JPS62111296A - Voice recognition method and apparatus - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a method and device for recognizing speech based on a predetermined word base, and further relates to a method and device for training a speech recognition device.

(Prior art) Research on speech recognition that has been done so far is based on [hidden Markov model J].
A mathematical model called ``model'' is used.

In this study, a mathematical model is established for each word to be recognized, and the parameters of the mathematical model are defined by using a suitable training procedure.

When an unknown word is spoken, its observations are made in some predetermined way. In this case, for each given word model, the maximum posterior probability that would cause that observation can be calculated. Then, the word corresponding to the model that gives the highest probability can be recognized as the most likely word corresponding to the unknown word. This concept can be extended to the recognition of sequences of words spoken in a related manner.

A Markov model is a finite state machine with a set of states. For example, see the four-state model in FIG. This Markov model can be in any one state at any time. At the end of each time interval, the model transitions from the current state to another state or to the current state itself. If a model only stays in one state, it is impossible to predict with certainty what state it will be in next. Rather,
For all possible combinations of transitions from one state to another, or to the current state itself, the transition probabilities are known.

At individual time intervals, the model produces observables.

This observable is typically a multidimensional measurable. Hidden Markov models do not have enough knowledge about the state to make reliable predictions about what the observed results will be. Instead, the state determines the prior probability distribution of the hypothesis measurements that the model will take in that state.

(Problems to be Solved by the Invention) The above-mentioned speech recognition research has the following drawbacks. That is, when the recognition device receives a sample of sounds that constitute the speech to be recognized, it is necessary to perform a large number of calculations in a very short period of time. This problem is particularly serious when recognizing continuous speech. In general, the requirement to perform a large number of calculations in a short period of time has made recognition equipment very expensive.
This limits many applications, such as personal computers and voice (eg, telephone) data entry.

(Means for Solving the Problems) According to a first feature of the present invention, there is provided an apparatus for recognizing words within a predetermined linguistic range, which has the following components.

(b) Means for sequentially sampling sounds to obtain a set of signals each representative of the characteristics of those sounds.

(b) Means for storing data representative of a number of finite state machines for each word of a given word work and for forming individual models of those words.

the data describes states forming a finite state machine;
A probability function is assigned to each state. In this case, at least one probability function is assigned to two or more states.

Each probability function describes the probability that a signal representing a characteristic of the sound will pretend to be some observed value if the model generates the actual sound and if every model is in the state to which the probability function is assigned. do.

(c) Calculate the probability that a given set of signals will occur if the model that produced the actual sound and any model are in any given state. A means of determining from a combination.

(d) A means of determining, from calculated probabilities and properties of finite state machines, the maximum likelihood that a sample of consecutive sounds will represent given words.

(e) Means for providing an output indicating one of the predetermined words as the most likely spoken word based on the maximum probability detected.

According to a second feature of the invention, there is provided a method for recognizing words within a given language, comprising the following steps.

(a) Sequential sampling of sounds to obtain a set of signals each representative of the characteristics of those sounds.

(b) Storing data representative of a number of finite state machines for each word of a given strategy and forming individual models of those words.

the data describes states forming a finite state machine;
A probability function is assigned to each state. In this case, at least one probability function is assigned to two or more states.

Each probability function describes the probability that a signal representing a characteristic of the sound will pretend to be some observed value if the model generates the actual sound and if every model is in the state to which the probability function is assigned. do.

(c) If the model that generated the actual sound and any model were in any given state, the probability that a given set of signals would be generated is calculated using the signal and the probability function. A means of determining from a combination.

(d) Determining, from the calculated probabilities and the properties of the finite state machine, the maximum likelihood that a sample of consecutive sounds will represent the given words.

(e) providing an output indicating one of the predetermined words as the most likely word spoken based on the maximum likelihood detected;

The main advantages of the invention are as follows. In other words, the number of probability functions can be made very small. For example, 1024 probability functions were required for a common word containing 64 words (i.e., 64 models for each of the 16 states), whereas 100 probability functions were required. It's over. This is because most probability functions are usually assigned to a large number of states, and therefore to a large number of finite state machines. Therefore, the number of calculations W, the number of required periodic operations, and the storage capacity used can be reduced, and the cost can also be reduced.

Finite state machines are also known as hidden Markov models.

A means to determine the maximum likelihood is the Viterbi algorithm (S.E. Levinson.

L, R, Rato i ner, M, M, 5oudh
i, “Introduction to the application of probability function theory of Markov processes in automatic speech recognition J, Be1l SystemTechn
ical Journal, Vol. 62, No. 4, Part 1. April 1983, pp. 1035-1074) or the "forward-backward" algorithm (see S. E. Levinson et al., supra). It is possible to have the means to do so.

According to a third feature of the present invention, there is provided an apparatus for recognizing words within a predetermined range of words, which has the following components.

(b) Means for sequentially sampling sounds to obtain a set of signals each representative of the characteristics of those sounds.

(b) Means for storing data representative of a number of finite state machines for each word of a given 3R business and forming individual models of those words.

the data describes states forming a finite state machine;
A probability function is assigned to each state. In this case, at least one probability function is assigned to two or more states.

Each probability function describes the probability that a signal representing a characteristic of the sound will pretend to be some observed value if the model generates the actual sound and if every model is in the state to which the probability function is assigned. do.

(c) Means for calculating a distance vector from the probability function and each set of these signals when an open signal occurs, and means for generating a control signal each time the vector is calculated. Each vector represents the probability that any given set of signals would occur if the model produced an actual sound and if any model were in any given state.

(d) logic circuit means specially constructed to receive one of the control signals and use the current distance vector to determine a value representing the minimum cumulative distance of each of the finite states;

(e) Means for providing an output indicating one of the predetermined words as the most likely spoken word based on the calculated minimum cumulative distance.

The present invention can reduce the time required for calculations and the complexity of calculations using specially designed hardware logic circuit means for calculating cumulative distances.

In the hidden Markov model, the finite state machine is
It can have multiple states connected by transitions. Each state and each transition has a corresponding signal set probability function and transition penalty, respectively. A state penalty can then be calculated for any set of observed signals, and a probability function can be expressed as the inverse of the probability of said set of observed signals. Each finite state machine represents a word with a word guide to be recognized. And the migration penalty is
It depends on the characteristics of the word. However, the state Benarti
It depends on the probability function and the sound currently being received.

