KR100304788B1 - Method for telephone number information using continuous speech recognition - Google Patents

Abstract

The present invention sets the sentence structure predicted pronunciation during phone number guidance, and by incorporating a plurality of sentences among sentences that can be made from the sentence structure to build the voice data necessary for training to induce more natural pronunciation of the user, The present invention relates to a telephone number guidance method using continuous speech recognition to ensure accurate phone number guidance with reliable speech recognition based on continuous speech recognition. When it is off-hook, it outputs the system introduction comment and requests voice input of the desired department and professor name, and the user has a sentence structure that predicts the continuous voice and utterance inputted when the user inputs the name of the department and professor. Compare training models trained with a network Recognize. After that, the recognition result of the recognized input voice is transmitted to the telephone number guide so that the telephone number of the professor concerned is transmitted.

Description

Method for telephone number information using continuous speech recognition}

The present invention relates to phone number guidance using continuous speech recognition, and in particular, sets a sentence structure in which a pronunciation is predicted when guiding a phone number, and has a plurality of sentences among sentences that can be made from the sentence structure. The present invention relates to a telephone number guidance method using continuous speech recognition that induces a more natural pronunciation of a user by constructing and provides accurate phone number guidance with reliable speech recognition based on Korean continuous speech recognition.

If human-machine dialogue is interfaced with the voice of natural dialogue, its application spreads across many fields.

The machine basically exchanges information by the promised code, and in the dialogue between human and machine, there is a restriction to convert the voice into a code that the machine understands. This process is called 'voice recognition'.

The ultimate goal of speech recognition is to realize the natural dialogue between man and machine between man and machine.

There are many methods of speech recognition, but the most widely used methods include a Hidden Markov Model (HMM), Dynamic Time Warping (DTW), and a speech recognition method using an artificial neural network.

The HMM is a recognition method using statistical characteristics of speech, and it is possible to find out various pronunciation patterns of various people through statistical processing, and to grasp temporal characteristics of speech well. However, a significant amount of learning data is required to satisfy the statistical hypothesis, and the introduction of variation sounds to absorb variations caused by articulations increases the complexity of learning.

In addition, since the DTW can recognize whether it is recognized through comparison with the reference pattern, it is based on a relatively simple principle, and in order to recognize the speech, the DTW must attempt to match the reference pattern through a process called time stretching. There is a drawback that the amount of computation and memory usage increase rapidly.

In addition, since neural networks are learned by relatively simple principles, complex algorithms are not required, and real-time implementation is possible at the time of recognition, and it is easy to implement in hardware. On the other hand, it is difficult to reflect the temporal characteristics of the voice signal and has a disadvantage of long learning time.

Looking at the conventional telephone number guide system applying the voice recognition method as described above, if the customer speaks the name of the company or service to make a call, 'voice automatic connection system' that automatically recognizes the customer's voice and connects to the business There is this. Speech recognition technology applied to this system recognizes predetermined words (about 150 words) in speaker independent manner.

In other words, call the list (list of voice recognition service guide, florist, flower delivery, flower shop, florist, real estate, Bokdeokbang, realtor, moving, express, director, packaging director, moving center, Chinese cuisine, Chinese restaurant, If one pronounces one of the words of Chinese food, Jjajangmyeon, Jjajangmyeon, Jjajangmyeon, jokbal, jokbal, etc.), it is recognized and connected.

However, such a phone number guide system is a phone number guide system using isolated word recognition technology. In the case of isolated word recognition, when a pronunciation other than a desired word is input, it is difficult to recognize it. Therefore, there is a disadvantage that users can only use in a limited environment.

Next, there is a voice recognition system called 'Stock information system' by voice recognition. This system can be used by 120 people at the same time. Even if the software (S / W) and the hardware (H / W) are separated and the software version is updated, the hardware is designed to be small. Instead of using two DSPs (digital signal processor) for 1-channel voice recognition, we can redesign the DSP for feature extraction and the DSP for search at an appropriate ratio so that 1. ** DSP can be used for 1-channel voice recognition. It was made. It is designed to be operated even if the electronic phone button, which is the input method used in the telephone information system, is pressed, and it is prioritized to operate by telephone button when the telephone button is input at the same time as voice. Considering the characteristics of the stock market where the number of listings is adjusted every day, the words to be recognized are automatically loaded into the system at a certain time every day, and the words to be recognized are updated. As the speech recognition technology, isolated word recognition using continuous speech recognition technology is used.

