JP2001083981A - Speech recognition system and method and recording medium readable by computer having recorded voice recognition program therein - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition system and a method capable of suppressing a cost lowly, and simultaneously capable of shortening a recognition response time, by memorizing a word dictionary in plural recording mediums, and a recording medium readable by a computer, having a speech recognition program recorded therein. SOLUTION: In this speech recognition system equipped with an acoustic analytical means 1, a reference model memory means 2, a dictionary memory means for memorizing plural partial dictionaries obtained by dividing a word dictionary, comparison data memory means 4, and a model comparison means 5 for executing comparison processing by consulting a reference model and the word dictionary relative to a characteristic vector, and for outputting a recognition result, the dictionary memory means is composed of a first dictionary memory means 6 capable of being consulted at high speed, for memorizing a partial dictionary having a high frequency of use, and a second dictionary memory means 7 incapable of being consulted at high speed, for memorizing a residual partial dictionary having a low frequency of use. Hereby, a cost can be suppressed lowly, and simultaneously a recognition response time can be shortened.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition system and method for performing large vocabulary recognition such as address search by voice, and a computer-readable recording medium storing a voice recognition program.

[0002]

2. Description of the Related Art When speech recognition of a large vocabulary is performed, a calculation amount is generally reduced by utilizing a technique such as a beam search. Here, Japanese Patent Application Laid-Open No. 10-2544
A conventional speech recognition apparatus will be described using the speech recognition apparatus disclosed in Japanese Patent Publication No. 79 as an example.

FIG. 15 shows, for example, Japanese Patent Application Laid-Open No. H10-25447.
FIG. 9 is a block diagram illustrating a configuration of a model matching processing unit of a conventional speech recognition device disclosed in Japanese Patent Application Laid-Open No. 9-No. Hereinafter, a description will be given using an HMM (Hidden Markov Model) as a speech recognition method, an address as a recognition target, and a node as a recognition unit as a place name.

In FIG. 15, reference numeral 1 denotes a sound analysis means for inputting a voice signal, performs sound analysis and converts it into a time series of feature vectors, 2 a standard model storage means for storing a standard model to be recognized, and 3 an address. Storage means for storing the matching data as a work area in the matching processing, and reference processing for the feature vectors from the acoustic analysis means 1 with reference to the standard model and the dictionary. And outputs a recognition result.

Next, the operation of the conventional speech recognition apparatus will be described with reference to the drawings.

In this description, the address shown in FIG. 16 is to be recognized. Prior to recognition, it is assumed that the standard model is stored in the standard model storage means 3 and a dictionary expressing addresses is stored in the dictionary storage means 3.

FIG. 17 shows the contents of the dictionary storage means 3 representing the addresses shown in FIG. In FIG. 17, the place name of the address to be recognized is indicated by a square, and the connection is indicated by an arrow. Each word is divided and stored in a partial dictionary indicated by a frame with a rounded corner, and can be taken into the model matching unit 5 for each partial dictionary and used for a matching operation. Triangles indicate entry points to partial dictionaries, for example, partial dictionaries net5
"Enoshima" indicates that it is connected to each of "1 chome", "2 chome" and "3 chome" of the partial dictionary net9.

The structure of the partial dictionary as an example of the partial dictionary net2 shown in FIG. 18 will be described. This partial dictionary net2
Has one entry point (entry0) and has "Kanagawa" of node number node2 inside and node number nod
It has two nodes of "Kagawa prefecture" of e3.

[0009] Kanagawa is then connected to entry point entry0 of partial dictionary net3, and Kagawa is then connected to partial dictionary net4.
Connected to entry point entry0 of When model matching is performed on a node in Kanagawa Prefecture, it means that the HMM parameter of the node number node2 of the standard model storage means 2 is used.

FIG. 19 is a diagram showing the contents of the standard model storage means. HMM parameters corresponding to each node are stored. It is assumed that the parameters required for the HMM collation operation, such as the number of states and the transition probability between states, are included in advance as the parameters of the HMM.

When the recognition is started, first, the net1 including the silence, which is the first node, is read from the dictionary storage means 3, and
A necessary work area is set in the collation data storage means 4. This is shown in FIG. When the recognition process proceeds and the subsequent node collation process is required, as shown in FIG.
The partial dictionary of net2 following t1 is read from the dictionary storage means 3, and a necessary work area is obtained in the collation data storage means 4. As the recognition process proceeds, the dictionary storage means 3
More necessary partial dictionaries are read and collation data storage means 4
Work area is taken up.

Next, the operation of model matching will be described.

When a sound signal is input to the sound analysis means 1 shown in FIG. 15, sound analysis is performed at fixed time intervals, converted into a feature vector and output. While the audio signal is being input, the feature vector is repeatedly sent from the acoustic analysis unit 1 to the model matching unit 5.

In the model matching means 5, the processing shown in the flowchart of FIG. 22 is repeatedly performed every time a feature vector is input. Step 501 is an intra-node operation for performing a collation operation on the state in the node;
Step 502 is an evaluation value determination process for determining an evaluation value of a beam search, step 503 is a beam search operation process for performing a beam search process, and step 504 is an inter-node operation process for performing an operation between nodes.

FIG. 23 is a diagram for explaining details of the data structure in the collation data storage means 4 for the node n.
20 and 21 show the contents of the collation data storage means 4 in node units, but FIG. 23 describes the contents of the nodes.

The node n has three of S _n (1), S _n (2), and S _n (3).
It is assumed that it is composed of states. The state S _n (0) at the left end of the node n is a pseudo state used for the operation between nodes, and three states of _Sn (1), _Sn (2), and _Sn (3) correspond to the model. Represents an entity.

A _n (i, j) represents a penalty based on the transition probability from state i to j, and b _n (i) represents a penalty based on the output probability of state i. The penalty increases as the probability decreases. The penalty based on the transition probability and the penalty based on the output probability are parameters used as standard in recognition using the HMM, and a detailed description thereof will be omitted. These parameters are stored in advance in the standard model storage means 2 as shown in FIG. 19, and are read from the standard model storage means 2 when the partial dictionary including the node n is read in the dictionary storage means 3, A work area is set in the collation data storage means 4 as shown in FIG.