The means for storing data may store a minimum cumulative distance for each machine. The logic circuit means are adapted to determine for each transition of the current state a value that depends on the minimum cumulative distance of the starting state of that transition and the transition penalty, and to determine the minimum value of that value. The present invention can be configured such that the minimum cumulative distance of the current state can be determined from the minimum value of the values and the state penalty for the current state. Each minimum cumulative distance can in fact be identified as indicating the maximum probability that a sequence of sounds forming part of a certain word in the vocabulary was generated.

Apart from the logic circuit means, one or more of the other means mentioned above may be implemented in a general purpose computer. Preferably, it is based on a microprocessor and is particularly permanently programmed for the necessary operations.

According to a fourth aspect of the invention, there is provided an apparatus for analyzing data relating to the presence of one of a number of conditions, the apparatus having the following components:

(b) A means of receiving and receiving a set of signals representing the data to be analyzed.

(b) A means of storing data representing a large number of finite state machines and modeling each of the conditions individually. This means
It has a number of functions, at least one of which is assigned to two or more states of the finite state machine. Also, if a model generates a real signal and any model is in the state to which its finite state is assigned, then each function is the probability that a signal representing the I quality of the data pretends to be an arbitrary observed value. Describe.

(c) When data is generated, from the function and data pair, if the model generates the actual signal and any model is in any given state, then any given signal A means of determining an indicator that represents the probability that a pair will occur.

(d) Using the above marks, each □ of each finite state machine
Logic circuit means configured to determine a value of a minimum cumulative distance of a state.

The means to sequentially sample sound is analog, digital, or analog via an amplifier with a wide dynamic range.
It can include a microphone connected to the converter. In general, these means can include, for example, a bank of digital filters to analyze and separate the incoming signal into frequency spectra. The intensities of this spectrum form each of the signal sets. The above-mentioned means can also be used for linear prediction (J, D.

Markel, A.H., Gray, [Linear Prediction of Speech J, Spr incler-Ver I ag, 1
976).

Preferably, the means for calculating the probability includes means for calculating, at each frame (ie a predetermined number of sample periods), a distance vector having a number of elements. Each of the elements is calculated from a probability function of one state and one of the signal sets.

The apparatus and method according to the first four aspects of the invention typically require a word model defined by machine states and transition probabilities. And the open state requires details of the probability function. These states and functions define words in a given K vocabulary as “
obtained through training. This vocabulary can be changed by retraining if necessary.

According to a fifth feature of the invention, therefore, selecting a number of states for a finite state machine to represent words within a predetermined vocabulary and obtaining data for characterizing the states. A method is provided having the following steps. .

(b) Sequentially sampling the sounds forming each car gR in the word cliche to obtain a set of signals representative of the characteristics of the sounds forming the word.

(b) Obtaining data defining the state for each word from the set of signals.

(c) Obtaining one probability function for each state from the set of signals.

(d) merging several probability functions to obtain a number of said functions less than a predetermined number, the merging comprising:
Steps carried out in accordance with regulations relating to suitability for annexation and a predetermined number.

(e) Calculating the probability of transition from each state to all allowed subsequent states in each finite state machine.

(f) Entering data that identifies each word when it is spoken.

(g) Data that determines the conditions for forming each machine,
and storing data identifying the words represented by the machine.

(H) Storing data defining each merged probability function.

The invention also includes an apparatus for carrying out the method of its fifth aspect.

Preferably, the step of obtaining state-defining data includes determining the signal for each word according to criteria related to merging suitability and maximum and minimum numbers of states for that word, to provide state-defining data for each word. merging some of the subsequent sets.

The merging suitability of two successive sets of signals can be pre-evaluated by calculating the logarithm of the ratio of the likelihood that each set of signals for potential merging arose from a separate set. If this logarithm exceeds a threshold, then the sets are merged. The threshold value can be predetermined experimentally or obtained using some test statistic. Similarly, the merging suitability of two probability functions is the logarithm of the ratio of the probability that each probability function for potential merging arises from the same probability function to the probability that each probability function arises from a separate probability function. It can be evaluated in advance by calculation. If this logarithm exceeds a threshold obtained in the same way as the other thresholds, then these functions are merged.

Preferably, the step of obtaining the set of signals includes the known process of dynamic time warping.

The device according to the invention can be configured for training and/or recognition.

Typically, this device is formed by one or more mask-programmed logic arrays and one or more microprocessors. Such an array or processor forms a plurality of the above-mentioned means.

According to further features of the invention, there is provided a speech recognition device having the following components.

(b) A means of sampling sound.

(b) Means for storing data representing a number of finite state machines corresponding to words to be recognized.

(c) means for determining, from the output of the sampling means and the stored data, the probability that a series of sound samples will represent said word;

The invention also includes methods that accommodate this feature.

Additionally, an apparatus and method is included for determining stored data for this feature from known spoken words.

(Example) Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

First, we will explain how to obtain the parameters of the hidden Markov model of a word from an example of when the word is spoken during training.

In FIG. 1a, in operation 10 one of the words included in the vocabulary is input into the training device.

Briefly, the training device may include a microphone, an amplifier with appropriate dynamic gain, an analog-to-digital converter, and a bank of digital filters. The words to be entered are spoken into a microphone. Next, perform operation 11. Here, one feature vector is formed for each frame period. One frame has, for example, 128 sampling periods of the analog-to-digital converter. Each feature vector is one (for example, one
0 elements).

This element corresponds to an intensity in each frequency band spread over the audio frequency band, for example from 0 to 4.8 kHz. Each feature vector can also be obtained using a bank of filters and means for extracting energy values at the filter outputs. Therefore, one feature vector can be expressed as follows.

[xl, ×2)X 3 X'F] T
Here, ■ means conversion to a more ordinary column vector, ie, a vertical vector. When there are 10 elements, [xl, x2) x3 ``10]■.

In general, even if the same words are pronounced, their pronunciations are not the same. The important reason for this difference is
'The time scale of one pronunciation is nonlinearly distorted compared to the time scale of another pronunciation. For this reason, we carry out a process of dynamic time stretching (
1 work 12). This operation matches the two timescales by matching the feature vector of the most recently spoken pronunciation of the word to the feature vector average of the composite template. The feature vector average of this composite template is
Unless it is the first pronunciation, it is already derived from a previous pronunciation of the word, as described below.

Dynamic time warping is a known process. This step can be implemented using dynamic programming algorithms and will therefore only be briefly described here. Actually,
The most recently spoken pronunciation feature vectors are divided into groups of one or more vectors. Each group corresponds to one template vector. The correspondence is obtained such that the resulting total spectral distance between the word vector and the corresponding template vector is minimized.

