KR101041035B1 - Method and Apparatus for rapid speaker recognition and registration thereof - Google Patents

Abstract

The present invention relates to a registration method for fast speaker recognition, and a registration method for fast speaker recognition according to an embodiment of the present invention provides a method for concealing a parameter of a subspace distribution clustering hidden Markov model (SDCHMM) with respect to an entire space. Converting to a cove model (HMM); Converting the hidden Markov model (HMM) into a linear spectral domain; And adapting the Hidden Markov Model (HMM) of the linear spectral domain to the speaker using maximum likelihood linear spectral transformation. According to the method, the speaker model can be quickly adapted to a small amount of adaptation data in a small mobile terminal such as a mobile phone, and it can effectively handle addition noise or channel noise, and can greatly reduce the amount of computation in the recognition process. There is an effect that the small terminal itself can handle speaker recognition without the need for help from other devices such as servers.

Description

Fast speaker recognition method and device, Registration method and device for fast speaker recognition {Method and Apparatus for rapid speaker recognition and registration}

The present invention relates to speaker recognition, and more particularly, to a registration method and apparatus for fast speaker recognition, and a fast speaker recognition method and apparatus.

A number of schemes for speech or sound detection and recognition systems have been proposed and implemented to some extent. These systems theoretically match the user's utterance against the registered speaker's pronunciation to allow or deny access to the device or system's resources according to the user's identity, or to identify or customize the registered speaker. Command libraries can be invoked.

Speaker recognition technology is a technology that determines whether the speaker is a registered speaker or whether the registered speaker is correct.

Large systems that include large resources can have many potential users. Therefore, a large number of registered speakers may require a significant amount of storage and processing overhead to recognize the speaker. As the number of speakers increases, simple and fast systems designed to quickly differentiate between different speakers will experience the performance saturation of the speaker recognition system.

A speaker-dependent system refers to a system that performs decoding and alignment of pronunciation on a decoded script model, for example, hidden Markov models (HMM) adapted to different speakers. The performance of most speaker-dependent systems can be saturated and degraded not only when the number of speakers is high but also when the number of speakers is small in fast and simple systems.

For example, a text-independent system such as frame-by-frame feature clustering and classification may be considered as a fast matching technique for speaker to speaker class identification. However, the number of speaker classes that can be processed with a substantial amount of processing overhead within an acceptable response time and the number of speakers in each class are limited. In other words, the classifier for each frame requires a relatively small amount of data for each registered speaker and less processing time for a limited number of speakers, while increasing the number of models reduces the difference between speaker models. Therefore, the ability to distinguish is limited. Attempts to reduce information about speaker pronunciation may degrade the system's ability to identify individual registered users as the number of users increases. When the number of speakers is quite large, the speaker recognition system or engine can no longer distinguish some speakers. This state is called saturation.

On the other hand, complex systems that use speaker-dependent model-based decoders adapted to individual speakers to provide speaker recognition are considerably slower because they have to run models in parallel or sequentially to achieve speaker recognition. Ask for time. Also, these models are typically difficult to learn and adapt because they require large amounts of data to form the model.

In a template matching system, storage conditions are somewhat relaxed, which is not only speaker dependent but also text dependent, depending on the specific pronunciation of each registered speaker specific to its speaker recognition and / or authentication function. However, such a system is inherently incapable of being transparent to the user, because such a system requires a relatively long registration and initial recognition (eg logon) process and often stops using the system periodically for authentication. It requires you to do it. More importantly, this system is responsible for each speaker's pronunciation variation (intraspeaker variation) that may be caused by each speaker's aging, fatigue, illness, stress, prosody, psychological state, and other conditions. More sensitive

More specifically, the speaker dependent speech recognizer builds a model for each speaker during the registration phase during operation. The speaker and its pronunciation are then recognized by the model that produces the greatest similarity or the lowest error rate. Enough data is needed to adapt each model to its own speaker so that all pronunciations are recognized. For this reason, most speaker dependent systems are also text dependent and use template matching to reduce the amount of data to be stored in each model. In contrast, systems using, for example, HMMs or similar statistical models typically generate a cohort model based on a group of speakers, rejecting speakers that are not overly likely.

