KR20050037125A - Apparatus for recognizing a voice by using hidden markov model - Google Patents

Abstract

The present invention relates to a speech recognition apparatus using an HMM, comprising: a sampling and quantization block for converting an analog speech signal into a digital speech signal, a noise removing unit for removing ambient noise from the converted digital speech signal, and HMM outputting speech recognition results through a feature vector extraction unit for extracting features of digital speech signals in the form of a vector, and an architecture HMM algorithm implemented in hardware with probability parameters determined by learning about the extracted feature vectors in advance. It includes a speech recognition block using. Therefore, voice recognition can be performed independently without a computer, and can be conveniently installed in a device to be used. Also, by implementing a parallel structure, real-time voice recognition is possible. It is effective that it can be applied to various fields such as a terminal and a PDA.

Description

Speech recognition device using Hidden Markov model {APPARATUS FOR RECOGNIZING A VOICE BY USING HIDDEN MARKOV MODEL}

The present invention relates to a speech recognition apparatus using a Hidden Markov Model (HMM), and more particularly, to a speech recognition apparatus implemented using a hardware architecture of a systolic array structure using the HMM used as a speech recognition algorithm.

In general, voice recognition devices are applied to products such as voice recognition phones, voice recognition computers, voice recognition cars, etc., and the public's interest in such voice recognition is increasing, but voice recognition technology is still deep in our lives. It is not available until now, and there are few products using voice recognition around us, and there is no advantage in using the products.

On the other hand, in foreign countries, communication services and products using voice recognition have appeared. In other words, considering the fact that the study of voice recognition in foreign countries has been more than 40 years, the appearance of such application products is late compared to other users.

In reality, in Korea, speech recognition technology is still incomplete. In other words, the speech recognition technology has not been able to adequately meet the needs of the general user in many parts, that is, the number of words, speaker independence, recognition method, and environment. However, the speech recognition matching is performed by using the HMM algorithm. It is meeting the needs.

Here, the HMM algorithm was introduced by Baum and his colleagues in the late sixties and early seventies as the most widely used method for speech recognition, and was applied to speech processing by Baker, Jelinek and his colleagues in the seventies. do.

In other words, under the assumption that the speech signal can be modeled as a Markov model, the HMM obtains the stochastic parameters of the Markov model using the speech corpus in the training process, then creates a reference Markov model and the reference Markov model most similar to the input speech during the recognition Recognize the voice in a way to select it.

The emergence of full-fledged applications using this speech recognition technology began in the late 1980s, and in Korea, it can be seen after the early 1990s.

Therefore, as the speech recognition apparatuses currently used are mostly software-oriented, in order to apply the actual speech recognition apparatuses to real life, it is necessary to develop a hardware architecture that implements the evaluation and decoding portions of the speech recognition apparatuses as chips.

Accordingly, the present invention has been made in accordance with the needs of the above-described hardware architecture development, the object of the present invention is to implement a speech recognition device using a HMM that can be implemented in a hardware architecture of the systolic array structure using the HMM used as a speech recognition algorithm In providing.

In the present invention for achieving the above object, a speech recognition apparatus using the HMM comprises a sampling and quantization block for converting an analog speech signal into a digital speech signal, a noise removing unit for removing ambient noise from the converted digital speech signal; Speech recognition results through a feature vector extraction unit that extracts the features of the noise-free digital speech signal in the form of a vector, and an architecture HMM algorithm implemented in hardware with probability parameters determined by learning the extracted feature vectors in advance. Characterized in that it comprises a speech recognition block using the HMM to output the.

Hereinafter, with reference to the accompanying drawings will be described in detail the operation of the preferred embodiment according to the present invention.

1 is a block diagram of a speech recognition apparatus using an HMM according to the present invention. After receiving a user's voice signal S1 through a microphone S2, the user's digital voice is sampled and quantized. Sampling and quantization block 10 for converting to a signal, a noise removing unit 20 for removing the ambient noise except the user's digital voice signal to increase the recognition rate of the converted digital voice signal, and the user of the noise is removed A feature vector extractor 30 for extracting a feature of the digital speech signal in the form of a vector so as to be used in a matching part that is actually recognized in the digital speech signal, and a probability that is determined through learning in advance after receiving the extracted feature vector. In performing the speech recognition through the HMM algorithm with parameters, it is hard to match the portion of the speech recognition that is matched. The voice recognition block 40 using the HMM outputs a voice recognition result through an architecture implemented by wear.