A template vector is formed in operation 13 by incorporating each group of feature vectors into a corresponding template vector according to the match obtained by dynamic time warping. Each template vector has F elements (10 elements in the previous example). Each of these elements is formed by weighted averaging of each element of the template vector with the elements in the group of feature vectors corresponding to the template vector. Therefore,
The template vector is expressed as a feature vector average as follows.

[xxx--1-maF] T 1, 2, 3 Furthermore, each template vector includes various elements. This is calculated using the following formula.

Here, i takes a value from 1 to F, σi is the variance of the set of feature vectors used and informs its template vector, and is the mean square of the - set of feature vectors, σmin is a positive constant.

Roughly speaking, a template vector is related to the number N of feature vectors that are merged into it.

That is, this number N is the sum of the number of corresponding feature vectors in each group obtained by time warping for all repetitions of a certain word. Each time operation 13 is performed, the mean value, mean square, and number of feature vectors are updated. First, when a certain word is input, operation 11
A composite template is directly formed from the lucky feature vectors obtained in . The feature vector average of each template vector is
Set equal to the corresponding feature vector obtained from the spoken word. The number N of feature vectors is set to 1.

When a predetermined number (for example, 10) of voice samples are input for a certain word, a decision 14 prevents further repetition of operation 10. As a result, the composite template provides data that defines the state of the S-in-law Markov model that models the word. Each of the template vectors can be seen as representing one model state. However, in this case, there are many more states than required, and each state is not fully characterized by one template vector.

The next group of operations in FIG. 1a are for making the number of states smaller.

The first step is to divide the template. The quantity λ is calculated g1 in operation 15 for each of the pair of subsequent template vectors. Pairs of template vectors are overlapping in the sense that each template vector appears in two pairs except for the first and last template vector.

Therefore, λ is n and n- for n-2 to nm.
Calculated for every pair of 1's. Here, nm is
is the number of template vectors in that template. n and n
Let λ(n> be the value of λ calculated for a pair of model vectors of Suppose that the probability that the observed value is Ls (in this case, the observed quantity is a feature vector that contributes to the template vector), and that the resulting observed quantity is generated from two different most likely distributions, Therefore, λ is a measure of how well two states are suitable for merging.Assuming a Gaussian multivariate distribution with a diagonal variance matrix, we find that:

Here, subscripts 1 and 2 relate to two consecutive template vectors. The subscript f relates to the fth feature of one feature vector. N is the number of feature vectors merged into the template vector. The subscript y is applied to the virtual template vector that would be formed if two consecutive template vectors were merged into one.

The correction value λ-(n) is calculated from n-2 to n-nm in operation 15.
It is calculated using the following formula.

λ-(n>-2λ(n)-λ(rl-1>-λ(n+1
) λ(1) and λ(n, 71> are set to arbitrarily large values. Values of λ−(n> The second difference is used (to indicate the location of the split), which we will now discuss.

The template vectors for one model are stored sequentially. A number of split indicators are then entered between the template vectors to indicate how the template vectors are to be merged. In this case, no marker is placed next to the template vector to be merged. Once a certain number of indicators have been inserted, the number of final states in the model is equal to the number of indicators plus one. Therefore, judgment 16 is performed and the number of states determined so far is the maximum value (for example, 8
) to determine whether it is less than. If so,
Decision 17 is performed to determine whether the maximum value of λ'(λ') not yet used to insert a split indicator is ax less than the threshold. If it is small, no splitting is necessary. If λ-max is greater than the threshold, operation 18 inserts a new split indicator at a position between the brochure ax type vectors corresponding to that value of λ-.

Then, repeat decision 16. However, in decision 17,
If λ' is less than the threshold ax , then decision 19 is performed to determine whether the number of states defined by the indicator is less than or equal to the minimum value. If it is small, it is necessary to roughly insert the splitting mark even if λ-max is smaller than the threshold. Therefore, one day of operation is carried out. This loop (ie, decision 16.17.19 and operation 18) continues until the number of states is less than or equal to the minimum value (eg, 3).

If, in decision 19, the minimum number of states criterion is satisfied before or after inserting the indicator corresponding to λ' that is less than the threshold, then in operation 20 of FIG.
Perform merging of template vectors.

To merge template vectors, the average value of the feature vector mean and mean square is determined, taking into account the total number of feature vectors that give rise to each template vector to be merged. Each merged vector created in this way, together with its standard deviation, defines a probability density function (PDF). however,
Assume that the probabilities follow a Gaussian distribution. However, the probability density function formed in this way is provisional. 1, since the next step in FIG. 1b is to merge the temporary probability density function with the probability density functions already stored for the other words (if any). If it is not appropriate, roughly memorize the probability density function. Before storing the probability density function, in operation 20, a tentative probability density function is stored as a tentative model of the word.

If the word currently used for training is not the first word in the predetermined vocabulary, when the temporary probability density function is memorized, in step 22, the probability density function of each flight and each memorized Calculate one λ between the probability density functions.

Then, training is started for the next single gl in the predetermined vocabulary.

Since the temporary probability density function and the stored probability density function are stored in the same format as a template vector, the calculation of λ is performed in the same way as when the template vectors are merged.

After completing operation 22, operation 23 selects the minimum value of λ for one probability density function. Then, in decision 24, a comparison is made with a threshold value indicating whether merging is desirable.

If the threshold indicates that it should be merged, then step 2
5, the tentative probability density function and the stored probability density function corresponding to λmax are merged by the averaging step described above. Then, in operation 26, the merged probability density function (now,
A new value of λ is calculated between the stored value) and the probability density function of each remaining flight.

In decision 24, if the maximum value of λ is greater than the threshold,
Decision 31 is performed to determine whether the number of probability density functions already stored is smaller than the maximum allowed f[ (for example, 48). And, if so, in operation 32 a temporary probability density function with the current maximum value of λ under test is stored. However, as shown in decision 31, if no more probability density functions can be stored, the probability density functions are merged even if the value of the provisional probability density function is greater than the threshold. There must be. Therefore, operation 25 will be performed.

Each probability density function is merged in operation 25, so operation 29
, the model for the current word is strengthened by using a label indicating the merged probability density function instead of the hypothetical probability density function.

Decision 27 is then performed to determine whether there are any more tentative probability density functions that should be considered for merging. If it exists, return to operation 23.

If there are no more provisional probability density functions to be merged, then in operation 28 the probability of transitioning from one state to another or returning to the same state is calculated (arrows 36 and 37 in Figure 3). ). The probability of transitioning from one state to the next can be calculated by dividing the number of pronunciation exemplars used to train the word by the total number of feature vectors merged into the first mentioned state. . The probability of transitioning from that state to itself is
It is obtained by subtracting the probability thus obtained from 1.