In general, model-based methods are the most common, a Gaussian mixture model (GMM) or HMM. The speaker recognition process to which this is applied is divided into two types: the first is the speaker registration process and the second is the recognition process. The registration process is the process of receiving the speaker's voice and registering the speaker's model. This process is generally done by making the speaker independent model into speaker dependent model using adaptive technique. The maximum likelihood linear regression (MLLR) or maximum a posterior (MAP) algorithm is mainly used for adaptation. These adaptive algorithms perform better with more voices from registered speakers. However, in order to recognize a speaker in a mobile environment where the environment changes frequently, a method of quickly adapting a speaker's voice using a small amount of adaptation data for tens of seconds to several minutes is required for speaker registration. The recognition process is to select the speaker of the model with the highest likelihood value by calculating the likelihood of the speaker's speech to be recognized in each of the adapted models. However, if there are too many speaker-dependent models registered in the speaker recognizer, the calculation volume is increased because there are too many models to be compared when the test voice is received. Especially in devices with low computing power, such as mobile devices, this problem is very important.

Existing adaptive algorithms require a few minutes or tens of minutes for the speaker to register in order to show high performance. Also, the existing speaker adaptation method is not effective in the new noise environment. On the other hand, even in speaker recognition, there is a problem in that there are too many calculations because the number of models to be compared is increased in the determination process when there are many registered speakers.

Accordingly, the first technical problem to be achieved by the present invention is to provide a registration method for fast speaker recognition capable of robust adaptation even with a small amount of adaptation data and effectively processing addition noise or channel noise.

The second technical problem to be achieved by the present invention is to provide a fast speaker recognition method that can effectively handle addition noise or channel noise even when the number of registered speakers does not increase greatly in the recognition process.

The third technical problem to be achieved by the present invention is to provide a registration apparatus for fast speaker recognition capable of robust adaptation even with a small amount of adaptation data, and which can effectively handle addition noise or channel noise.

The fourth technical problem to be achieved by the present invention is to provide a fast speaker recognition apparatus that can effectively handle addition noise or channel noise even when the number of registered speakers does not increase greatly in the recognition process.

In order to achieve the first technical problem, a registration method for fast speaker recognition according to an embodiment of the present invention may include the parameters of the subspace distribution clustering hidden Markov model (SDCHMM) for the entire space. Converting); Converting the hidden Markov model (HMM) into a linear spectral domain; And adapting the Hidden Markov Model (HMM) of the linear spectral domain to the speaker using maximum likelihood linear spectral transformation.

Advantageously, in the step of converting to the linear spectral domain, the Hidden Markov Model (HMM) is converted into a log spectral domain using Discrete Cosine Inverse Transformation (IDCT), and the Hidden Markov Model (HMM) of the log spectral domain. ) Can be converted into a linear spectral domain.

Preferably, the registration method for fast speaker recognition according to an embodiment of the present invention converts the adaptive hidden Markov model (HMM) into a spectrum domain to adapt the subspace distribution clustering hidden Markov to the speaker. The method may further include extracting a model (SDCHMM) parameter. In this case, in the step of extracting the subspace-distributed clustering hidden Markov model (SDCHMM) parameter adapted to the speaker, the adaptive hidden Markov model (HMM) is transformed into a spectrum domain adapted for the entire space. Obtain a Hidden Markov Model (HMM) and apply a link structure transformation to the Hidden Markov Model (HMM) adapted for the entire space to determine the parameters of the subspace distribution clustering Hidden Markov Model (SDCHMM) adapted to the speaker. Can be extracted. More specifically, obtaining an adaptive hidden Markov model (HMM) for the entire space comprises converting the adapted hidden Markov model (HMM) to a log spectral domain and the adapted concealment of the log spectral domain. The discrete cosine transform (DCT) may be applied to the Markov model (HMM) to obtain a hidden Markov model (HMM) adapted to the entire space.