In the intra-node operation processing (step 501),
In other words, every time a feature vector is input,
The model matching operation shown in the following equation 1 is performed using the transfer probability.
And S _n(i) is updated. I_nIs the number of states of node n
Represents The output probability is based on the input feature vector.
Probability calculation for distributions representing acoustic features in each state
Is required.

S _n (i) = max (S _n (i) + a _n (i, i)), (S _n (i−1) + a _n (i−1, i)) + b _n (i) i = 1, I _n ··· formula 1

In the evaluation value determination process (step 502),
The evaluation value En for beam search is calculated by the following equation 2.
ode (n) and evaluation value Earc (n) for performing transition between nodes
And a beam search reference value Ebestnode. Ebestno
de is represented by the evaluation value of the best beam search among the nodes.

Enode (n) = max (S _n (i)) _{1 <i <In} Earc (n) = S _n (I _n ) Ebestnode = max (Enode (n)) _{1 <n <N} 2

Beam search calculation processing (step 503)
Then, when the following condition is not satisfied, the collation operation of the node is inactivated to reduce the amount of operation. Intra-node arithmetic processing is not performed on the inactivated node. Tnode is a threshold value for beam search, and uses a predetermined value.

Enode (n)> Ebestnode−Tnode Equation 3

Subsequently, the inter-node arithmetic processing (step 50)
In 4), if the condition of the following expression 4 is satisfied and the work area is not in the collation data storage means 4, the partial dictionary including the new node is read from the dictionary storage means 3, and the work area of the new node is read. Is secured, and the newly read node is activated to perform an inter-node matching operation.

Earc (n)> Ebestnode−Tarc Equation 4

FIG. 24 shows a state where the work area of the node n + 1 is taken when the subsequent node is set to n + 1. Immediately after the reading, the score of each state of the node n + 1 is initialized to the lowest value. Then, node n and node n + 1
Between nodes are collated. The collation processing between nodes is represented by the following Expression 5.

S _{n + 1} (0) = S _n (I _n ) Equation 5

The next node n + 1 is the collation data storage means 4.
In the case where the work area has been set up and has been deactivated, the work area is activated again, and the collation operation between nodes shown in Expression 5 is performed.

The model matching means 5 repeatedly performs the matching process as described above every time a feature vector is input.
When the processing for all the feature vectors of the audio signal is completed, the node sequence having the highest score is output as the recognition result.

[0030]

In the above-described conventional speech recognition apparatus, when recognizing a large vocabulary such as a nationwide address in Japan, the data capacity of the dictionary storage unit 3 reaches 40 megabytes (Mbytes). There is a problem that the cost of the speech recognition device is increased to store the data on a high-cost recording medium such as an EPROM.

Also, a CD-ROM or a DVD-R
When the data is stored in a low-speed recording medium such as the OM, there is a problem that the dictionary reading time is long and the recognition response time until a recognition result is obtained is long.

The present invention has been made in order to solve the above-mentioned problems. By storing a word dictionary on a plurality of recording media, it is possible to shorten the recognition response time while keeping costs low. It is an object of the present invention to obtain a computer-readable recording medium recording a system and a method and a speech recognition program.

[0033]

According to a first aspect of the present invention, there is provided a voice recognition system which receives a voice signal, performs a voice analysis, converts the voice signal into a time series of feature vectors, and outputs the time series. A standard model storage unit that stores a standard model, a dictionary storage unit that stores a plurality of partial dictionaries obtained by dividing a word dictionary, a collation data storage unit that stores collation data as a work area in the collation processing, and a sound analysis unit. In the speech recognition system including a model matching unit that performs a matching process on the feature vector with reference to the standard model and the word dictionary and outputs a recognition result, the dictionary storage unit can be referred to at high speed. First dictionary storage means for storing frequently used partial dictionaries, and second dictionary storage means for storing remaining infrequently used partial dictionaries that cannot be referenced at high speed. Are those that you configured.

According to a second aspect of the present invention, in the speech recognition system, the dictionary storage means indicates whether the partial dictionary is stored in the first dictionary storage means or the second dictionary storage means. Is further included in the storage medium storing means.

In the speech recognition system according to a third aspect of the present invention, the first dictionary storage means and the second dictionary storage means correspond to each node, and the next partial dictionary to be connected is the first dictionary. The storage unit stores a partial dictionary including recording medium information indicating which of the storage unit and the second dictionary storage unit is stored.

In the speech recognition system according to a fourth aspect of the present invention, the first dictionary storage means and the second dictionary storage means store partial dictionaries having a strong dependency on the same recording medium as a group. When the model matching means refers to a certain partial dictionary, a group including the partial dictionary is read out collectively and a matching process is performed.

According to a fifth aspect of the present invention, in the speech recognition system, the dictionary storage means further includes a group storage means for storing a group including the partial dictionary transferred together when referring to the partial dictionary. The model matching means performs the matching process by individually referring to the partial dictionaries in the group stored in the group storage means.

In the speech recognition system according to a sixth aspect of the present invention, the dictionary storage means further includes a transfer amount measurement means for measuring an amount of the partial dictionary read from the second dictionary storage means to the collation data storage means. The model collating means compares the measured amount by the transfer amount measuring means with a specified amount, and controls so as not to read a partial dictionary having the specified amount or more.

According to a seventh aspect of the present invention, in the voice recognition system according to the first aspect, the first dictionary storage unit may add a partial dictionary composed of only a head portion of a group to be stored in the second dictionary storage unit to a new group. The second dictionary storage means stores a partial dictionary composed of the remainder from which the leading part has been omitted as a new group, and when the model matching means refers to a certain partial dictionary, Are read out collectively and collation processing is performed.