Finally, in operation 30, the transition probabilities are stored together with the stored probability density function labels. This forms a complete model of the word. The steps of Figures 1a and 1b are repeated for all words. As a result, additions are made to the stored model and the stored probability density function is modified. However, since the stored probability density function is stored in the model by the label, no further processing of the completed model is required.

. The flowcharts of FIGS. 1a and 1b can be implemented by programming a computer, such as a microprocessor. Such programming is a well-known technique and will not be described here. If the processing required to pronounce each word is required to be carried out during training in approximately the same amount of time that the word is spoken, then at least in part a specially designed computer can be used. It will be necessary to use it. An example of this type of computer is discussed herein in connection with a recognition system that uses a trained model.

For each word (e.g., 64 words) included in a given vocabulary to be recognized, a respective model can be obtained by repeating the word (e.g., repeating it 10 times into the microphone of the training device). can. The training device is in combination with a data input device such as a keyboard. This allows you to write and input each word as you repeat it during training, and store it as a label for that model.

Next, speech recognition using the Tomo S3 Markov model obtained through training will be described.

Here we will use the same analog circuit that we used in training. The output of the analog-to-digital converter is then separated into the frequency domain using digital filtering.

In each frame period, the filter output is first formed as a single feature vector in operation 42 (FIG. 2). This feature vector is created by extracting each probability density function obtained during training and using it together with the elements of the feature vector, as described above, so that the speech sample at that time is generated by the state corresponding to this probability density function. The probability is calculated (operation 43). Each such calculation yields a distance vector [d i , d 2 . d
a.

...dm]T is obtained.

The meaning of "distance" will be explained below.

Here, m is the total number of probability density functions, which is 48 in the above example. The distance vector obtained in this manner is stored in operation 44. In a -dimensional Gaussian distribution,
The probability that an arbitrary model state produces an observable equivalent to a certain feature vector is given by the following equation.

Here, σ is the standard WA difference, X is the observed amount, and ★ is the average observed amount.

Suppose a distribution has J dimensions. In the speech recognition method of the present invention, J corresponds to the number of elements included in the feature vector. At this time, the above probability becomes as follows.

In the above formula, ■ means the product obtained by multiplying the terms obtained by changing j according to a known mathematical convention. The values for xj are obtained from the feature vector, and the values for xj and Oj are obtained from the template vector l-. This probability (herein referred to as the "distance" of probability)
By taking the natural logarithm of (known as ) and multiplying by 12, we get the following equation:

Here, Jln(2π) is omitted. I mean,
This is because it is a constant and is not needed to compare relative "distances". Each element of the resulting distance vector provides a measure of the degree to which the speech sample is unlikely to be given the corresponding probability density function. Each of these elements is then calculated as having a corresponding probability density function proportional to the logarithm of the reciprocal of the probability of a given speech sample.

By using ``distance'', the number of calculations and divisions can be reduced. This can speed up recognition and reduce the cost of special purpose circuitry to implement the flowchart.

After obtaining this distance vector, which gives the probability that the current pronunciation is true from each of the possible model states, each model in turn calculates the probability that the pronounced word falls under that model. considered in determining the probability that the Thus, in operation 45, a model is selected, and in operation 46, the minimum distance for each state is determined in that model. This process will be described in detail below with reference to FIGS. 4 to 10.

The process described above continues until each model has been processed, as indicated by decision 47. Operation 48 is then performed. Here, for all states of all models,
Find the smallest of the minimum distances. This value is
In operation 49, it is used to normalize all of the minimum cumulative distances found. (Also regarding operations 48 and 49,
This will be explained with reference to the figures and FIG. ) The normalized distances are then stored in operation 50 as a cumulative distance vector. Therefore, this vector is the array of elements D1
．． D2. D3. ...has Dn.

Here, n is the maximum number of states in the model. One column corresponds to each model.

The smallest of the minimum cumulative distances is also stored. Next, perform operation 51 to minimize. From the last state of each model with , determine and provide its corresponding word as the most likely word that would have been spoken in the current sampling period.

Such a method is described in the following literature:
It can be easily extended to recognizing combined words. Namely, J.S., Br1dle, M.

D. Brown, R.M., Chamberlain [Algorithm for Combined Word Recognition J JS
Please refer to RLI Research Report 1010. here,
A route is determined through the model that shows the minimum cumulative distance, and a traceback pointer indicating this route is used.

One way the traceback pointer can be determined is
This will be explained below. The following two papers also describe suitable methods for recognizing combined words, using minimum cumulative distance and traceback pointers.

``Partial Traceback and Dynamic Programming'' JP, F. Brown, J. C., Sp.

hrer, P. H., Hocksch, J.

K. Baker, IEEE on Acoustic Speech and Signal Processing.
Proceedings of the International Conference, pages 1629-1632, 1982
Year. and “Using a Single-Stage Dynamic Programming Algorithm for Combined Word Recognition JHerm
ann Net, IEE on Acoustic Speech and Signal Processing
E-newsletter, ASSP-32, No. 2@, April 1984,
Pages 263-271.

The process of FIG. 2 is repeated for each frame period. For this purpose, a high speed specialized computer is usually required. This type of machine will now be described.

A device that can be used for both training and speech recognition is shown in FIG. Here, microphone 60 is coupled via an amplifier 61 to the input of an analog-to-digital converter 62. Amplifier 60 has automatic gain control and has a wide dynamic range. However, this amplifier and microphone 60 will not be described in detail. This is because suitable circuits for this part of the device are well known. The processor 63 is
For example, it may be an Intel 8051 type microprocessor and provides a sampling frequency of 8 kHz to converter 62 and front end processor 64 via line 71. Therefore, each cycle of the process shown in FIG. 2 completes in 125 microseconds. Processor 64 performs the digital filter operation 40 described above.
is carried out, and the passband is 0 kHz and 1 kHz in arithmetic terms.
kHz, preferably logarithmically divided between 1 kHz and 4.6 kHz. Therefore, the feature vector is
The feature vectors can be obtained on an eight bus 65 where the feature vectors are sent to a logic circuit 66. The logic circuit 66 is
Can be a mask programmed array. The initial connection procedure between processor 64 and logic circuit 66 is controlled on line 70.

Processor 63, logic circuit 66, random access
Memory (RAM>67 is connected to the address and data bus 68.
69, they are connected to each other.

The host computer 72 has a bus 73 and a control line 74.
is connected to the processor 63 by.