In order to achieve the second technical problem, a fast speaker recognition method according to an embodiment of the present invention is a Gaussian extracted subspace distribution clustering hidden Markov models (SDCHMMs) adapted from one or more speakers. Comparing with subspaces of; And outputting which one of the speakers corresponds to the registered speaker according to the comparison result. Here, the subspace distributed clustering hidden Markov model (SDCHMM) adapted to one or more speakers converts the hidden Markov model (HMM) transformed from the parameters of the subspace distributed clustering hidden Markov model (SDCHMM) into the linear spectral domain. And adapting the Hidden Markov Model (HMM) of the linear spectral domain to the speaker using the maximum likelihood linear spectral transformation, and converting the adapted Hidden Markov Model (HMM) to the cepstrum domain. .

Advantageously, in the step of outputting which speaker corresponds to any of the registered speakers, the maximum for said any speaker of subspace distribution clustering hidden Markov models (SDCHMMs) adapted to said one or more speakers. Subspace distribution clustering hidden Markov model (SDCHMM) with likelihood can be extracted.

In order to achieve the third technical problem, the registration apparatus for fast speaker recognition according to an embodiment of the present invention uses the parameters of the subspace distribution clustering hidden Markov model (SDCHMM) as a hidden Markov model (HMM) for the entire space. A model converting unit for converting); Converts the hidden Markov model (HMM) to a linear spectral domain, and converts the speaker-adapted hidden Markov model (HMM) to a spectral domain in a linear spectral domain, Domain conversion unit for obtaining HMM); A speaker adaptor adapted to adapt a hidden Markov model (HMM) transformed into a linear spectral domain to a speaker using a maximum likelihood linear spectral transformation; And a link structure transform unit for applying a link structure transform to the hidden markov model (HMM) adapted for the entire space to extract parameters of the subspace distribution clustering hidden markov model (SDCHMM) adapted to the speaker.

Advantageously, the domain transform unit comprises a discrete cosine inverse transform unit that converts the hidden Markov model (HMM) into a log spectral domain; And a time domain transform unit converting the hidden Markov model (HMM) of the log spectral domain into a linear spectral domain.

Advantageously, the domain transform unit comprises: a log spectral transform unit for converting the adapted hidden Markov model (HMM) into a log spectral domain; And a discrete cosine transform unit for obtaining a hidden Markov model (HMM) adapted to the entire space from the adapted hidden Markov model (HMM) of the log spectral domain.

In order to achieve the fourth technical problem, a fast speaker recognition apparatus according to an embodiment of the present invention includes a memory unit for storing subspace distribution clustering hidden Markov models (SDCHMMs) adapted to one or more speakers; A microphone unit for receiving a voice of an arbitrary speaker; A likelihood calculator for comparing subspace distribution clustering hidden Markov models (SDCHMMs) stored in the memory unit with subspaces of Gaussian extracted from the input speech; And a maximum likelihood selector for outputting to which of the registered speakers the corresponding speaker corresponds to the comparison result. Here, the subspace distributed clustering hidden Markov model (SDCHMM) stored in the memory unit converts the hidden Markov model (HMM) transformed from the parameters of the subspace distributed clustering hidden Markov model (SDCHMM) into the linear spectral domain, A maximum likelihood linear spectral transformation is used to adapt a hidden Markov model (HMM) of the linear spectral domain to the speaker and to convert the adapted hidden Markov model (HMM) to a cepstrum domain.