However, the speech recognition method in the securities information system also uses a continuous speech recognition technology, but it is a system that recognizes only keywords, and there is a disadvantage in that speech recognition such as continuous sentences is impossible.

Accordingly, an object of the present invention is to provide a speech recognition method capable of solving various problems occurring in the service apparatus and method using the conventional speech recognition as described above, and enabling continuous speech recognition and processing Korean natural language. ,

More specifically, by setting the sentence structure predicted pronunciation during phone number guidance and by constructing the voice data necessary for training with a plurality of sentences among the sentences that can be made from the sentence structure to induce more natural pronunciation of the user In addition, the present invention provides a phone number guidance method using continuous speech recognition to ensure accurate phone number guidance with reliable speech recognition based on Korean continuous speech recognition.

The present invention for achieving the above object,

In the phone number guide method through voice recognition,

When the user dials and the telephone number guide system is in an off-hook state, outputting a system introduction comment and sending a comment requesting voice input of a desired class and professor's name;

When the user inputs a continuous voice including the name of the department and the professor after the voice input request, recognizes the continuous voice by comparing the inputted continuous voice with a training model trained with a sentence set in advance and predicted;

Transmitting the recognition result of the recognized input voice to a telephone number guide to control the telephone number of the corresponding professor to be transmitted;

After sending the phone number of the corresponding professor consists of sending the end comment.

In the above, if a voice is not input within a predetermined time while requesting voice input of a desired class and professor's name, checking whether the set time has elapsed; Resending a comment requesting voice input of the desired department and professor's name if the set time has not elapsed; And transmitting the termination message when a set time has elapsed in the absence of the voice input.

The recognizing of the voice may include setting a sentence structure in which the user's pronunciation is predicted, setting a sentence that may be made from the set sentence structure, and 10 sample sentences among the set sentences. Constructing the voice data required for training, storing the constructed voice data as a training model in a training model unit, and comparing the input voice with the stored training model when a voice is input from the user. Characterized in that it consists of a step of recognizing.

1 is a block diagram of a continuous speech recognition system to which the present invention is applied;

2 is a word network for setting a sentence structure in which the pronunciation of the user is predicted;

3 is a sentence sample table set up for collecting voice data;

4 is a flowchart showing a phone number guide method using continuous speech recognition according to the present invention;

5 is a flow chart illustrating in more detail the continuous speech recognition process of FIG. 4;

6 is a flow chart showing an endpoint detection process in the present invention;

7 is an explanatory diagram for explaining a method for obtaining a zero crossing rate in the present invention;

8 is an explanatory diagram for explaining a starting point and an end point detection process by energy;

9 is an explanatory diagram for explaining the use and overlapping of windows;

10 is a scale diagram of a pitch felt by a person expressed as a function of physical frequency;

11 is an explanatory diagram for explaining a mel scale filter bank;

12 is a waveform showing a mixed Gaussian probability distribution function,

Figure 13 is a table of context independent phoneme models.

<Explanation of symbols for main parts of the drawings>

10: analog / digital converter 20: endpoint detection unit

30: feature extraction unit 40: Viterbi decoder

50: training model unit 60: speech recognition unit

70: phone number guide

Hereinafter, with reference to the accompanying drawings, preferred embodiments of the present invention according to the technical spirit as described above in detail.

1 is a block diagram of a continuous speech recognition system to which the present invention is applied.

As shown in the figure, an analog / digital converter 10 for sampling an input analog voice at a predetermined sampling frequency and converting it into a digital signal, extracting only a voice section from the digital data output from the analog / digital converter 10 A feature extraction unit 30 for extracting feature vectors from the voice data through the endpoint detection unit 20, the endpoint detection unit 20, and a Viterbi decoder 40 for Viterbi decoding the feature vectors obtained by the feature extraction unit 30. The training model unit 50 sets a sentence structure in which a pronunciation is predicted at the time of use of a user, and stores a training sentence pattern in which voice data necessary for training is constructed with a predetermined sentence among sentences that can be made from the sentence structure. , And compares the sentence pattern stored in the training model unit 50 with the input speech data obtained from the Viterbi decoder 40 to be continuous. Voice recognition unit 60 for recognizing a voice and outputting the recognition result, and a telephone number guide unit 70 for guiding a phone number corresponding to the recognition result output from the voice recognition unit 60 by voice.