It is anticipated that components 62 through 67 will form a device that is manufactured and sold as a peripheral device for use with a host computer. This allows the host computer 72 to provide input facilities during the training period (where spoken words must be identified) and to provide a display displaying the words to be recognized. If the device is used solely for recognition, RAM 67 may be partially replaced with read-only memory and host computer 72 may be replaced with a simple or intelligent display for displaying recognized words. . Logic circuit 66 implements the Piterbi algorithm.

Therefore, operations equivalent to operations 45-48 are performed to determine the minimum cumulative distance. The remaining operations shown in FIGS. 1a and 1b are performed by processor 63. RAM67 is
Operation 29 including the composite model of operation 13 and the probability density function
The model is memorized.

The RAM 67 also stores the distance vector of operation 44 and the distance vector of operation 5.
A cumulative distance vector of 0 is stored.

Programming these processors is a known technique and is programmable given the algorithms of FIGS. 1a and 1b, and will not be described further. However, one way in which operations 46, 48, 49 can be performed is described below. We then describe an example of a Viterbi engine for performing these operations.

The described example shows how to obtain a traceback pointer. However, it is of course not necessary to determine these pointers if the recognition process used does not use them, as shown optionally in connection with operation 51 of FIG.

The finite state machine of FIG. 4 represents a single word within a given word zero of the recognizer. This is the three normal states A, B, and C.
has. However, in reality, instead of these three states, it will typically have many more states.

Additionally, this finite state machine has start and end dummy states (SD and ED). Normal model states are separated by transitions as indicated by arrows. In the speech model, only left-to-right transitions and transitions from one state to itself are used. In a simple example for a word consisting of three sounds, these sounds must occur in order from left to right. Alternatively, one sound can remain in the same state for the duration of one or more time frames.

The information provided by the Viterbi engine to processor 63 so that words can be indicated as recognized relates to the maximum possible paths through the model and the lengths of these paths. The maximum likelihood is determined by assigning two types of penalties to the model. That is, transition penalty - tp (sl, s2). This is separate for each transition between or into itself as indicated by the arrow.

State penalty -5p(I,t). It is assigned by index 1(S) to one normal state in a particular iteration, 11 of the Viterbi engine.

The transition penalty and index need not remain constant for a particular Viterbi engine model. However, in this example, we will keep these constant.

Table 1 of FIG. 5 shows how the values associated with the index change over time. In the first iteration (0), index 11. I2. ...The value associated with Ii is formed by the elements of the distance vector for the first frame (0). Then, the subsequent distance vectors are in columns 1, 2, etc. of Table 1. ...form i. Only the current distance vector is maintained by the Viterbi engine.

Next, a method for calculating the minimum distance passing through each model will be explained. At each iteration, the Piterbi engine calculates the minimum cumulative distance for each normal state. This calculation starts at the rightmost state and proceeds to the left. Therefore, for iteration O and state C, the minimum cumulative distance from the three paths is determined. Then, for each path, some stored cumulative distance relative to the initial state of the path is added to the transition penalty for that path.

The obtained minimum value is penalized to exist in the obtained state fiQ using the index assigned to state C and the current distance vector (i.e. the currently available column of Table 1). Added. The Minimum Cumulative Distance (MCD) obtained in this way for each state is held in a single storage device that holds one row of Table 2 of FIG. 6 at any given time. In the first iteration O, all values in row O are initialized to the maximum value. therefore,
When iteration 1 is performed, the cumulative distances for states A, B, and C are available and the minimum cumulative distance mentioned above is updated. For example, to update the minimum cumulative distance for state C, the previous cumulative distance O for state HASB,C is available. In iteration 1, for state C, after determining which has the smallest cumulative distance, use that distance to update M, CD. At the same time, the traceback pointer (TB) (which is initialized to the value O at iteration O) also increments the previous traceback pointer for that state by one when calculating the new minimum cumulative distance. Update by doing. Therefore, each iteration for each state (S)
>, one minimum cumulative distance (MCD (s, t)
) and one traceback pointer (TB (s,
t)) is obtained as follows.

Here, t relates to the current iteration. DV(1(
s), t) are the elements of the distance vector obtained using the index I, (S) for the state S in the iteration tk:. tp(x.

S) is the transition penalty for transitioning from state qx to state s. MCD (s, t) is the minimum cumulative distance to state S in an iterative likelihood.

TB (S, t) is the traceback pointer for state S at iteration t.

min x () represents the minimum of the values in C) for all valid values of X. Argm; nX() is
Represents the value of X that minimizes the value in parentheses.

As previously mentioned, the Vitel e-engine obtains a minimum cumulative distance in each frame and obtains a traceback pointer for each state. For each finite state machine, these minimum cumulative distances and corresponding traceback pointers are sent as data to the decision logic circuit to indicate which words were clearly spoken. In addition to the operations described above, the Viterbi engine obtains the minimum of the minimum cumulative distances for all models at the end of each complete iteration over all models. Then, in the next iteration, this minimum value is subtracted from all of the stored minimum cumulative distance values. By performing this scaling operation, it is prevented, as far as possible, that the value of the stored cumulative distance increases beyond the size of the storage location in the storage device. The following can be seen from the above equations. That is, the minimum cumulative distance is formed by an addition operation, and therefore the minimum cumulative distance must increase. By subtracting a certain minimum value from all the obtained values, the power increase trend can be suppressed. However, if a certain minimum cumulative distance stored reaches the maximum storage capacity, the minimum cumulative distance is not increased and is held at that value. In order to avoid storing a misleading cumulative distance value, once the maximum value is reached, the above-described subtraction step is not performed. The maximum value will eventually come out of the storage device automatically (.

This is because in each iteration the minimum cumulative distance is stored and the maximum value is eventually removed.

The two dummy states in Figure 5 are used to simplify the transition from one finite state (representing a single word) to another. In Figure 7, two different words are
have similar models A, B, C and A'', B-, C-, respectively. However, as already mentioned, the transition penalty -
and the index are different. The normal transition from the end of the first model to the beginning of the second model is indicated by arrow 110 leading from state C to state -. but,
If part of the word is not pronounced, arrows 111 and 1
Transitions such as those shown at 12 often occur, ie, those that omit the last state of the first model or the first state of the second model. Transitions from state B to state B- also often occur. Considering all the possible transitions between the 64 words contained in the given quantities of a typical recognizer, if implemented on the basis of FIG. 5, would be extremely complex. However, this problem can be avoided by using the dummy state described below. When the Viterbi engine completes updating the normal state of a word, it updates the ending dummy state (ED) by obtaining the minimum cumulative distance of that state. The update is performed in the same way as when updating the normal state. However, there is no state penalty for the termination dummy state. A traceback pointer is also stored for the exit dummy state. This traceback pointer is also 1q in the same way as other traceback pointers.