Based on the above configuration, the operation of the speech recognition apparatus using the HMM according to the present invention will be described in more detail.

First, the sampling and quantization block 10 receives the user's voice signal S1 through the microphone S2, and then converts the provided analog voice signal into a digital voice signal by sampling and quantizing the noise signal 20. To provide.

The noise removing unit 20 removes the surrounding noise in order to increase the recognition rate among the converted digital voice signals, and then provides only the remaining digital voice signals to the feature vector extracting unit 30.

The feature vector extractor 30 extracts the features of the digital speech signal in the form of a vector so as to be used in a matching part where actual recognition is performed on the user's digital speech signal from which the noise is removed, and extracts the features of the digital speech signal into the speech recognition block 40 using the HMM. to provide.

The speech recognition block 40 using the HMM is a core technology of the present invention, and is an architecture part that implements a matching part in hardware. The speech recognition result is output through the HMM algorithm with the probability parameter determined through.

On the other hand, the part of outputting the speech recognition result through the HMM algorithm in the speech recognition block 40 using the HMM will be described in more detail.

In other words, the HMM algorithm uses a more complex HMM than the conventional reference Markov model, to accommodate various changes in speech patterns.

In other words, since HMM is a double probability process that estimates an unobservable process through a process for generating an observable symbol, it is an appropriate modeling method for expressing a multivariate and unknown process such as speech.

In particular, the HMM is a group of states connected to each other by state transitions, and each transition has two kinds of probabilities. That is, one is a transition probability that governs a change of state, and the second is an output probability that defines a conditional probability that each output symbol is emitted from a finite number of objects when a transition is made.

As a result, in expressing each feature of the speech pattern with state transition probability and output probability, the state transition probability causes the current state to depend only on the state immediately before it, and the output probability is performed independently of each other.

In addition, when performing the three tasks that must be performed when implementing a voice recognition system using the HMM, defining several elements used are as follows.

- next -

N is the number of model states, each set of states As the state at time t It is defined as

M is the number of output symbols in each state, and each set of symbols is It is defined as

A = Is the state transition probability, Is,

Is the probability of state observation, Is,

Is the initial state probability, Is,

The output symbol string is to be.

As mentioned above, the defined HMM is This is simply indicated by three things in order for this HMM to be used in practice.

First, as an evaluation task, a forward algorithm is used to calculate the probability of an observation sequence given a model and an observation sequence.

Second, as a decoding operation, an optimal state sequence is selected when an observation sequence is required as a recognition process. In this case, the Viterbi algorithm is used.

Here, the Viterbi algorithm is

Is indicated by

here, Represents the best score (highest probability) accumulated along the path with the highest probability up to time (t) and state (i).

In addition, the process of calculating the optimal state string cost through a recursive method is as follows.

1, Initialization, ego,

2, Recursion, As the multiplication is very difficult and complicated in hardware implementation, if you design the structure with only the adder by taking -log on both sides, the matching cost can be configured with only the simple adder and the comparator. ,

3, Termination to be.

This will be described with reference to an example of the speech recognition process illustrated in FIG. 2.

That is, referring to FIG. 2, the voice waveform S2-1 of "Hello everyone" is shown, of which the HMM (S2-4) of the part corresponding to "multiple" is shown.

In other words, "several" is represented by the concatenation of the phoneme HMMs "o", "ㅕ", "ㄹ", and "ㅓ", and the feature vector (S2) in each state (S2-2) of the model "ㅕ". -3) shows the process of generating.

In other words, to be a x _1, x _2, x ₃ occurring in state S _1, and the x _4, x ₅ occurs in the state S _2, indicates that the x _6, x ₇ balsaengdoem in state S _3.

