship between the Gaussian distribution that constitutes UBM and the Gaussian distribution that constitutes ^ UBM, there is no relationship between the distribution numbers of both Gaussian distributions. is there. In other words, ^ even known Gaussian distribution number in UBM, this number is different from the distribution number of the Gaussian distribution at UBM, since it is impossible to obtain the Gaussian distribution number in UBM, feature vector x _t (t = 1,2, ···, L ) can not be obtained c-th mean vector m _c to subtract from. To solve this problem, by using a Gaussian distribution number in ^ UBM, according to the above (22), approximated -m _c of the c-th mean vector m _c.

Finally, the i-vector calculation unit 15 calculates i-vectorw _u by the following equations (23), (24), and (25).

[I-vector extraction processing according to the embodiment]
FIG. 5 is a flowchart illustrating an example of the i-vector extraction process according to the embodiment. First, the first basic feature quantity extraction unit 11A of the i-vector calculation apparatus 10 extracts a feature quantity vector series X _u (first basic feature quantity) from the input speech data u by the above equation (15). (Step S11). Next, the second basic feature quantity extraction unit 11B extracts a feature quantity vector series ^ X _u (second basic feature quantity) from the input speech data u by the above equation (16) (step S12). In addition, the execution order of step S11 and step S12 may be before and after, or may be simultaneous.

Next, ^ gamma _t (c) calculating unit 12, ^ the parameter set ^ Omega the UBM, by using the feature quantity vector sequence ^ X _u, ^ X _u feature vector ^ x _t (t of each frame = 1,2, ···, L) is ^ posterior probability is generated from c-th Gaussian distribution UBM ^ gamma _t and (c), calculated as above (19) (step S13).

Next, -m _c calculating unit 13, to give the -m _c in the above equation (21), a posterior probability ^ gamma _t (c), from the voice data for UBM learning represented by the formula (13) Using the D-dimensional and Q-frame feature vector time series O and the posterior probability ^ γ _t (c), calculation is performed as in the above equation (22) (step S14).

_{Next, ^ N u, c, ^} F u, c calculating unit 14, a feature vector sequence X _u, the posterior probability ^ γ _t (c), and a -m _c, 0-order with respect to the input audio data u First-order Baum-Welch statistics ^ N _{u, c} and vector ^ F _{u, c} are calculated as shown in the above equations (20) and (21) (step S15).

Next, i-vector calculating unit 15, the above equation (23), (24) and (25), to calculate the i-vectorw _u (step S16). The i-vector calculation device 10 outputs the i-vectorw _u calculated in step S15. i-vectorw _u, for example, as shown in FIG. 2, the basic feature vector and i-vectorw _u is expanded feature vectors, which are connected, for example DNN-HMM acoustic model is input to the obtained HMM state posterior probability vector It can be applied to a speech recognition apparatus that performs speech recognition using

  Note that any speech enhancement processing technique can be applied as a method of reducing noise and channel distortion in the i-vector extraction procedure according to the above embodiment. Various speech enhancement processing techniques are described in detail in Non-Patent Document 8, for example. Alternatively, as a method of reducing the influence of noise and channel distortion, a processing technique using a bottleneck feature amount obtained from a DNN-HMM acoustic model may be used instead of the voice enhancement processing technique. The bottleneck feature amount is described in detail in Non-Patent Document 2, for example.

  Further, the mixed distribution model learned using the feature vector time series ^ O is not limited to the UBM based on the GMM but may be an HMM.

[Evaluation experiment]
The conventional technique compared with the embodiment is the conventional technique described in Non-Patent Documents 4 and 5. The following (Table 1) and (Table 2) are tables showing evaluation experiment results when the i-vector calculated by the i-vector calculation apparatus 10 of the embodiment is input to the acoustic model of DNN. The percentage in each table is the word error rate (WER).

  In (Table 1), the left side of the “+” symbol represents “type of feature quantity used in the <first procedure> of i-vector extraction”, and the right side of the “+” symbol represents “i-vector extraction”. <Type of feature amount used in <second procedure> ". “Noisy MFCC” is a noise MFCC, “Bottleneck” is a Bottleneck feature, and “VTS enhanced” is a vector tailor expansion enhancement.

  Table 1 shows that for any combination, WER reduction was seen over baseline DNN.

  Table 2 shows the WER depending on the difference in the size of the bottleneck feature amount used in the ^ UBM mixed distribution model learning during i-vector extraction. (Table 2) shows that WER reduction was seen over baseline DNN for any size.

  Each process performed in the i-vector calculation apparatus 10 and the speech recognition apparatus including the i-vector calculation apparatus 10 is realized by a processing apparatus such as a CPU and a program that is analyzed and executed by the processing apparatus. May be. In addition, each process performed in the i-vector calculation apparatus 10 and the speech recognition apparatus including the i-vector calculation apparatus 10 may be realized as hardware by wired logic.