However, the traceback pointer corresponding to the chosen minimum cumulative distance is not incremented prior to storage.

When the Viterbi engine has processed all models, the ending dummy with the smallest of the minimum cumulative distances is the starting dummy state (SD) of the chosen word model.
Use to update. Other starting states are updated to some maximum value.

Traceback pointers for all starting dummies are set to zero. The starting state is chosen for update with the smallest minimum cumulative distance based on certain grammar rules. According to the grammar rules, it can be determined whether a word included in the vocabulary to be recognized can follow a previous word. If no such grammar exists, or if the grammar is ignored, all starting states are updated by obtaining the smallest minimum cumulative distance. In this method, the minimum path and cumulative distance for transitioning from any finite state machine to another is recorded for each machine.

The Viterbi Engine can be constructed using a general purpose computer. However, for continuous speech recognition, it is preferable to construct a dedicated computer with discrete integrated circuits or a single clip with specially metallized gates or logic arrays in order to perform the processing in a sufficiently short time. A special purpose computer, an example of which will be described later, has, for example, a computing unit with an output device connected to a number of latches. Some latches are connected to the random access memory address bus; others are connected to the data bus.
connected to the bus.

To make a computer like the one described in the previous paragraph flexible for this application, we must avoid hard wiring that does not allow us to change the number of states or the transitions between states in our finite state machine model. For this purpose, RAM6
Allocate part of 7 as a "skeleton". This "skeleton" determines the number of states included in each model and the transition paths between states. Various parts of the RAM are shown in FIG. One of these parts 115 is a "skeleton". Portions 116-118 contain most of the data for each of the three models. There is one similar part for each model. The model in Figure 4 is RAM1
16, we can see that: The three lower storage locations contain three entry transition penalties -1p1 to t, 3, corresponding to transitions to state RC from the state of count or from itself. The fourth memory location contains the index. With this index, the state penalty can be obtained from Table 1. part 1
16 is state A, E3. It is divided into other storage locations for EEo. State C has one more memory location than states A and B because there are three entry transitions.

The state ED requires one less memory location than the state A because there is no penalty for this state. Status S
D has no transition to this state and the associated penalty -
Therefore, there is no storage location in portion 116. When performing an iteration, the Viterbi engine uses pointers to move sequentially through portions 116, 117, 118, and other parts of the model, as described in the section. One pointer is the skeleton RA
It is also used to move through the M section 115. In the skeleton RAM portion 115, for each state, an offset for each transition penalty is stored.

Other parts of the RAM, one for each word, are set apart to store the minimum cumulative distance and traceback pointers associated with each state of each word. Examples of these parts are 120.121.1 in Figure 6.
22. For the words shown in FIG. 4, the RAM portion 120 is divided into five pairs of memory locations, one for each state, for which the cumulative distance and traceback pointer are stored. including.

Three pointers, indicated by arrows 123-125 and held by latches in the dedicated computer, are used to update the minimum cumulative distance. The first pointer 123 initially points to the first transition penalty of the last state of the first model to be updated. In this example,
It is held by the bottom memory location of RAM portion 116. When the Viterbi engine transitions to state C,
It must be determined which cumulative distance is the minimum. To do that, the Viterbi engine must find the cumulative distance for each of the paths to state C. Then you have to decide which is the smallest. After the transition penalty -1 is reduced by 1q,
Arrow 124 points to an offset in skeleton RAM portion 115. This offset determines the cumulative distance position in the state from which the first transition begins. Therefore, if the starting state is the current state, the offset is zero and the cumulative distance is that given by pointer 125. The distance obtained is stored and pointers 123 and 1-24 are incremented by one. This adds a second migration penalty to the cumulative distance given by offset 2. Offset 2 may be "2" in this case. In that case, the value held in the memory location with an address two greater than the address of pointer 125 is read. That is, it is the cumulative distance of state B. The distance obtained in this way - if it is smaller than the one obtained previously - is stored. pointers 123 and 12
4 is again incremented to give the third transition and the cumulative distance to be added to it. The obtained distance is then stored again if it is smaller than the already stored distance.

In this method, the minimum value of the cumulative distance obtained is stored. Pointers 123 and 124 are incremented again and then
Pointer 123 provides an index to obtain the appropriate state penalty. Index 124 indicates that the iteration for state C has reached its end. next,
The minimum cumulative distance obtained together with the appropriate traceback pointer is stored at the location given by pointer 125. All of these pointers are then incremented. The cumulative distances for states B and C are then updated in the same manner as pointers 123.124.125 were incremented. However, the cumulative distance and traceback pointer to the end state (ED) is obtained when all states in a model are updated (excluding their starting state) when all models are updated as described above. . Then, the minimum cumulative distance maintained by any ending dummy state is entered into the chosen starting state (SDS) based on the grammar rules.

By using "skeleton", the model can be modified in three ways. First, by changing the offset, the transition between states can be changed. Second, if we change the number of offsets between two indicators, the number of transitions to a certain state changes. This indicator indicates that a state update is complete. The number of states in each model can be changed by increasing the number of groups of offsets stored in RAM portion 115. The number of models can be many parts, for example 116
．． 117.118 or 120.121.122).

The flowchart in Figure 9 shows one iteration of the Viterbi engine, but the main points of its operation are:
It should be clear from the chart.

In operation 130, a distance vector is read into a memory location in RAM that represents the current column of the first table.

Then loop 131 is entered. Here, for one transition of the last normal model of the first model, the cumulative path 11
tcDn is calculated. If the value of CD, is less than the reset minimum cumulative distance value for that state (i.e., the previously calculated (t!!(MCD)), then in operation 132, set MCD to CD, Set to .

At the same time, the corresponding traceback pointer for the cumulative distance of the starting state for that state is incremented by one. Decision 133 determines whether a transition to the current state exists. If not, then loop 134 is performed. Here, scaling is performed by subtracting the smallest minimum distance from the previous iteration, adding the state penalty at that state,
Store the minimum MCD.

At the end of loop 134, we store MCD and TB for that state, and then loop 131 for the next state.
repeat. When performing this iteration, MCD is reset to the new state and CDn, tp, C
OD, TB relate to the new state. If decision 135 determines that the new state is the last state, loop 13
Repeat steps 1 and 134. However, if necessary, the variables are reset again to represent the new model and new state. When all models have been processed, as indicated by decision 137, - the starting dummy state of all models is updated;
Various MCs for all states of all models [)
and TB are awaited so that they can be output to the processor 63. Once output to the processor, the determined word is indicated as recognized. The next iteration is then initiated at operation 130.