According to the method, the speaker model can be quickly adapted to a small amount of adaptation data in a small mobile terminal such as a mobile phone, and it can effectively handle addition noise or channel noise, and can greatly reduce the amount of computation in the recognition process. There is an effect that the small terminal itself can handle speaker recognition without the need for help from other devices such as servers.

Hereinafter, with reference to the drawings will be described a preferred embodiment of the present invention. However, embodiments of the present invention illustrated below may be modified in various other forms, and the scope of the present invention is not limited to the embodiments described below.

In the present invention, a subspace distribution clustering hidden markov model (SDCHMM) is used for a speaker recognizer used in a mobile environment. SDCHMM divides the entire multivariate Gaussian into feature spaces called subspaces, and then quantizes the distributions belonging to each subspace to quantize the subspace distribution prototype, or codeword. ) To represent the original full space distribution.

In other words, find a combination of subspace prototypes that most closely resembles the original overall spatial distribution and link them. As such, the SDCHMM expresses the original total spatial distribution using a small number of subspace codewords.

1 is an exemplary diagram of an SDCHMM composed of 13 subspaces.

SDCHMM, consisting of 13 subspaces, represents one total spatial distribution by combining the most similar subspace distributions in 13 subspace distribution prototypes. For example, the distribution on the left of the two groups of segments shown in FIG. 1 is 13 connecting lines including the first column of Sp13, which is expressed as a combination of subspace distributions. Further, the distribution on the right side of the two groups of segments shown in FIG. 1 is 13 connecting lines including the third column of Sp13, which is expressed as a combination of another subspace distribution.

The SDCHMM can represent all distributions of the HMM as a small number of subspace distribution prototypes, reducing memory and computation, making it a good model for mobile devices.

One way to adapt SDCHMM is to adapt the prototype and change the link structure without changing the prototype.

2 is an exemplary diagram of a link structure adaptation method for the SDCHMM.

By adapting the SDCHMM using the link structure adaptation method, it is possible to generate a different speaker model by only changing the link, which can be powerful for adaptive processing. That is, adapting a structure that links fewer prototypes can be more robustly adapted with less data than adapting many different multivariate Gaussian distributions, as is the case with conventional HMMs. In addition, adapting the SDCHMM using the link structure adaptation method can reduce the amount of computation since the likelihoods of the prototypes are computed only once when the likelihoods are calculated during the recognition process. This is because each speaker model shares one prototype and links it.

In speech recognition and speaker recognition, a model is created to measure how high the similarity between the input speech and the model is. Similarity is called likelihood when using an HMM-based model. The likelihood is expressed as in Equation 1.

에서 데이터

가 발생할 확률을 의미한다.This is the model

Data from

Means the probability of occurrence.

When building the model, try to maximize the likelihood of the training data (voice). Creating a model is called learning. When learning to likelihood to be maximum, the concept of maximum likelihood (ML) is used. When learning, enough data should be used. Otherwise, exceptions or data that were not found at the time of learning would occur in the test situation and interfere with accurate recognition.

The transformation of a well-trained model with less data is called adaptation. Similar to the adaptive learning, the model is transformed to maximize the likelihood of the adaptive data. Both MLLR and ML-LST are adapted using ML techniques.

However, there is no way to maximize the likelihood directly. Equation 2 shows a Q-function used to maximize the likelihood.

와

는 각각 이전 모델과 재추정된 모델을 의미하며, 수학식 2가 최대가 되면 수학식 3이 만족된다. Of equation (2)

Wow

Denotes a previous model and a reestimated model, respectively, and when Equation 2 becomes the maximum, Equation 3 is satisfied.

와

를 각각 적응되기 전의 모델과 적응 후의 모델이라고 가정하고 수학식 2가 최대가 되도록 하면 적응 전의 모델보다 적응 후의 모델에서의 적응 데이터

의 우도가 크게 된다. 이와 같이 수학식 3을 만족시키는 과정을 반복하면 점점 우도가 커지게 되고 최대 우도에 근접하게 된다.