According to the speech recognition method of the present invention, a speech signal is inputted, acoustic analysis is performed, and the speech signal is converted into a time series of feature vectors. In a voice recognition method for performing a matching process while referring to a plurality of divided partial dictionaries and outputting a recognition result, in the matching process, first, a first dictionary that stores a frequently referenced partial dictionary that can be referred to at high speed. In the first partial dictionary reading step of reading the partial dictionary from the storage means, and in the collation processing, the second dictionary storage means or the first dictionary storage means for storing the remaining partial dictionaries which cannot be referred to at high speed and are not used frequently. Reading a partial dictionary from the dictionary storage means.

In the speech recognition method according to a ninth aspect of the present invention, in the second partial dictionary reading step, a partial dictionary to be connected next is stored in either the first dictionary storage unit or the second dictionary storage unit. A partial dictionary is read from the first dictionary storage unit or the second dictionary storage unit based on recording medium information indicating whether the dictionary is stored.

According to a tenth aspect of the present invention, in the second partial dictionary reading step, when referring to a certain partial dictionary, the partial dictionaries having a strong dependency are stored as a group. A group including the partial dictionary is read out from the second dictionary storage means.

In the speech recognition method according to an eleventh aspect of the present invention, in the second partial dictionary reading step, when a certain partial dictionary is referred to, a group including the partial dictionary is collected from the second dictionary storage means. The data is read out and stored in the group storage means.

According to a twelfth aspect of the present invention, there is provided a computer-readable recording medium storing a speech recognition program, comprising: a sound analysis step of inputting a sound signal, performing sound analysis, converting the sound signal into a time series of feature vectors, and outputting the result; A standard model storage area for storing a standard model to be recognized, a dictionary storage area for storing a plurality of partial dictionaries obtained by dividing the word dictionary, a collation data storage area for storing collation data as a work area in the collation processing, For a feature vector from the analysis procedure, a matching process is performed while referring to the standard model and the word dictionary, and a model recognition procedure for outputting a recognition result. The dictionary storage area, a first dictionary storage area that stores a partial dictionary that can be referred to at high speed and that is frequently used, Impossible reference speed is obtained by construction of a second dictionary storage area for storing the rest of the used infrequently moiety dictionary.

According to a thirteenth aspect of the present invention, in the computer-readable recording medium storing the voice recognition program, the dictionary storage area stores the partial dictionary in either the first dictionary storage area or the second dictionary storage area. It further includes a recording medium storage area for storing recording medium information indicating whether the information is stored.

According to a fourteenth aspect of the present invention, there is provided a computer readable recording medium storing a speech recognition program, wherein the first dictionary storage area and the second dictionary storage area correspond to each node and are connected next. The storage unit stores a partial dictionary including recording medium information indicating whether the partial dictionary to be stored is stored in the first dictionary storage area or the second dictionary storage area.

According to a fifteenth aspect of the present invention, in the computer-readable recording medium recording the speech recognition program, the first dictionary storage area and the second dictionary storage area are on the same recording medium and have a strong dependency. The partial dictionaries are stored as a group, and when the model collation procedure refers to a certain partial dictionary, a group including the partial dictionary is collectively read and collation processing is performed.

A computer-readable recording medium on which a speech recognition program according to claim 16 of the present invention is recorded, wherein the dictionary storage area includes a group including a partial dictionary transferred together when referring to the partial dictionary. Is further provided, and the model collation procedure performs collation processing by individually referring to partial dictionaries in the group stored in the group memory area.

A computer readable recording medium having recorded thereon a speech recognition program according to claim 17 of the present invention, wherein a transfer amount measurement for measuring an amount of a partial dictionary read from said second dictionary storage area to said collation data storage area. The method further includes a step of comparing the measured amount in the transfer amount measuring step with a specified amount, and controlling not to read a partial dictionary having the specified amount or more.

A computer-readable recording medium having recorded thereon a speech recognition program according to claim 18 of the present invention, wherein the first dictionary storage area starts from only the head of a group to be stored in the second dictionary storage area. The second dictionary storage area stores a partial dictionary composed of the remaining part from which the leading part has been omitted as a new group, and the model collation procedure includes a step of: When a dictionary is referred to, a group including the partial dictionary is read out collectively and a collation process is performed.

[0051]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 A speech recognition system according to Embodiment 1 of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a speech recognition system according to Embodiment 1 of the present invention. In the drawings, the same reference numerals indicate the same or corresponding parts.

In FIG. 1, reference numeral 1 denotes a sound analysis means for inputting an audio signal, performs sound analysis and converts the sound signal into a time series of feature vectors, 2 denotes a standard model storage means for storing a standard model to be recognized, and 4 denotes a matching process. A collation data storage unit 5 that stores collation data as a work area performs collation processing on the feature vector from the acoustic analysis unit 1 with reference to the standard model and the dictionary, and outputs model recognition results.

In FIG. 6, reference numeral 6 denotes a first dictionary storage unit using a recording medium that can be referred to at high speed, 7 denotes a second dictionary storage unit that uses a recording medium that cannot be referred to at high speed, and 8 denotes a partial dictionary. Storage means for storing the recorded recording medium (storage means).

Hereinafter, the first dictionary storage means 6 will be referred to as an EEPROM.
(Electrical Erasable Programmable Read Only Memor
y), the second dictionary storage means 7 is stored in a CD-ROM (Compact Dis
The description will be made as “k-Read Only Memory”. EEPROM
Is a recording medium that can be referenced at high speed, but has the disadvantage that the cost is very high. Further, the CD-ROM has an extremely large capacity and a low cost, but has a drawback that the reading speed is low.

When recognizing a large vocabulary such as an address, the size of the dictionary becomes enormous, and storing all the data in the EEPROM increases the cost. Plus everything on CD
-When stored in the ROM, the number of times the CD-ROM is referred to increases, and the time required for recognition increases due to the overhead caused by the reference, which causes a problem in recognition responsiveness as a speech recognition system. Therefore, two types of storage media for storing dictionaries are used, a frequently used partial dictionary is selected and only the selected partial dictionary is stored in the EEPROM, and the infrequently used partial dictionary is stored on a low-cost CD-ROM. Shall be.