  In addition, among the processes described in the embodiment, all or a part of the processes described as being automatically performed can be manually performed. Alternatively, all or some of the processes described as being manually performed among the processes described in the embodiments can be automatically performed by a known method. In addition, the above-described and illustrated processing procedures, control procedures, specific names, and information including various data and parameters can be changed as appropriate unless otherwise specified.

(About the program)
FIG. 6 is a diagram illustrating an example of a computer that realizes an i-vector calculation apparatus and a speech recognition apparatus including the i-vector calculation apparatus according to the embodiment by executing a program. The computer 1000 includes a memory 1010 and a CPU 1020, for example. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. In the computer 1000, these units are connected by a bus 1080.

  The memory 1010 includes a ROM 1011 and a RAM 1012. The ROM 1011 stores a boot program such as BIOS. The hard disk drive interface 1030 is connected to the hard disk drive 1031. The disk drive interface 1040 is connected to the disk drive 1041. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1041. The serial port interface 1050 is connected to a mouse 1051 and a keyboard 1052, for example. The video adapter 1060 is connected to the display 1061, for example.

  The hard disk drive 1031 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the program that defines each process of the i-vector calculation device 10 and the speech recognition device including the i-vector calculation device 10 is stored in, for example, the hard disk drive 1031 as a program module 1093 in which a command executed by the computer 1000 is described Remembered. For example, a program module 1093 for executing information processing similar to the functional configuration in the i-vector calculation device 10 and the speech recognition device including the i-vector calculation device 10 is stored in the hard disk drive 1031.

  The setting data used in the processing of the embodiment is stored as program data 1094 in, for example, the memory 1010 or the hard disk drive 1031. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1031 to the RAM 1012 as necessary, and executes them.

  Note that the program module 1093 and the program data 1094 are not limited to being stored in the hard disk drive 1031, but may be stored in, for example, a removable storage medium and read out by the CPU 1020 via the disk drive 1041 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and the program data 1094 may be read by the CPU 1020 via the network interface 1070.

  The embodiments are included in the invention disclosed in the claims and equivalents thereof, as well as included in the technology disclosed in the present application.

10 i-vector calculating device 11A first basic feature amount extracting section 11B second basic feature amount extracting unit 12 ^ γ _t (c) calculating unit 13 -m _c calculating unit _{14 ^ N u, c, ^} F u, _c calculation unit 15 i-vector calculation unit 1000 computer 1010 memory 1020 CPU

Claims (6)

A first feature quantity extraction unit for extracting a first feature quantity vector from input speech;
A second feature amount extraction unit that extracts a second feature amount vector from speech that has been subjected to noise or channel distortion reduction processing on the input speech;
Based on the parameters of the mixture distribution model obtained by learning the speech on which noise or channel distortion reduction processing has been performed on the speech including noise or distortion, the second feature vector is included in each distribution of the mixture distribution model. A posterior probability calculation unit for calculating a posterior probability indicating a corresponding probability;
An average vector calculation unit that calculates an average vector of each distribution in the mixed distribution model from the noise or distortion-containing speech and the posterior probability;
A statistic calculator that calculates a zero-order Baum-Welch statistic and a first-order Baum-Welch statistic for the input speech from the first feature vector, the posterior probability, and the average vector; ,
A feature vector calculation apparatus comprising: a feature vector calculator that calculates a feature vector from the zero-order Baum-Welch statistic and the first-order Baum-Welch statistic. The feature quantity vector calculation apparatus according to claim 1, wherein the reduction process is a voice enhancement process. The feature vector calculation apparatus according to claim 1, wherein the reduction process is a process using a bottleneck feature quantity. Wherein a first feature vector as an input the extended feature vectors obtained by connecting the feature quantity vector calculated by the feature quantity vector calculation unit to predetermined acoustic model, the speech to the speech recognition processing the input speech The feature vector calculating apparatus according to claim 1 , further comprising a recognition processing unit . A feature vector calculation method executed by a feature vector calculator,
A first feature amount extraction step of extracting a first feature amount vector from the input speech;
A second feature amount extracting step of extracting a second feature amount vector from the speech in which noise or channel distortion reduction processing has been performed on the input speech;
The probability that the second feature vector corresponds to each distribution of the mixed distribution model is determined based on the parameters of the mixed distribution model obtained by learning the voice subjected to the reduction processing on the voice including noise or distortion. A posterior probability calculation step for calculating a posterior probability to be shown;
An average vector calculation step of calculating an average vector of each distribution in the mixed distribution model from the speech including the noise or distortion and the posterior probability;
A statistic calculation step of calculating a zeroth-order Baum-Welch statistic and a first-order Baum-Welch statistic for the input speech from the first feature vector, the posterior probability, and the average vector; ,
A feature vector calculation method comprising: calculating a feature vector from the zeroth-order Baum-Welch statistic and the first-order Baum-Welch statistic. According to claim 1, the feature quantity vector calculation program for causing a computer to function as a feature quantity vector calculation apparatus according to 3 or 4.

Images