In addition, one feature vector (S2-3) is extracted from the speech signal at frame shift time intervals, and the transition probability indicates the frequency of transition from state s _i to s _j , and the observation probability density function b _i (x _t ) is The probability of observing feature vector x _t in state s _i .

In addition, the transition probability in the HMM having only the transition from left to right can be obtained from the ratio of the number of frames staying in each state and the number of frames moving to the next state.

Third, as a learning task, by adjusting each parameter for optimal modeling as a problem about learning required for the training process, it is solved by the Baum-Welch algorithm.

The present invention implements an evaluation task and a decoding task in hardware among the three tasks performed to actually use the HMM, and the learning task is implemented in software, which is an existing method.

Referring to Figure 3, the overall configuration diagram of the hardware architecture of the HMM algorithm according to the present invention, the transition probability ( , transition probability) and output probability ( memory S3-1 storing the observation probability and the left registers S3-2-0 and S3-2 that temporarily store the probability values stored in the memory S3-1 and provide them to the processing element. -j-1, S3-2-j, S3-2-j + 1, S3-2-N-1) and registers on the right side (S3-4-0, S3-4-j-1, S3-4 -j, S3-4-j + 1, S3-4-N-1) and N elements, each of which is connected to an element one level below and two levels below, from the left and right registers. Processing elements S3-3-0, S3-3-j-1, S3-3-j, S3-3-j + 1, S3-3 which calculate the provided probability values and provide them to the matching system S3-5. N-1) and a matching system (S3-5) for finally matching and determining the probability value calculated in the processing element.

First, the internal structure of any of the processing elements S3-3-0, S3-3-j-1, S3-3-j, S3-3-j + 1, S3-3-N-1 See FIG. 4.

That is, state transition probabilities are added to the cost values calculated at the previous time t-1 through the respective adders S4-2, 3, 4, and the result ( , , ) Is input to the comparator S4-1.

The comparator S4-1 outputs the minimum value of the three inputs provided through the respective adders S4-2, 3, and 4 as an output, and the state observation probability is equal to the exported minimum value. Is added through the adder (S4-5) to the cost value of the current time t ( )

The cost value calculated by the adder (S4-5) ) Passes through the right side register S4-6 shown in FIG. 3 for the calculation of the next time t + 1 and then into the input of the comparator S4-1 and its two processing elements in its processing element.

5 is a diagram illustrating an HMM model algorithm according to the present invention.

That is, the state transition probability is We are using a left-right model with the same properties.

In other words, the state never goes backwards and never changes significantly. Here, to show the transition as shown in FIG. Is 2, the initial state should always start at 1, and the matching should start at the origin of FIG. The conditions of

6A to 6C illustrate processing elements S3-3-0, S3-3-j-1, S3-3-j, S3-3-j + 1, and S3-3-N-1 shown in FIG. Is a diagram showing the data flow of the.

That is, referring to FIG. 6A, the feature vectors t ₀ , t ₁ ,..., T _N-1 corresponding to one frame provided by the feature vector extractor 30 of FIG. 1 are stored in the memory S3-1. Enter the area A, and the corresponding transition probability a parameter is serially registered on the left side of the register (S3-2-0, S3-2-j-1, S3-2-j, S3-2-j + 1, S3). -2-N-1) to update the processing elements (S3-3-0, S3-3-j-1, S3-3-j, S3-3-j + 1, S3-3-N-1) do. In this case, the required number of clocks is 3N + 1 clocks when there are N processing elements.

Next, in the second drawing of FIG. 6A, the feature vectors t ₀ , t ₁ ,..., T _N-1 corresponding to one frame provided by the feature vector extractor 30 are stored in the memory S3-1. Entering the area B, the corresponding transition probability b parameter is serially registered on the right side of the register (S3-4-0, S3-4-j-1, S3-4-j, S3-4-j + 1, S3). -4-N-1) to update the processing elements (S3-3-0, S3-3-j-1, S3-3-j, S3-3-j + 1, S3-3-N-1) do. At this time, the required clock number is N + 1 clocks.

As such, when both a and b parameters are updated in the processing element, as shown in the third figure of FIG. 6A, the matching cost value is calculated using the updated value.