FIG. 2 is a diagram showing the configuration of the word dictionary of the speech recognition system according to the first embodiment. In FIG. 2, the partial dictionaries net1, net2, net9, net1 enclosed by solid lines
2 is stored in the first dictionary storage means 6 (EEPROM),
The remaining partial dictionaries surrounded by dotted lines are stored in the second dictionary storage means 7 (CD-ROM).

Since the partial dictionaries net1, net2, and net12 are always required for recognition, they are recorded in the first dictionary storage means 6, which is a recording medium that can be referenced at high speed. Also, the partial dictionary ne
Since t9 is connected from many partial dictionaries, it is also recorded as having a high possibility of being referred to. In FIG. 2, the remaining partial dictionaries indicated by squares surrounded by broken lines are stored in the second dictionary means 7, and are read as needed.

The basic operation of the speech recognition process is the same as that of the conventional example, and the description is omitted here. With the above configuration, if the utterance is Kanagawa Prefecture, the node of Kanagawa Prefecture satisfies the condition of Expression 4, and the subsequent partial dictionary net3 can be read. Further, in the case of the utterance of Kagawa Prefecture, the partial dictionary net4 in which the node of Kagawa Prefecture satisfies the condition of Expression 4 can be read.

Which recording medium contains these partial dictionaries is determined by referring to the recording medium storage means 8. FIG.
Indicates the contents of the storage medium storage means 8 and is assumed to be stored in the storage medium storage means 8 prior to recognition.

An example of another word dictionary will be described. FIG. 4 is a diagram showing an example of another word dictionary of the speech recognition system according to the first embodiment. This means that if a user residing in Kanagawa Prefecture uses this speech recognition device, the frequency of searching for an address in Kanagawa Prefecture will increase, so the partial dictionaries net3, net5, and net6 containing the names of places in Kanagawa Prefecture will be the first dictionary It is configured to be stored in the storage means 6 (EEPROM).

While using the speech recognition system, the use frequency of each partial dictionary is learned, and the first dictionary storage means 6 (EE
It is also effective to rewrite the contents so that the partial dictionary of (PROM) has a high reference frequency. A speech recognition system having these functions is also included in the scope of the present invention.

In the above description, the EEPROM and the CD-RO
M has been described as an example of a storage medium, but instead of an EEPROM, a flash ROM, an EPR
OM, ROM, RAM, etc. Also, instead of CD-ROM, CD-RW (rewritable CD), DVD-RO
M, DVD-RAM, DVD-RW (rewritable DV
D), a hard disk or the like may be used, and the same effect can be obtained.

The classification of the partial dictionaries used in the above description is merely an example, and other classification methods may be used.

In the above description, a method using two types of storage media has been described. However, three or more types of storage media may be used, and the same effect can be obtained.

Further, in the above description, nodes are described as words, but nodes are phonemes, phonemes, semi-syllables, syllables,
A unit such as a morpheme may be used, and the same effect can be obtained.

Although the speech recognition method has been described as an HMM, the same effect may be obtained by a DP (Dynamic Programming) matching or a speech recognition method using a neural network.

Embodiment 2 Embodiment 2 A speech recognition system according to Embodiment 2 of the present invention will be described with reference to the drawings. FIG. 5 is a block diagram showing a configuration of a speech recognition system according to Embodiment 2 of the present invention.

The speech recognition system according to the first embodiment has a disadvantage that the table size of the recording medium storage means 8 increases when the number of partial dictionaries increases. Here, as a method for solving this, a speech recognition system in which the following recording medium is stored in a partial dictionary will be described.

In FIG. 5, reference numeral 6 denotes a first dictionary storage unit that uses a recording medium that can be referenced at high speed, and 7 denotes a second dictionary storage unit that uses a recording medium that cannot be referenced at high speed.

The other parts are the same as those of the conventional speech recognition apparatus, and the description is omitted. Hereinafter, as in the first embodiment, the first dictionary storage unit 6 will be described as an EEPROM, and the second dictionary storage unit 7 will be described as a CD-ROM. However,
Unlike the first embodiment, the information of the recording medium of the partial dictionary to be connected next is recorded in the partial dictionary.

An example of the partial dictionary according to the second embodiment is
This is explained using net2 as an example. FIG. 6 shows the structure of the partial dictionary net2. Compared with the structure of the partial dictionary of the conventional speech recognition apparatus shown in FIG. 18, in FIG. 6, information on the recording medium storing the next partial dictionary is added to the CD-ROM. Based on this information, it is possible to determine which storage medium has the subsequent partial dictionary.

By adopting such a configuration, the recording medium storage means 8, which is required in the first embodiment, becomes unnecessary.

In the above description, the EEPROM and the CD-RO
M has been described as an example of a storage medium, but instead of an EEPROM, a flash ROM, an EPR
OM, ROM, RAM, etc. Also, instead of CD-ROM, CD-RW (rewritable CD), DVD-RO
M, DVD-RAM, DVD-RW (rewritable DV
D), a hard disk or the like may be used, and the same effect can be obtained.

The classification of the partial dictionaries used in the above description is merely an example, and other classification methods may be used.

In the above description, a method using two types of storage media has been described. However, three or more types of storage media may be used, and the same effect can be obtained.

Further, in the above description, the node has been described as a word, but the node is a phoneme, a phoneme, a half syllable, a syllable,
A unit such as a morpheme may be used, and the same effect can be obtained.

Although the speech recognition method has been described as an HMM, it may be a DP (Dynamic Programming) matching or a speech recognition method using a neural network, and the same effect can be obtained.

Embodiment 3 The configuration of the speech recognition system according to the third embodiment is the same as the configuration of the speech recognition system according to the second embodiment, and a description thereof will not be repeated.

The CD-ROM has a problem that it takes time for one reference. In the first and second embodiments, C
The number of partial dictionaries on the D-ROM increases, and CD-RO
There has been a problem that the number of references to M is increased. In order to avoid this problem, the third embodiment is characterized in that partial dictionaries having a strong dependence stored on the same recording medium are stored as a group, and are read out collectively when referred to.