Referring to Figures 2 to 13 attached to the phone number guide method of the present invention using a continuous speech recognition system to which the present invention configured as described above is as follows.

The telephone number guide method of the present invention is a telephone number guide method of a medium scale (for example, a university, a specific company, a public office, etc.), and in the present invention, the telephone number guide of a university is described as an embodiment.

First, a sentence structure in which pronunciation is predicted when a user uses the system is set through a word network as shown in FIG. Also, among the sentences that can be made through the sentence structure described above, 10 sentences (sentence sample set for collecting voice data, as shown in FIG. 3) are constructed, and the voice data necessary for training is constructed. The data is stored in the training model unit 50.

In this state, if the continuous voice is input through the voice input unit, it is processed, and the processed data is compared with the training model to recognize the input continuous voice.

That is, as shown in Figure 4, the user attempts to dial to the system side to be guided by the telephone number, the ring signal arrives in the system (S10), the system side on the system side (for example, ' The system is an automatic telephone number guide system '), and asks for the name of the desired department and the professor (S20 ~ S30).

Thereafter, the system checks the presence or absence of a voice input (S40). At this time, if there is no voice input from the user within a set time (for example, 10 seconds), a termination message is transmitted and the voice recognition process and the phone number guide process are terminated ( S50, S90).

Unlike this, when a user inputs a voice, the input continuous voice is recognized through a continuous voice recognition process (S60).

Here, continuous speech recognition is implemented by the system as shown in FIG. 1, which will be described in more detail as follows.

First, the analog-to-digital converter 10 samples an input analog voice signal with a sampling clock of 8 KHz and quantizes the converted analog voice signal to convert it into a digital voice signal.

Since the phone line is limited to 300 ~ 3400Hz, the sample is converted to a digital signal by sampling at the minimum sampling frequency of 8kHz.

The converted digital voice signal is input to the endpoint detector 20, and the endpoint detector 20 extracts only a voice section from the converted digital signal using an energy value and a zero crossing rate value.

That is, the end point detection is a process of finding a section of speech only from the input signal when a word is input. This is very important in many areas of speech processing and has a significant impact on recognition performance, especially in the case of isolated word recognition. Since the end point detection is processed at all stages prior to signal conversion and processing of the voice signal, its inaccuracy may influence the recognition performance. In addition, it eliminates unnecessary signal information, providing the benefit of reduced computation.

The features of speech used for endpoint detection include Zero Crossing Rate, Energy, and Auto Correlation Coefficients. It is classified into an explicit, implicit, and hybrid method according to the connection relationship between the endpoint detection part and the recognition part.

The end point detection method applied in the present invention employs a zero crossing rate and energy among voice features, and adopts a method that can be independently adopted in front of a recognition part.

6 is a flowchart illustrating the endpoint detection process.

As shown, the digitally converted speech signal is pre-emphasized into blocks of 10 msec units. Pre-emphasis is characterized by better distinction between voice and noise. Each block is processed according to the current state to determine whether the block is voice or silent, and the state value is changed accordingly.

The first 10 blocks of data (100msec) are assumed to be the silence period, and the reference value is determined by calculating the statistical characteristics of energy and zero crossing rate in that interval. The reference value of energy is determined to be proportional to the average of each block energy value as shown in Equations 1 to 4 below. In this case, the offset is the minimum value set to prevent the reference value from becoming '0', and the value of THEL is to prevent the reference value from becoming too large.

The original meaning of the zero crossing rate refers to the number of times the signal is changed to negative and positive values based on 0, but this algorithm is defined as follows to reduce the influence of noise. First, divide SilenceU and SilenceL which are proportional to ThreshEL and divide signal value into 4 areas.

Count only the number of times that change as shown in 2. After calculating the zero crossing rate of each block, the reference value of the zero crossing rate is calculated as shown in Equation 7 below. THZCR is an experimental value to prevent it from becoming too small.

If the voice does not start and is silent, the reference value is recalculated using 10 blocks of data at 0.5 second intervals. By comparing the newly calculated value Th2 and the previous value Th1, a new reference value is determined as shown in Equation 8 below.

In addition, if the current state is noise, the energy is first compared to find the starting point of the voice.

8 is an explanatory diagram for explaining a process of detecting a start point and an end point by energy.