Next, FIG. 6B is a diagram in which the process of FIG. 6A is repeatedly performed while the feature vectors t ₀ , t ₁ , ..., t _N-1 are shifted by one space, and when the feature vectors of one frame are all finished. Do until.

Next, as shown in FIG. 6C, the feature vector, which has completed the repetition process, returns to its original position and prepares for matching with the next reference. Determine the largest value and the corresponding reference is the speech recognition result we want to recognize.

FIG. 7 is a diagram illustrating a process of creating a next frame. When the process of FIG. 6 is completed in one frame, the process needs to be changed to the next frame for the next recognition. This process removes the front feature vector of the frame and the rest of the feature vector. Shift by one space and attach the new feature vector to the end to the next frame. As such, when the next frame is reached, the voice recognition result is extracted while repeating the process of FIG. 6.

As described above, the present invention is implemented by a hardware architecture of a systolic array structure using HMM used as a speech recognition algorithm, so that speech recognition can be performed independently without a computer, and also conveniently installed in a device to be used. It is possible to realize real-time voice recognition by implementing the parallel structure, and thus it is effective to be applied to various fields such as voice recognition home appliances, automobiles, toys, mobile communication terminals and PDAs.

1 is a block diagram of a speech recognition apparatus using an HMM according to the present invention;

2 is a diagram illustrating an example of a speech recognition process.

3 is an overall configuration diagram of the hardware architecture of the HMM algorithm according to the present invention,

4 shows the internal structure of a processing element according to the invention,

5 is a diagram illustrating an HMM model algorithm according to the present invention,

6 is a diagram illustrating a data flow of the processing element illustrated in FIG. 3.

7 is a diagram illustrating a process of creating a next frame according to the present invention.

<Description of the symbols for the main parts of the drawings>

10: sampling and quantization block 20: noise canceller

30: feature vector extractor 40: speech recognition block using HMM

S1: analog voice signal S2: microphone

Claims (6)

In the speech recognition device using the Hidden Markov Model (HMM), A sampling and quantization block for converting an analog speech signal into a digital speech signal, A noise removing unit for removing ambient noise from the converted digital voice signal; A feature vector extraction unit for extracting a feature of the noise-free digital voice signal in the form of a vector; Speech recognition block using HMM that outputs speech recognition result through architecture HMM algorithm implemented by hardware with probability parameter determined through learning about the extracted feature vector Speech recognition apparatus using a HMM, characterized in that it comprises a. The method of claim 1, The speech recognition block using the HMM, A memory for storing transition probabilities and output probabilities obtained through learning in advance, A register for temporarily storing a probability value stored in the memory; A processing element which is composed of a plurality of elements, each of which is connected to an element below one level and two levels below, and calculates a probability value provided from the register; Matching system for finally matching and determining the calculated probability value Speech recognition apparatus using HMM characterized in that it comprises a. The method of claim 2, The internal structure of the processing element, A first adder each adding a state transition probability to a cost value calculated at a previous time; The added result ( , , Comparator that selects and outputs the minimum value State observed probability to the output minimum Second adder to add Speech recognition apparatus using a HMM characterized in that it comprises a. The method of claim 3, wherein Cost value of the current time t added by the second adder ) Is a speech recognition device using an HMM, characterized in that it enters its own processing element and the input of the two processing elements for the calculation of the next time. In the speech recognition result extraction method of the processing element, Storing the feature vectors t ₀ , t ₁ , ..., t _N-1 corresponding to one frame in memory; Serially updating certain parameters of the probability of transition to the processing elements via registers on the left and right sides, After the parameter is updated, calculating a matching cost value using the updated value; Extracting a speech recognition result by repeatedly performing the storing and updating process while shifting the feature vector one by one until the feature vector corresponding to the one frame is finished. Speech recognition method using HMM, characterized in that it comprises a. The method of claim 5, wherein When the feature vectors corresponding to the one frame are all finished, the first feature vector of the frame is removed, the remaining feature vectors are shifted by one space, and the new feature vectors are attached to the rear, and the voice recognition is performed on the feature vectors corresponding to the next frame. Speech recognition method using HMM characterized in that to extract the result.