In the case of address recognition, since the place name portion has a tree structure, the dependency of the partial dictionary is clear. For example, CD
FIG. 7 shows a dictionary in which place names under the same prefecture stored in the ROM are grouped into one group. In the drawing, a partial dictionary group surrounded by a chain line is represented as a group.

FIG. 8 is a diagram showing the structure of the partial dictionary net2. 6, group numbers grp3 and grp4 are added. Thereby, it is possible to know the group number of the subsequent partial dictionary. The same effect can be obtained by creating a table for storing independent group numbers as in the recording medium storage unit 8 of the first embodiment.

Next, the operation will be described. It is assumed that the partial dictionaries net1 and net2 are fetched on the collation data storage means 4. If the utterance is “Nagao-cho, Okawa-gun, Kagawa prefecture”, and the node “Kagawa prefecture” satisfies the condition of Expression 4, the group grp4 connected next is referred to.

FIG. 9 shows the contents of the collation data storage means 4 at this time. In the figure, a node surrounded by a solid line indicates that the node is activated, and a node surrounded by a broken line indicates that the node is deactivated. Since the group grp4 includes the partial dictionaries net7, net8, net10, and net11, they are read out collectively and a work area is set in the collation data storage means 4. However, even the nodes in the group grp4, the nodes other than “Shido-cho” and “Nagao-cho” do not meet the condition of Expression 4, so that no inter-node transfer occurs, so that they are not activated at this time.

If the recognition process proceeds further and the node of Okawa-gun meets the condition of the expression 4, the collation process for the partial dictionary net8 is necessary. However, since the partial dictionary net8 has already been incorporated in the work area, the second dictionary is renewed. It is not necessary to refer to the storage means 7 (CD-ROM).

With the above configuration, there is an effect that the partial dictionaries which will be required in the collation processing in the future are read out collectively, and the number of times of reference to the second dictionary storage means 7 (CD-ROM) can be reduced.

In the above description, the case where a CD-ROM is used as a storage medium has been described as an example.
CD-RW (rewritable CD), DVD-
A ROM, a DVD-RAM, a DVD-RW (rewritable DVD), a hard disk, or the like may be used, and the same effect can be obtained.

The classification of the partial dictionaries used in the above description is merely an example, and other classification methods may be used.

In the above description, a method using two types of storage media has been described. However, three or more types of storage media may be used, and the same effect can be obtained.

Further, in the above description, the node has been described as a word, but the node is a phoneme piece, a phoneme, a half syllable, a syllable,
A unit such as a morpheme may be used, and the same effect can be obtained.

Although the speech recognition system has been described as an HMM, a DP (Dynamic Programming) matching or a speech recognition system using a neural network may be similarly effective.

Embodiment 4 In the third embodiment, the number of times of reference to the second dictionary storage unit 7 (CD-ROM) can be reduced by grouping the partial dictionaries. There is a drawback that the amount of memory of the collation data storage means 4 is large because it is necessary to take the data. This is because the data format of the work area of the collation data storage unit 4 is considerably larger than the data format of the partial dictionary. In the fourth embodiment, by providing an intermediate storage area that can be referenced at high speed,
After the partial dictionaries of the group of the second dictionary storage means 7 are transferred, only the partial dictionaries referred to in the collation processing have a work area in the collation data storage means 4.

FIG. 10 is a diagram showing a configuration of a speech recognition system according to Embodiment 4 of the present invention.

In the figure, reference numeral 6 denotes a first dictionary storage means which uses a high-speed referable recording medium for storing partial dictionaries organized in groups, and 7 denotes a high-speed storage means for storing partial dictionaries organized in groups. A second dictionary storage unit 9 using a non-referenceable recording medium is a high-speed reference group storage unit that transfers and stores partial dictionary groups in groups from the second dictionary storage unit 7. Reference numeral 5 denotes a model matching unit that performs a matching process by individually referring to the partial dictionaries of the group storage unit 9.

Hereinafter, the second dictionary storage means 7 is stored in the CD-RO
M, the group storage means 9 will be described as a RAM.

Next, the operation will be described. During the recognition process, the node in Kagawa Prefecture satisfies the condition of Equation 4 and the group
Suppose grp4 is referenced. FIG. 11 shows the contents of the group storage means 9 and the collation data storage means 4 at this stage. In the figure, the upper part (a) shows the contents of the group storage means 9, and the lower part (b) shows the contents of the collation data storage means 4, respectively.

First, the partial dictionary group of the group grp4 is stored in the group storage unit 9 from the second dictionary storage unit 7 (CD-ROM).
(RAM). Subsequently, the group storage means 9
Only the necessary partial dictionary net4 on the (RAM) is referred to, and a work area is set in the collation data storage means 4.

With such a configuration, the amount of memory of the collation data storage unit 4 can be suppressed while the number of times of reference to the second dictionary storage unit 7 (CD-ROM) is reduced.

In the above description, RAM, EEPROM and C
Although the case where a D-ROM is used as a storage medium has been described as an example, a flash ROM, an EPROM, a ROM, a RAM, or the like may be used instead of a RAM or an EEPROM.
-CD-RW (rewritable CD) instead of ROM,
A DVD-ROM, a DVD-RAM, a DVD-RW (rewritable DVD), a hard disk, or the like may be used, and the same effects can be obtained.

The classification of the partial dictionaries used in the above description is an example, and other classification methods may be used.

In the above description, a method using three types of storage media has been described. However, four or more types of storage media may be used, and the same effect can be obtained.

Further, in the above description, the node has been described as a word, but the node is a phoneme, a phoneme, a half syllable, a syllable,
A unit such as a morpheme may be used, and the same effect can be obtained.

Although the speech recognition method has been described as an HMM, the same effect can be obtained by a DP (Dynamic Programming) matching or a speech recognition method using a neural network.