As shown in the drawing, the energy of the previous block and the energy of the current block are compared with the reference value in a reverse direction, and the first point above the ThreshEL before exceeding ThreshEL is used as the starting point of the voice by energy. Once the starting point due to energy is found, the zero crossing rate of the previous block is compared backwards to find the point where the threshold ThreshZCR is exceeded for the first time. This point is used as the start point of the voice and the end point is obtained in the same manner to determine one voice pulse section.

When a word is spoken, a number of voice pulses can be connected, and some of the pulses found may be noise.

1) certain voice intervals-6 or more blocks with energy above ThreshBig,

2) noise section-block length is 3 or less,

3) Uncertainty-Other.

The uncertainty section is divided into speech if a certain voice is followed within 400msec. If silence is continued for more than 500msec after the voice interval is set, it is regarded as the end point.

In addition, preemphasis is used to remove other factors affecting the vocal tract's resonance frequency and bandwidth, which have good results in modeling low amplitudes and high frequencies in the spectrum. Can be. In addition, the characteristics of the saints themselves are well reflected by removing the characteristics of the sound source and the radiation. Equation 9 below is a commonly used transfer function of the high frequency emphasis.

On the other hand, in order to analyze the voice signal, a stationary section must be assumed, so that the block is divided into small sections. These blocks are called frames, and distortion of frequency components occurs due to discontinuity of data at the end of the frame. This can be reduced by covering the window and attenuating the amplitude of the data at both ends of the frame. Generally, rectangular, triangular, hanning, and hamming windows are used. In the present invention, a hamming window is used. The voice signal covering the window is expressed as follows.

s (m + n): audio signal, w (m): window signal

Here is the Hamming window function:

A 25msec (analysis section, or 20msec) long Hamming window is moved by 10msec (moving section) to the high frequency stressed voice. 9 illustrates the use and overlapping of windows.

The best way to represent the characteristics of speech is by frequency analysis of the speech. Frequency analysis methods for speech feature extraction include the use of filter banks, LPC (Linear Predictive Coding) coefficient extraction, and Capstral Coefficients extraction. The LPC coefficient extraction method is the most common method that assumes speech utterance as a process based on a transfer function that models human vocal tracts, and it is known that the cepstrum coefficient has better recognition performance than the LPC coefficient. In the present invention, as a filter bank modeling a human auditory organ, Mel-Capstrum, which is known to be strong against ambient sleep and channel distortion in speech recognition, is used as a speech feature parameter.

Cognitive psychological studies have shown that the frequency scale at which humans perceive the sound, whether pure tone or voice, is not linear with the actual frequency. This study was carried out by experiments that define the subjective pitch of pure tones. The frequency scale that a subject feels subjectively is called mel, and it is distinguished from the physical frequency f, which is expressed in actual Hz. Other subjective pitch values are obtained by measuring the frequency of the tone felt at a point where the pitch of the tone corresponding to 1 kHz is doubled or halved. Figure 10 shows the subjective pitch scale expressed as a function of actual frequency. In Fig. 10, the upper curve is the relationship between the subjective pitch and the frequency in the linear scale, and the lower curve is the relationship between the subjective pitch and the frequency in the log scale.

Another subjective factor of recognizing the frequency of a signal is a critical band of frequency bandwidth called loudness. For a voice signal having a constant sound pressure, the band of the signal may be increased until the width of the critical band, that is, the loudness is increased, while maintaining the loudness of the band at a constant value. In other words, the human auditory organ analyzes the audio spectrum on a non-linear frequency scale. Mel-Busstrum uses this information to filter Fourier transformed speech signals at equal intervals on the mel scale. Analyze with The physical frequency and the mel frequency are given by the following relationship.

In other words, in the process of extracting the feature vector of speech recognition, the filter bank is non-linearly distributed by warping the frequency to a mel scale reflecting the frequency characteristic perceived by a subjective subject. It is called Mel Frequency Spectral Coefficients (MFSC). Fig. 11 shows triangular filter banks distributed in such a mel scale. In the present invention, a total of 24 triangular filter banks are used.

The elapsed value of the mel scale triangular threshold filter bank is taken, and the result is converted to DCT (Discrete Cosine Transform) to obtain Mel Frequency Cepstral Coefficients (MFCC). In the present invention, 13 MFCCs were obtained through DCT conversion from 24 logarithmic MFSCs. DCT relation is as follows.