The circuit 140 shown in FIG. 10 can be used for the Viterbi engine (ie, logic circuit 66). For example, circuit 140 may be constructed from a discrete integrated circuit or a custom gate array. The Viterbi engine uses a 16-bit address bus 142
and the 8-bit data path code 43, the RAM 67
connected to. RAM67 is area 1 in Figure 8! ! 115
~120 and similar regions.

When an iteration of the Viterbi engine is performed, a "start" signal is sent to terminal 144 connected to controller 145. Controller 145, in one example, includes a sequencer and memory (neither shown). In response to a "start" signal, the sequencer cycle through all its states addresses the controller memory.

The controller memory outputs a continuous pattern of binary bits on terminal 146. viterbi engine 140
The enabled terminals of all circuits (other than controller 145 ) are each connected to a respective terminal of controller 145 . This allows different circuits or groups of circuits when a pattern of bits appears.

The first operation uses a parameter base address held permanently by a latch.

This addresses the parameter array in RAM 67. The buffer 148 is connected to the controller 145
enabled by. And RAM is bus 14
Numbered by 2. The parameter array is 4
Holds the addresses of pointers. The four pointers are 123 (penalty-'pointer, Pent), 1
24 (skeleton pointer, Tpt), 125 (
a cumulative distance pointer (CDpt), and an index pointer ([)vec) for reading the first table. Is each of these pointers a buffer? 154 and latch 155 in serial 8-bit bytes to latch 150.1.
51, 152, and 153 respectively. The controller 145 again sends the necessary control signals (hereinafter referred to as
Controller functionality is treated herein as understood, except where special performance is required).

If the cumulative distance is to be decreased by a certain amount, that decrease is also initialized by reading another element of the parameter array and inputting the result to latch 158. To complete the initialization, two latches 156 and 157 representing the minimum previous cumulative distance plus transition penalty and the minimum cumulative distance for the current iteration are set to maximum, respectively.

Next, skeleton pointer 124 in latch 151 is used to read the first offset from RAM 67.
use. This first offset is
is loaded into latch 160 via . The pointer 124 is then incremented by applying it to one input of the arithmetic unit (ALU>161) via the buffer 162.
This occurs simultaneously when buffer 163 is used to force ``1''. Buffer 1159 loads incremented pointer 124 into latch 151.

Cumulative distance pointer 125 is sent from latch 152 to the ALU. In the ALU, the cumulative distance pointer 125 is
Added to the required offset held in latch 160. The result is then sent to latch 140 that holds a temporary pointer (Vpt). Vpt is used to read the cumulative distance (held in RAM 67) of the state from which the first transition (eg, from A to C in FIG. 1) departs. This is because the offset is the incremental address of this distance in region 120 from pointer 125. The read distance is the latch 16
Loaded within 5. buffer 162) latch 163,
By using ALU and buffer 1159, Vp
t is incremented by one. Therefore, vpt will indicate a traceback pointer corresponding to the cumulative distance held by latch 165. This pointer is loaded into latch 166. Next, the penalty -
- Pointer 123 is used to load the appropriate transition penalty into latch 160 from RAM 67. The ALU then adds the contents of latches 160 and 165, namely the cumulative distance and the transition penalty -1. Then, the result is sent to the latch 16 via the buffer 167.
Load into 0. Latch 156 typically holds the minimum cumulative distance obtained so far for a given state. However, for the first migration, the cumulative distance is set to the maximum, as already mentioned. therefore,
Typically, the ALU compares the contents of latches 156 and 160. A flag is then set on control line 168 if the contents of latch 160 are less than the minimum value previously obtained. By control line 168,
The controller can read from latch 160 into latch 156 and the traceback pointer from latch 166 into latch 169 . Therefore, we were able to obtain the best cumulative distance plus the migration penalty obtained so far and the corresponding traceback pointer.

Next, to begin studying another transition to the first state,
Use skeleton pointers. Then, the sum of the accumulated distance and the obtained displacement (tepenalty) is the sum of the latch 1
56 to determine if it is less than the value held at 56. If so, the sum, along with the corresponding traceback pointer, is added to latches 156 and 1.
69. This process continues until skeleton pointer 124 crosses the first indicator. The first indicator is of the type known as an end of state (EO3) indicator. This kind of indicator is in the buffer 154.
, the indicator is detected by detector 170 and controller 145 is instructed to proceed in the manner described below.

In the ALU, from the contents of launch 156, the minimum cumulative distance obtained in the last iteration and held in latch 158 as described above is retrieved. The result is then written into latch 165.

The address pointed to by penalty pointer 123, the index in table 1, is used to write that index into latch 160. Then, in ALLJ, these contents are added to the contents of latch 153. The latch 153 will now hold the base of that area of the RAM 67 as a result of the initialization described above.
Retain address. This RAM 67 contains the elements of the distance vector (ie, the current column of Table 2). The above result is loaded into latch 164 and
It becomes t. This Vpt is used to load the appropriate elements of the distance vector into latch 160. A
At the LU, the contents of latch 160 and the contents of latch 165 (which holds the scaled minimum cumulative distance) are added to obtain the updated cumulative distance. This updated cumulative distance is
The memory location indicated by pointer 125 held at 52 is entered. The traceback pointer in latch 169 is then incremented by one in the ALU and the result is input into latch 160.

The contents of latch 152 are also incremented by one within the ALU. The address so formed buffers the traceback pointer in latch 160? It is used for inputting to the RAM 67 via 171. To prepare for processing the next state, the contents of this latch are again incremented by one.

This completes the operations for one state.

The skeleton pointer is used again to read RAM and other normal conditions are handled in a similar manner. Eventually, the detector@170 obtains the indicator between the last normal state and the termination dummy.

This indicator is of the type known as an end of word (EOW). When the next state (state after R) is processed, the following operations are omitted. That is, operations that use the index and operations that increment the traceback pointer. Also, the termination dummy operations are in the opposite direction through the model (see Figure 4), and since these operations require an offset between the pointer 125 and the cumulative distance read, that offset is subtracted instead of incremented.

When the next mark after the EOW mark encounters skeleton pointer 124, initialization of latches 151 and 156 is performed. Then the next word model is processed. Invalid migration penalties are maintained in memory locations above area 116 of RAM 67. Then, after the last word in the edict has been processed, when the controller attempts to continue processing, the read first transition penalty is obtained by the detection device 170, and this value is the invalid penalty -
It is said that An end-of-word (EOV) control signal is then delayed to the controller, signaling the end of the repetition.

The controller then extracts the minimum cumulative distance obtained in this iteration for use in the next iteration.