Embodiment 5 FIG. If the number of references to the CD-ROM or the transfer amount becomes too large during the recognition, the overhead of the recognition process increases, and the reaction time until a recognition result is obtained may be long. For this reason, in the fifth embodiment, the upper limit of the number of times of reference to the CD-ROM or the transfer amount is set to secure a specified reaction time.

Here, the number of times of reference to the group is measured,
A description will be given of an example of a voice recognition system that prohibits reference when a predetermined value is reached. The case where the upper limit value of the number of times of reference to the CD-ROM is specified will be described as an example.

FIG. 12 is a block diagram showing a configuration of a speech recognition system according to Embodiment 5 of the present invention. In the figure, reference numeral 10 denotes a transfer amount measuring means for measuring the number of times of reference of the second dictionary storage means 7. The other means are the same as those in the third embodiment, and the description is omitted here.

The collation operation between nodes has been performed under the condition of Expression 4, but when it is necessary to newly refer to the group in the second dictionary storage means 7 (CD-ROM), the condition expressed by Expression 6 below is used. Shall be performed.

Earc (n)> Ebestnode−Tgrp Equation 6

At the beginning of recognition, the threshold value Tgrp is calculated by the following equation (7).
Is the condition.

Tgrp = Tarc Equation 7

However, the second dictionary storage means 7 (CD-RO
If the number of references in M) exceeds the specified value, the following equation 8
Set as follows.

Tgrp = ∞ Equation 8

After being set as in the above equation 8, the equation 6
Is not satisfied, the reference request to the subsequent node does not occur, and the second dictionary storage unit 7 (CD-RO
The reference count of M) can be limited to a prescribed value.

The conditions for restricting reference to the second dictionary storage means 7 (CD-ROM) are as follows.
-ROM), the amount of transfer of the partial dictionary, or both the number of references and the amount of transfer of the partial dictionary can be limited, and the same effect can be obtained.

In the above description, the EEPROM and the CD-RO
M has been described as an example of a storage medium, but instead of an EEPROM, a flash ROM, an EPR
OM, ROM, RAM, etc. Also, instead of CD-ROM, CD-RW (rewritable CD), DVD-RO
M, DVD-RAM, DVD-RW (rewritable DV
D), a hard disk or the like may be used, and the same effect can be obtained.

The classification of the partial dictionaries used in the above description is merely an example, and other classification methods may be used.

In the above description, a method using two types of storage media has been described. However, three or more types of storage media may be used, and the same effect can be obtained.

Further, in the above description, the node has been described as a word, but the node is a phoneme, a phoneme, a half syllable, a syllable,
A unit such as a morpheme may be used, and the same effect can be obtained.

Although the speech recognition method has been described as an HMM, the same effect can be obtained by a DP (Dynamic Programming) matching or a speech recognition method using a neural network.

Embodiment 6 FIG. When the upper limit of the number of times of reference to the second dictionary storage means 7 (CD-ROM) is set as in the above-described fifth embodiment, if the group is incorrectly selected, there is a high possibility that recognition becomes impossible. In the case of "Kanagawa prefecture" and "Kagawa prefecture" shown as examples of the present invention, since only one syllable is different and similar, it is particularly prone to error. Therefore, in the sixth embodiment, after performing a part of the collation processing of the municipal name in addition to the prefecture name, the group stored in the second dictionary storage means 7 (CD-ROM) is selected. It is characterized in that the accuracy of the selection is increased. First dictionary storage means 6 which stores not only the prefecture name but also a part of the following municipal names as another partial dictionary
(EEPROM).

The configuration of the speech recognition system according to the sixth embodiment is the same as that described in the third embodiment, and a description thereof will not be repeated. Since the storage method of the first and second dictionary storage means 6 and 7 has the feature of the sixth embodiment, it will be described below.

FIG. 13 shows an example of the configuration of the dictionary according to the sixth embodiment. Second dictionary storage means 7 (C
The first two syllables of the groups grp3 and grp4 stored in the D-ROM) are stored in the first dictionary storage means 6 (EEPROM) as another partial dictionary net3h and net4h.

In FIG. 13, "Fujisawa City" is divided into "Fuji" and "Sawashi". Similarly, "Kamakura City", "Takamatsu City" and "Okawa-gun" also separate the first two syllables. FIG.
Then, the partial dictionaries stored in the first dictionary storage means 6 (EEPROM) are surrounded by solid lines, and the second dictionary storage means 7 (C
The partial dictionary stored in the D-ROM) is enclosed by a broken line. The partial dictionary net3h belongs to the group grp3h, and the partial dictionary net4h belongs to the group grp4h.

With such a configuration, “Kagawa Kawao” can be used in response to the utterance of “Nagao-cho, Okawa-gun, Kagawa”.
Since the group with the municipal name of “Kagawa Prefecture” can be selected using the information up to, the possibility of erroneously selecting grp3 which is the group with the municipal name of “Kanagawa Prefecture” can be reduced. As a result, there is an effect that the recognition rate can be improved.

In the above description, the first syllable of the group is stored in the first dictionary storage means 6 (EEPROM), but the present invention is not limited to the two syllables.

As shown in FIG. 14, the entire "Fujisawa City", "Kamakura City", "Takamatsu City" and "Okawa-gun" may be separated and stored in the first dictionary storage means 6 (EEPROM). ,
It works similarly.

In the above description, the EEPROM and the CD-RO
M has been described as an example of a storage medium, but instead of an EEPROM, a flash ROM, an EPR
OM, ROM, RAM, etc. Also, instead of CD-ROM, CD-RW (rewritable CD), DVD-RO
M, DVD-RAM, DVD-RW (rewritable DV
D), a hard disk or the like may be used, and the same effect can be obtained.

The classification of the partial dictionary used in the above description is an example, and other classification methods may be used.

In the above description, a method using two types of storage media has been described. However, three or more types of storage media may be used, and the same effect can be obtained.

Further, in the above description, the node has been described as a word, but the node is a phoneme, a phoneme, a half syllable, a syllable,
A unit such as a morpheme may be used, and the same effect can be obtained.