Cepstral Mean Subtraction (CMS) is widely known as a method for compensating for distortion caused by a channel such as a telephone line. This method has a good result in improving speech recognition performance. CMS has the effect of high pass filtering the spectral coefficients.

4 out of channel-compensated MFCC, c _t [k] with c _t [0] as log power and c _t [1], c _t [2], ..., c _t [12] Get a set of features.

MFCC: c _t [1], c _t [2], ..., c _t [12] 12-dimensional vector

First derivative of MFCC: dc _t [1] dc _t [12] 12-dimensional vector

Second derivative of MFCC: ddc _t [1], ddc _t [2], ..., ddc _t [12] 12-dimensional vector

1st and 2nd derivative of log power: dc _t [0], ddc _t [0] 2 dimensional vector

Next, when the feature of the voice is extracted by the process as is well known, the Viterbi decoder 40 decodes the feature vector.

This process yields the optimal state sequence.

This Viterbi algorithm finds one state sequence that maximizes P (q | 0, λ).

Q denotes a state.

This equation represents the highest probability value in state s _i at time t along a sequence of states. By the cyclic method, the probability value at the next time t + 1 is expressed as follows.

The optimal state sequence can be obtained by the following procedure using the above equation.

1) Initialization

2) Repeat calculation

3) Termination

4) Find the optimal state sequence

In this process, the Viterbi decoding word is stored in the training model unit 50 as a reference pattern through the HMM modeling and training process.

The most common method used for speech recognition is by pattern matching. That is, it is divided into a training process of extracting the features of the speech patterns (words and phonemes) from the speech, and a recognition process of finding the most similar reference speech patterns by comparing the characteristics of the stored reference speech patterns when an unknown voice is input. Can be.

The HMM algorithm used in the present invention has been used in speech recognition algorithms since late 1970 as a pattern matching method. Recently, due to the high recognition rate and fast processing speed, it is widely used in a large capacity speech recognition system. The basic idea of the HMM algorithm is to recognize that the speech can be modeled as a Markov model by obtaining the parameters of the Markov model during training, creating a reference Markov model, and finding the most similar reference Markov model in the recognition process.

As Markov model, Hidden Markov model is applied to accommodate various changes of speech pattern. The Hidden Markov Model is a dual stochastic process consisting of a stochastic process for state selection and a stochastic process for output probability of generating a speech pattern in each state. That is, each feature of the voice pattern is expressed by the selection probability of the state and the output probability. Hidden here means that the state is hidden in the model regardless of the speech pattern. The HMM algorithm can be set to a pattern having a pronunciation length of less than a word, such as a phoneme or a syllable. The HMM algorithm is mainly used in a large-capacity speech recognition system because a word or a sentence can be input as an input voice.

The model (λ) notation using the HMM can be expressed as a transition probability (A) between states, an output probability (B) dependent on states, and an initial state existence probability (π) of states. It is also assumed that the change of state must follow the Markov process and the output of the observed symbols are independent of each other.

Variables used in HMM are as follows.

(1) N: number of states in the model

(2) M: number of observation columns in the state

(3) Probability of transition between states

(4) output probability dependent on state

(5) Probability of initial state

The model is defined as λ = (A, B, π).

To apply HMM to speech recognition, the following three problems must be solved.

1) The problem of calculating the probability P (O | λ) that a sequence will occur when a sequence of observed symbols is obtained.

This problem is necessary when using the Baum-Welch algorithm during training, but does not need to be calculated when using the segmental K-mean algorithm.

2) Calculation Problem of Optimal Sequence.

The second problem is to predict the sequence of hidden states by using HMM. This is calculated using the Viterbi algorithm. In the present invention, the Viterbi algorithm is applied in the training and recognition part, and the process of obtaining the optimal state sequence is required in the training process, and in the recognition part, the input word against the candidate word model is scored using the Viterbi algorithm. Used for

3) Reestimation problem of the model to maximize P (0 | λ).

The third problem is the method of reestimating the model to best match the observed value, but no analytic method is known. Currently, the most commonly used methods include the Baum-Welch reestimation method and the segmental K-means method. In the present invention, the segmental K-means method is used.

On the other hand, the output probability B in the HMM has a discrete probability distribution (DHMM) or a continuous probability distribution (CHM, SCHMM). Since the DHMM uses VQ to divide the acoustic space, the computation is reduced, but the recognition rate is lowered because of distortion errors. Therefore, the present invention uses CHMM.