Then, it sends an "end" signal to terminal 172. The decision logic circuit connected to this is usually a microprocessor)
then uses the accumulated distance and the traceback pointer sent to RAM 67 to perform the iteration. The decision logic circuit or associated microprocessor
No further operation is performed on the Viterbi engine until a further "start" signal is sent to terminal 144.

As shown in Figure 10, the upper input of ALtJ161 is 1
The lower input is 8 bits. latch 1
Since the output of 56 to latches 158, 165, and 169 is 8 bits, latch 174 is provided. This latch 174 is set to 8 zeros on the high valid side of the top input of the ALU when the latches described above are input to the ALU.
Force bit.

A detector 175 is connected to the output of ALLJ to detect saturation and control buffer 176. This ensures that whenever saturation occurs, the buffer 176 is in the buffer? 167. buffer 176
contains the maximum possible number.1. On the other hand, buffer 7167 can contain an overflow. Detector 175 also has an input from buffer 163. If this input indicates that the ALU is performing an operation with a maximum value, then saturation is also assumed to have been detected and the output of buffer 176 (i.e., the maximum value) is used as a result of that operation.

A conventional clock pulse system is operated to synchronize the various circuits operated by controller 145. In other words, the external clock signal is connected to the terminal 173.
A pulse is input. The arrow pointing to the right of each latch represents the clock terminal, and the circle pointing to the left represents the available terminal.

Although the device described above concerns the recognition of isolated words, it can be extended to the recognition of 'continuous speech' formed from a limited number of trained words.

That is, this is possible, for example, by using traceback pointers to track the various word model states discussed in detail (see P., F., Brown et al., supra, and NeV. Please refer to ).

It will be obvious that the invention may be practiced in many other ways than those described above. For example, other hardware can be specially designed to obtain minimum cumulative distance. Also, the distance vector can be subtracted by 1q from the linear prediction function. Roughly speaking, the Viterbi process can be replaced by a forward-backward algorithm.

The probability function can be obtained in other ways than the specific methods mentioned above. However, the number of probability functions must be smaller, preferably significantly smaller, than the number of states in all finite state machines.

[Brief explanation of drawings]

Figures 1a and 1b form a flowchart of the method according to the invention for selecting states and calculating the probability density function 1q for a certain word. FIG. 2 is a flowchart of the method according to the invention for recognizing words. FIG. 3 is a block diagram of an apparatus according to the invention. FIG. 4 is a diagram of a finite state machine used to explain one embodiment of the invention. FIG. 5 is an overview of a table showing how the input values for the Telbi engine used in one embodiment of the present invention are generated and accessed. Figure 6 shows how the Viterbi engine produces one line at a time.
1 is a table showing quantities that are calculated and stored; FIG. 7 is an example of how transitions between states representing different words occur. FIG. 8 is a diagram of storage locations within a Viterbi engine used in one embodiment of the present invention. FIG. 9 is a flowchart of a Viterbi engine used in one embodiment of the present invention. FIG. 10 shows the Viterbi system used in one embodiment of the present invention.
FIG. 2 is a block diagram of a dedicated circuit for the engine. 63...Processor, 64...Processor, 66...Logic circuit, "67...Random access memory (RA)
M), 72...Host computer, 145...
...Controller, 161... Arithmetic unit (, ALU>. Showa year, month, day 1. Display of incident Patent application No. 195109 of 1988)
No. 2) Title of the invention: Speech recognition method and device 3
, Relationship with the case of the person making the amendment Applicant 4, Agent

Claims (1)

[Scope of Claims] (1) A device for recognizing words within a predetermined vocabulary, which includes the following components. (b) Means for sequentially sampling sounds to obtain a set of signals each representative of the characteristics of those sounds. (b) means for storing data representative of a number of finite state machines for each word of a predetermined vocabulary and forming individual models of those words, said data representing a number of finite state machines; Describe the state that forms,
A probability function is assigned to each state, where at least one probability function is assigned to more than one state, and each probability function is determined based on whether the model generates the actual sound. Then, a means for describing the probability that a signal representing the characteristics of a sound will pretend to be some observed value if any model is in a state to which the probability function is assigned. (c) The probability that a given set of signals will be generated if the model that generated the actual sound and any model are in any given state is calculated by combining signal probability functions. From, the means to decide. (d) A means of determining, from calculated probabilities and properties of finite state machines, the maximum likelihood that a sample of consecutive sounds will represent given words. (e) Means for providing an output indicating one of the predetermined words as the most likely spoken word based on the maximum probability detected. (2) The device according to claim 1, wherein the means for calculating the probability includes means for calculating a distance vector from each set of the probability function and the signals when the various signals occur; If any finite state machine generates a real sound and occupies a state that is assigned a corresponding probability function, then each element of the distance vector has a corresponding probability function and the current signal represents the reciprocal of the probability of observing a set of having, equipment. (3) In the device according to claim 1 or 2, each finite state machine has a plurality of states connected by transitions, and each state and each transition correspond to each other. states and transition penalties, maintained as part of the data representing the finite state machine, each transition penalty being a property of the finite state machine,
Apparatus being part of a finite state machine, wherein each state penalty depends on a probability function assigned to the corresponding state and on the current set of said signals. 4. A method for selecting a number of states for a finite state machine to represent words within a predetermined vocabulary and obtaining data for characterizing the states, the method comprising the steps of: (b) Sequentially sampling the sounds forming each word in the vocabulary to obtain a set of signals representative of the characteristics of the sounds forming the word. (b) Obtaining data defining the state for each word from the set of signals. (c) Obtaining one probability function for each state from the set of signals. (d) merging several probability functions to obtain a number of said functions less than a predetermined number, the merging comprising:
Steps performed according to criteria relating to merging suitability and a predetermined number. (e) Calculating the probability of transition from each state to all allowed subsequent states in each finite state machine. (f) Entering data identifying each word as it is spoken. (g) Data that determines the conditions for forming each machine,
and storing data identifying the words represented by the machine. (H) Storing data defining each merged probability function. (5) In the method according to claim 4, the data defining the states to obtain one probability function for each state is obtained by merging several sets of signals for each word according to a criterion for merging suitability. A method comprising providing data defining a state for the word. (6) In the method of claim 5, successive sets of signals are merged, and the merging compatibility of the two successive sets is such that each set of signals for potential merging arises from a separate set. The method is pre-evaluated by calculating the logarithm of the ratio of the probabilities. (7) In the method according to claim 5 or 6, the merging suitability of two probability functions is determined by determining the probability that each probability function for potential merging arises from the same probability function. A method in which a probability function is pre-estimated by calculating the logarithm of the ratio to the probability resulting from separate probability functions.