Wow

Is assumed to be a model before adaptation and a model after adaptation, so that Equation 2 becomes the maximum.

Likelihood becomes large. As described above, if the process of satisfying Equation 3 is repeated, the likelihood is gradually increased and the maximum likelihood is approached.

3 illustrates a process of combining addition noise and channel noise in the time domain.

The clean voice uttered by a human is contaminated by channel noise caused by channel noise between an additive noise generated by ambient noise and an input device such as a microphone.

Since speaker recognition generally operates in the cepstral domain, most adaptive algorithms are designed to be performed in the cepstral domain. However, in the spectral domain, it is difficult to express summative noise or channel noise directly.

Maximum likelihood linear spectral transform (ML-LST) can be used in the spectral domain to clearly represent noise in the time domain.

4 illustrates a process of domain conversion of voice data for ML-LST.

Since speaker recognition is processed in the spectral domain and adaptation using ML-LST is applied in the spectral domain, domain transformation is required before adaptation. Inverse Discrete Cosine Transformation (IDCT) and exponential operation are applied sequentially to transform the model into the linear spectral domain. Adapted data from the linear spectral domain is returned back to the spectral domain for recognition.

5 illustrates a speaker registration process according to the present invention.

In the present invention, the speaker-independent model is converted into a speaker-dependent model using ML-LST and link structure adaptation. In ML-LST, the new model is transformed into a linear spectral domain before the adaptation, noise is applied, and then returned to the spectral domain. ML-LST is more effective in new noise environments than conventional MLLRs or MAPs because it adapts in the spectral domain to effectively handle additive or channel noise.

Equation 4 shows the noise mean vector μ in the cepstrum domain.

Here, C means N * M DCT (Discrete Cosine Transformation) matrix. M is the number of filter banks, N is the dimensionality of μ. A and b represent noise.

6 illustrates a detailed domain change process.

After converting the parameters of the SDCHMM into HMM for the entire space (610), the IDCH (620) and the exponential operation (630) in the spectral domain is converted into a linear spectral domain. ML-LST adaptation 640 and the application of noise are in the linear spectral domain.

Since Equation 4 is a model after adaptation, the likelihood of this model is maximized. In order to maximize the likelihood of Equation 4, Equation 4 is applied to the Q function of Equation 2, and a condition for obtaining the maximum Q function is obtained. Repeating this process results in ML-LST adaptation with maximum likelihood in the linear spectral domain.

After that, the log operation 650 and the DCT 660 are converted into the cepstrum domain. In Equation 4, the noises A and b are applied in the linear spectral domain and then converted back into the spectral domain through the log and DCT.

Finally, link structure transformation of the HMM in the cepstrum domain generates an adapted SDCHMM parameter.

7 is a block diagram of a registration apparatus for fast speaker recognition according to an embodiment of the present invention.

The model converter 710 converts the parameters of the SDCHMM into the HMM for the entire space.

The domain converter 720 converts the HMM into a linear spectral domain.

In addition, the domain converting unit 720 converts the HMM of the linear spectral domain that the speaker adapting unit 730 has adapted to the speaker to a spectral domain to obtain an HMM adapted to the entire space.

The speaker adaptation unit 730 adapts the HMM to the voice of the speaker to be registered. That is, the speaker adaptation unit 730 adapts the HMM converted into the linear spectral domain to the speaker using the ML-LST.

The link structure converter 740 extracts a parameter of the SDCHMM adapted to the speaker by applying the link structure transform to the HMM of the spectrum domain generated by the domain converter 720, that is, the HMM adapted to the entire space. These parameters constitute the speaker dependency model.

8 illustrates a speaker recognition process according to the present invention.

In the present invention, the speaker is recognized by calculating a likelihood between the speaker dependent model configured as described above and the Gaussian subspace of the input voice, and selecting a model having the maximum likelihood among them.

9 is a block diagram of a fast speaker recognition apparatus according to an embodiment of the present invention.