CHMM represents the output probability distribution by mixing M Gaussians and is expressed as follows.

In the above, M is the number of Gaussian mixtures in the state, C _im is the mixing gain of the Mth mixture, N (O _t , μ _im , _im is the Gaussian probability distribution function for the observation vector (O _t ), O _t is the observation vector, μ _im is the mean vector, _im denotes the covariance matrix and K denotes the order of the observed vectors. 12 is a waveform illustrating a mixed Gaussian probability distribution function.

When using the CHMM the number of parameters to be estimated from observations is very important. One way to reduce the number of parameters when representing a continuous probability distribution using a Gaussian distribution is to use a geometrical covariance matrix. In other words, all off-terms are treated as '0'. This reduces the amount of computation and reduces the number of parameters that need to be estimated.

Next, there are a beam-welch algorithm and a segmental k-means algorithm for estimating the parameters of the HMM by ML criterion.

The Beam-Welch algorithm finds a model that maximizes the observation sequence given a model, while the Segmental K-means algorithm finds a model that maximizes the observation sequence and the optimal state sequence given a model.

That is, the Beam-Welch algorithm finds λ that max P (O | λ), while the Segmental K-means algorithm obtains λ that maximizes max _s P (O, s | λ).

The reason for using max _s P (O, s | λ) is, firstly, that it takes a lot of computation and time since P (O | λ) takes into account all state transition paths when looking for likelihood. . Second, since the output probability B of the parameter has a very large dynamic range, the calculation of likelihood for all state transition paths may encounter numerical difficulties. Third, the modeling and decoding are all P _s max (O, s | λ) The modeled by using a max _s P (O | λ) and using the decoded P _s max (O, s | λ) that uses Baum- More natural than the Welch method.

The method of estimating the parameters using the segmental K_means algorithm is to find a model that satisfies the following condition.

To solve this, the EM algorithm is used.

The first step is to find the optimal state sequence that maximizes the observed sequence given the model.

The second step is to estimate the new model using the optimal state sequence obtained above.

The segmental k-means algorithm uses the Markov chain k-mean method to estimate the parameters of the HMM. Segmentation information is obtained through the Viterbi decoding process.

In fact, the parameter estimation method using the segmental k-means algorithm is as follows.

1) Initialization

All training vectors are evenly segmented into phoneme units and HMM state. By clustering the segmented segments of each phoneme, the parameters ( ) Is initialized.

2) segmentation

The parameters of CHMM obtained in process 1 or process 3 are re-segmented the training vectors by phoneme and HMM state through Viterbi decoding. The transition probabilities can be obtained from partitioning information obtained from Viterbi decoding.

In the above, s ^* is an optimal state sequence, and n _ij (s ^* ) represents the number of transitions from state i to state j in the optimal state sequence.

3) Clustering and Estimation

All training vectors corresponding to each state of the phoneme model are divided into M clusters by vector quantization (VQ). And using the training vectors divided into each cluster of state s _i , the parameters ( Estimate).

Here, each of the above parameters is obtained by the following equation.

Where V _im is the set of vectors corresponding to the mth blend of state s _i , and L _im is the number of vectors corresponding to V _im .

4) Repetition

Repeat steps 2 and 3 until the convergence condition is met.

On the other hand, the model used in the present invention is a context-independent phoneme model. Context-independent means that when creating a phoneme model, we will not consider the circumstances before and after the phoneme, which reduces the number of models. The number of context-independent phonemes adopted is 53 and the phoneme model is the left-to-right phoneme model. The adopted context-independent 53 phonemes are shown in FIG.

Next, the training process using the HMM consists of an input, initialization of the probability of transition, reestimation, and the construction of the pronunciation dictionary.

First, the input step receives a voice in sentence units, and when a sentence is input, receives a list of input files of a type in which a pronunciation dictionary indicating which phonemes are composed of the sentence is attached to the file representing the word. It is.

Therefore, it is possible to separate the sentence from the phoneme.

A total of four inputs are used for the 12-dimensional vector MFCC, the first derivative of the MFCC, the second derivative of the MFCC, and the first and second derivatives of the log power of the two-dimensional vector.

The next step is to uniformly initialize the transition probability of the HMM.

Next, the re-estimation re-estimates the parameters of the HMM using the segmental k-means algorithm described previously. The construction of the pronunciation dictionary uses 53 phoneme models when constructing the sentence model, and the notation is written out phonetically. . It also adds silence models at the beginning, end, and middle of the sentence.