Although the speech recognition method has been described as an HMM, the same effect can be obtained by a DP (Dynamic Programming) matching or a speech recognition method using a neural network.

[0131]

As described above, the speech recognition system according to the first aspect of the present invention comprises: a sound analysis unit that inputs a speech signal, performs sound analysis, converts the signal into a time series of feature vectors, and outputs the time series; Standard model storage means for storing a standard model of the above, dictionary storage means for storing a plurality of partial dictionaries obtained by dividing the word dictionary, collation data storage means for storing collation data as a work area in the collation processing, and the acoustic analysis means In the speech recognition system including a matching process for performing a matching process on the feature vector from the reference model with reference to the standard model and the word dictionary and outputting a recognition result, the dictionary storage unit can be referred to at high speed. First dictionary storage means for storing a partial dictionary frequently used in the first dictionary, and a second dictionary storage means for storing remaining partial dictionaries which cannot be referred to at high speed and remain infrequently used Since it is configured from a, an effect that it is possible to shorten the recognition response time while keeping costs low.

In the speech recognition system according to a second aspect of the present invention, as described above, whether the dictionary storage unit stores a partial dictionary in the first dictionary storage unit or the second dictionary storage unit Since the recording medium storage means for storing the recording medium information indicating the above is further included, it is possible to shorten the recognition response time while keeping the cost low.

As described above, in the speech recognition system according to claim 3 of the present invention, the first dictionary storage means and the second dictionary storage means correspond to each node, and Stores a partial dictionary including recording medium information indicating which of the first dictionary storage unit and the second dictionary storage unit is stored, so that it is possible to shorten the recognition response time while keeping costs low. It works.

In the speech recognition system according to a fourth aspect of the present invention, as described above, the first dictionary storage unit and the second dictionary storage unit collectively combine the partial dictionaries having a strong dependency on the same recording medium. When the model matching unit refers to a certain partial dictionary, the group including the partial dictionary is read out collectively,
Since the collation processing is performed, the recognition response time can be shortened while keeping the cost low, and the number of times of referring to each storage means can be reduced.

As described above, in the speech recognition system according to the fifth aspect of the present invention, the dictionary storage means stores a group including a partial dictionary transferred together when referring to the partial dictionary. A storage unit, wherein the model matching unit performs the matching process by individually referring to the partial dictionaries in the group stored in the group storage unit, so that the recognition response time is shortened while keeping costs low. This makes it possible to reduce the number of references to each storage means.

In the speech recognition system according to the sixth aspect of the present invention, as described above, the dictionary storage means transfers the partial dictionary read from the second dictionary storage means to the collation data storage means. The apparatus further comprises an amount measuring unit, wherein the model matching unit compares the amount measured by the transfer amount measuring unit with a specified amount and controls not to read a partial dictionary of the specified amount or more. This has the effect of shortening the response time.

In the speech recognition system according to a seventh aspect of the present invention, as described above, the first dictionary storage means includes:
The second dictionary storage unit stores a partial dictionary consisting only of the head portion of the group to be stored as a new group, and the second dictionary storage unit stores the partial dictionary consisting of the remainder left out of the head portion. The dictionary is stored as a new group, and when the model matching unit refers to a certain partial dictionary, the group including the partial dictionary is collectively read and the matching process is performed, so that the recognition rate can be improved. To play.

As described above, in the speech recognition method according to the eighth aspect of the present invention, a speech signal is input and subjected to acoustic analysis to convert the feature vector into a time series of feature vectors. In a voice recognition method of performing a matching process with reference to a plurality of partial dictionaries obtained by dividing a model and a word dictionary and outputting a recognition result, in the matching process, first, a partial dictionary that can be referred to at high speed and that is frequently used is used. A first partial dictionary reading step of reading a partial dictionary from the first dictionary storage means to be stored; and a second dictionary storage for storing the remaining infrequently inaccessible partial dictionaries in the matching process. Means for reading the partial dictionary from the first dictionary storage means or the first dictionary storage means, so that the recognition response time can be shortened while keeping the cost low. There is an effect that kill.

As described above, in the speech recognition method according to the ninth aspect of the present invention, in the second partial dictionary reading step, the next partial dictionary to be connected is the first dictionary storage unit or the second dictionary. Since the partial dictionary is read from the first dictionary storage unit or the second dictionary storage unit based on the storage medium information indicating which storage unit is stored, the recognition response time can be reduced while keeping the cost low. This has the effect that it can be performed.

As described above, in the speech recognition method according to the tenth aspect of the present invention, in the second partial dictionary reading step, when referring to a certain partial dictionary, a partial dictionary having a strong dependency is grouped together. Since the group including the partial dictionary is collectively read from the second dictionary storage unit that is stored as, the recognition response time can be shortened while the cost is kept low, and the number of references to the storage unit can be reduced. To play.

As described above, in the voice recognition method according to the eleventh aspect of the present invention, in the second partial dictionary reading step, when a certain partial dictionary is referred to, the second partial dictionary storage means reads the partial dictionary. Are collectively read and stored in the group storage means, so that the recognition response time can be shortened while the cost is kept low, and the number of references to the storage means can be reduced.

According to a twelfth aspect of the present invention, a computer-readable recording medium storing a speech recognition program receives a speech signal, performs acoustic analysis, converts the signal into a time series of feature vectors, and outputs the result. Acoustic analysis procedure, a standard model storage area for storing a standard model to be recognized, a dictionary storage area for storing a plurality of partial dictionaries obtained by dividing a word dictionary, and collation data storage for storing collation data as a work area in the collation processing A computer-readable recording of a speech recognition program including a region and a model matching procedure for performing a matching process on the feature vector from the acoustic analysis procedure with reference to the standard model and the word dictionary and outputting a recognition result In a simple recording medium, the dictionary storage area stores a partial dictionary that can be referred to at high speed and that is frequently used. Since it is composed of one dictionary storage area and a second dictionary storage area that stores the remaining partial dictionaries that cannot be referenced at high speed and that are used less frequently, the recognition response time can be shortened while keeping costs low. To play.