The memory unit 910 stores SDCHMMs adapted to one or more speakers. The SDCHMM stored in the memory unit 910 converts the HMM converted from the parameters of the SDCHMM into the linear spectral domain, and adapts the HMM of the linear spectral domain to the speaker using the maximum likelihood linear spectral transformation, as described above. This model is created by converting the HMM into a spectral domain. The memory unit 910 is configured of a physical memory block such as a volatile memory device or a nonvolatile memory device.

The microphone 920 converts an input speaker's voice into an electrical audio signal.

The likelihood calculator 930 compares the SDCHMMs stored in the memory 910 with Gaussian subspaces extracted from the voice input to the microphone 920.

The maximum likelihood selector 940 outputs whether the speaker who uttered the voice corresponds to any of the pre-registered speakers according to the comparison result of the likelihood calculator 930.

The invention can be implemented via software. Preferably, a program for executing a fast speaker recognition method or a fast speaker recognition method according to an embodiment of the present invention can be recorded and provided on a computer-readable recording medium. When implemented in software, the constituent means of the present invention are code segments that perform the necessary work. The program or code segments may be stored on a processor readable medium or transmitted by a computer data signal coupled with a carrier on a transmission medium or network.

Computer-readable recording media include all kinds of recording devices that store data that can be read by a computer system. Examples of the computer readable recording medium include ROM, RAM, CD-ROM, DVD 占 ROM, DVD-RAM, magnetic tape, floppy disk, hard disk, optical data storage, and the like. The computer readable recording medium can also be distributed over network coupled computer devices so that the computer readable code is stored and executed in a distributed fashion.

Although the present invention has been described with reference to one embodiment shown in the drawings, this is merely exemplary, and it will be understood by those skilled in the art that various modifications and variations may be made therefrom. And, such modifications should be considered to be within the technical protection scope of the present invention. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

The present invention relates to a speaker registration and recognition technique that can not only quickly adapt a speaker model to a small amount of adaptation data, but also effectively handle addition noise or channel noise, and greatly reduce the amount of computation in the recognition process. Small mobile terminals such as mobile phones or speaker-recognized devices with a lot of resources can be applied to any environment where it is difficult to collect enough data to register a speaker and must be quickly recognized with a small number of calculations.

1 is an exemplary diagram of an SDCHMM composed of 13 subspaces.

2 is an exemplary diagram of a link structure adaptation method for the SDCHMM.

3 illustrates a process of combining addition noise and channel noise in the time domain.

4 illustrates a process of domain conversion of voice data for ML-LST.

5 illustrates a speaker registration process according to the present invention.

6 illustrates a detailed domain change process.

7 is a block diagram of a registration apparatus for fast speaker recognition according to an embodiment of the present invention.

8 illustrates a speaker recognition process according to the present invention.

9 is a block diagram of a fast speaker recognition apparatus according to an embodiment of the present invention.

Claims (12)