The training model that has undergone the training process for continuous sentence recognition in the same manner as the well-known process is stored in the training model unit 50, and then through the voice recognition unit 60 and the phone number guide unit 70 of FIGS. In the same way, the telephone number guide service according to the actual continuous speech recognition is performed.

That is, as shown in Fig. 4, when the inquirer off-hooks the telephone and dials the telephone number of the inquiries in order to know the professor's telephone number of the corresponding department, the telephone number guide system is in the off-hook state (S10). When the phone number guide system is off-hook, the system sends out a system introduction comment (for example, 'this phone number guide system is a system for guiding the phone numbers of professors at 00 university') (S20). It sends out a voice request to enter the name of the desired department and the name of the professor (for example, 'enter the name of the desired professor and professor') (S30). In this state, if the user inputs the voice within a predetermined time, the process returns to the continuous voice recognition step (S60). Otherwise, if the user does not input the voice within the predetermined time, it is checked whether the set time has elapsed (S50). If the setting time has not elapsed as a result of the check, the process returns to the step S30 once again and sends out a comment for inputting the desired subject name and the professor's name. If the setting time has elapsed, the end comment (for example, , 'Please start again from the beginning', or 'I will end the telephone number guide service'), and ends the telephone number guide service (S90).

The continuous speech recognition in the case where the user inputs the desired subject name and professor's name within a predetermined time is as follows.

First, as is well known, a sentence structure in which a user's pronunciation is predicted is set (S61), and a sentence that can be made from the sentence structure is set (S62). Next, among the set sentences, 10 sample sentences (10 or more may be 10 or less, but in the present invention, 10 is described as an embodiment), voice data necessary for training is constructed (S63). Then, the constructed speech data is stored in the training model unit 50 as a training model, and when a voice is input from the user, the continuous voice is recognized by comparing the input voice with the stored training model (S65). That is, the voice input by the user is transferred to the voice recognition unit 60 through the analog / digital converter 10, the endpoint detector 20, the feature extractor 30, and the Viterbi decoder 40 in sequence. The voice recognition unit 60 recognizes the input voice by comparing the input voice with the reference model stored in the training model unit 50 and transmits the recognized result to the telephone number guide unit 70 (S70). ). Accordingly, the phone number guide unit 70 transmits the phone number of the corresponding professor according to the recognition result by voice. Thereafter, the user sends an audit comment about the use (for example, 'thank you for using') (S80), and sends out the termination comment, and ends the telephone number guide service (S90).

As described above, the present invention has the advantage of automatically guiding a phone number in a medium scale (university, hospital, government office, etc.) with continuous speech recognition in Korean, and also provides a more natural pronunciation to the user with continuous speech recognition. In addition, there is an advantage that can be induced, and also has the effect of making the service as convenient as possible free from the constraints (conditions restricting the user's voice input to input only pre-stored words) generated in the existing speech recognition application.

Claims (3)

In the phone number guide method through voice recognition, When the user dials and the telephone number guide system is in an off-hook state, outputting a system introduction comment and sending a comment requesting voice input of a desired class and professor's name; Recognizing a continuous voice by comparing a training model trained with a sentence that is set in advance and predicted when the user inputs a name of a corresponding department and professor after the voice input request; Transmitting the recognition result of the recognized input voice to a telephone number guide to control the telephone number of the corresponding professor to be transmitted; The phone number guide method using the continuous speech recognition, characterized in that the step comprising the step of transmitting the termination after the phone number of the corresponding professor. The method of claim 1, further comprising: checking whether a set time has elapsed if a voice is not input within a predetermined time while requesting voice input of the desired department and professor name; Resending a comment requesting entry of the desired department and professor's name if the set time has not elapsed; And returning to the step of transmitting the termination message when a set time has elapsed in the absence of the voice input. The method of claim 1, wherein the recognizing of the voice comprises: setting a sentence structure in which the user's pronunciation is predicted, setting a sentence that may be made from the set sentence structure, and among the set sentences. Constructing voice data necessary for training with 10 sample sentences; storing the constructed voice data as a training model in a training model unit; and inputting an input voice and a stored training model when a voice is input from the user. Telephone number guidance method using continuous speech recognition, characterized in that the step consisting of comparing the continuous speech recognition.