Converting the parameters of the subspace distribution clustering hidden Markov model (SDCHMM) into a hidden Markov model (HMM) for the entire space; Converting the hidden Markov model (HMM) to a log spectral domain; Converting a hidden Markov model (HMM) of the log spectral domain to a linear spectral domain; And Adapting the Hidden Markov Model (HMM) of the linear spectral domain to the speaker using maximum likelihood linear spectral transformation Including, a registration method for fast speaker recognition. delete The method of claim 1, Converting the adapted hidden markov model (HMM) into a spectral domain to extract the subspace-distributed clustering hidden markov model (SDCHMM) parameters adapted to the speaker; Way. The method of claim 3, wherein Extracting the subspace distribution clustering hidden Markov model (SDCHMM) parameter adapted to the speaker, Converting the adapted hidden Markov model (HMM) into a cepstrum domain to obtain a hidden Markov model (HMM) adapted for the entire space; And Applying a link structure transform to a hidden Markov model (HMM) adapted for the entire space to extract parameters of the subspace-distributed clustering hidden Markov model (SDCHMM) adapted to the speaker. How to register for The method of claim 4, wherein Obtaining a hidden Markov model (HMM) adapted for the entire space, Converting the adapted hidden Markov model (HMM) to a log spectral domain; And Applying a discrete cosine transform (DCT) to the adapted hidden markov model (HMM) of the log spectral domain to obtain an adapted hidden markov model (HMM) for the entire space. How to register. Comparing subspace distribution clustering hidden Markov models (SDCHMMs) adapted to one or more speakers with Gaussian subspaces extracted from the voice of any speaker; Outputting, according to the comparison result, to which of the registered speakers any of the speakers correspond to; The subspace distribution clustering hidden Markov model (SDCHMM) adapted to the one or more speakers is: Convert the Hidden Markov Model (HMM) transformed from the parameters of the subspace distribution clustering Hidden Markov Model (SDCHMM) into a linear spectral domain, and using the maximum likelihood linear spectral transform, the Hidden Markov Model (HMM) of the linear spectral domain ) Is a model generated by adapting the speaker to the speaker and converting the adapted hidden Markov model (HMM) into a cepstrum domain. The method of claim 6, The step of outputting to which of the registered speakers any speaker, Extracting a subspace distributed clustering hidden Markov model (SDCHMM) with a maximum likelihood for any of the subspace distributed clustering hidden Markov models (SDCHMMs) adapted to the one or more speakers. Fast speaker recognition method. A computer-readable recording medium having recorded thereon a program for executing the method of any one of claims 1 and 3 to a computer system. A model converter for converting the parameters of the subspace distribution clustering hidden Markov model (SDCHMM) into the hidden Markov model (HMM) for the entire space; Converts the hidden Markov model (HMM) to a linear spectral domain, and converts the speaker-adapted hidden Markov model (HMM) to a spectral domain in a linear spectral domain, Domain conversion unit for obtaining HMM); A speaker adaptor adapted to adapt a hidden Markov model (HMM) transformed into a linear spectral domain to a speaker using a maximum likelihood linear spectral transformation; And A link structure transform unit for extracting a parameter of a subspace distribution clustering hidden markov model (SDCHMM) adapted to the speaker by applying a link structure transform to the hidden Markov model (HMM) adapted to the entire space Registration apparatus for a fast speaker recognition, including. The method of claim 9, The domain conversion unit, A discrete cosine inverse transform unit for converting the hidden Markov model (HMM) into a log spectral domain; And And a time domain converter for converting the hidden Markov model (HMM) of the log spectral domain into a linear spectral domain. 11. The method of claim 10, The domain conversion unit, A log spectral converter for converting the adapted hidden Markov model (HMM) into a log spectral domain; And And a discrete cosine transform unit for obtaining a hidden markov model (HMM) adapted for the entire space from the adapted hidden markov model (HMM) of the log spectral domain. A memory unit for storing subspace distribution clustering hidden Markov models (SDCHMMs) adapted to one or more speakers; A microphone unit for receiving a voice of an arbitrary speaker; A likelihood calculator for comparing subspace distribution clustering hidden Markov models (SDCHMMs) stored in the memory unit with subspaces of Gaussian extracted from the input speech; And a maximum likelihood selector for outputting which of the registered speakers corresponds to the speaker according to the comparison result. The subspace distribution clustering hidden Markov model (SDCHMM) stored in the memory unit is Convert the Hidden Markov Model (HMM) transformed from the parameters of the subspace distribution clustering Hidden Markov Model (SDCHMM) into a linear spectral domain, and using the maximum likelihood linear spectral transform, the Hidden Markov Model (HMM) of the linear spectral domain ) Is a model generated by adapting the speaker to the speaker and converting the adapted hidden Markov model (HMM) into a cepstrum domain.

Images