According to a thirteenth aspect of the present invention, as described above, the computer-readable recording medium recording the speech recognition program is such that the dictionary storage area stores the partial dictionary in the first dictionary storage area or the second dictionary storage area. Since the recording medium storage area for storing the recording medium information indicating which of the storage areas is stored is further included, an effect is obtained that the recognition response time can be shortened while keeping costs low.

According to a fourteenth aspect of the present invention, there is provided a computer readable recording medium storing a speech recognition program, wherein the first dictionary storage area and the second dictionary storage area correspond to each node, as described above. Since a partial dictionary including recording medium information indicating whether the next partial dictionary is stored in the first dictionary storage area or the second dictionary storage area is stored, recognition can be performed while keeping costs low. This has the effect of shortening the response time.

According to a computer-readable recording medium having recorded thereon a speech recognition program according to claim 15 of the present invention, as described above, the first dictionary storage area and the second dictionary storage area are stored on the same recording medium. Memorize a partial dictionary with a strong dependency as a group,
When the model matching procedure refers to a certain partial dictionary, a group including the partial dictionary is read out collectively and the matching process is performed, so that the recognition response time can be shortened while keeping costs low, and the There is an effect that the number of times of reference can be reduced.

As described above, in the computer-readable recording medium storing the speech recognition program according to claim 16 of the present invention, when the dictionary storage area refers to a certain partial dictionary, The apparatus further includes a group storage area for storing a group including a partial dictionary, and the model matching procedure performs a matching process by individually referring to partial dictionaries in the group stored in the group storage area. , The recognition response time can be shortened, and the number of times of referring to each storage area can be reduced.

The computer-readable recording medium having recorded thereon the speech recognition program according to claim 17 of the present invention, as described above, stores the amount of the partial dictionary read from the second dictionary storage area to the collation data storage area. The method further includes a transfer amount measurement procedure for measuring, and the model matching procedure controls the comparison between the measurement amount and the specified amount by the transfer amount measurement procedure so as not to read a partial dictionary of the specified amount or more, thereby reducing costs. There is an effect that the recognition response time can be shortened while suppressing it.

According to a computer-readable recording medium having recorded thereon a speech recognition program according to claim 18 of the present invention, as described above, the first dictionary storage area is a group to be stored in the second dictionary storage area. The second dictionary storage area stores a partial dictionary composed of the remainder of the first part as a new group, and stores the partial dictionary composed only of the first part of the model dictionary as a new group. When the procedure refers to a certain partial dictionary, a group including the partial dictionary is read out collectively and collation processing is performed, so that the recognition rate can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a speech recognition system according to Embodiment 1 of the present invention.

FIG. 2 is a diagram showing a configuration of a word dictionary of the speech recognition system according to the first embodiment of the present invention.

FIG. 3 is a diagram showing contents of a recording medium storage means of the voice recognition system according to the first embodiment of the present invention.

FIG. 4 is a diagram showing another configuration of the word dictionary of the speech recognition system according to the first embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a speech recognition system according to Embodiment 2 of the present invention.

FIG. 6 is a diagram showing a structure of a partial dictionary of the speech recognition system according to Embodiment 2 of the present invention.

FIG. 7 is a diagram showing a configuration of a word dictionary of a speech recognition system according to Embodiment 3 of the present invention.

FIG. 8 is a diagram showing a structure of a partial dictionary of the speech recognition system according to Embodiment 3 of the present invention.

FIG. 9 is a diagram showing contents of a collation data storage means of the voice recognition system according to Embodiment 3 of the present invention.

FIG. 10 is a block diagram showing a configuration of a speech recognition system according to Embodiment 4 of the present invention.

FIG. 11 is a diagram showing contents of a group storage unit and a collation data storage unit of the voice recognition system according to the fourth embodiment of the present invention.

FIG. 12 is a block diagram showing a configuration of a speech recognition system according to Embodiment 5 of the present invention.

FIG. 13 is a diagram showing a configuration of a word dictionary of a speech recognition system according to Embodiment 6 of the present invention.

FIG. 14 is a diagram showing another configuration of the word dictionary of the speech recognition system according to Embodiment 6 of the present invention.

FIG. 15 is a block diagram showing a configuration of a conventional speech recognition system.

FIG. 16 is a diagram showing addresses to be recognized in a conventional voice recognition system.

FIG. 17 is a diagram showing a configuration of a word dictionary of a conventional speech recognition system.

FIG. 18 is a diagram showing a structure of a partial dictionary of a conventional speech recognition system.

FIG. 19 is a diagram showing the contents of a standard model storage means of a conventional speech recognition system.

FIG. 20 is a diagram showing contents of a collation data storage means of a conventional speech recognition system.

FIG. 21 is a diagram showing the contents of a collation data storage means of a conventional speech recognition system.

FIG. 22 is a flowchart showing the operation of a model matching unit of a conventional speech recognition system.

FIG. 23 is a diagram showing the contents of the HMM of the collation data storage means of the conventional speech recognition system.

FIG. 24 is a diagram showing the contents of an HMM in a collation data storage means of a conventional speech recognition system.

[Explanation of symbols]

1 acoustic analysis means, 2 standard model storage means, 4 collation data storage means, 5 model collation means, 6 first dictionary storage means, 7 second dictionary storage means, 8 recording medium storage means,
9 Group storage means, 10 Transfer amount measurement means.

Claims (18)

[Claims] 1. A sound analysis means for inputting a sound signal, performing sound analysis, converting the sound signal into a time series of feature vectors and outputting the time series, a standard model storage means for storing a standard model to be recognized, and a plurality of divided word dictionaries. Dictionary storage means for storing partial dictionaries, collation data storage means for storing collation data as a work area in the collation processing, and referring to the standard model and the word dictionary for the feature vector from the acoustic analysis means. A speech recognition system comprising: a model matching unit that performs a matching process and outputs a recognition result; wherein the dictionary storage unit is a first dictionary storage unit that stores a partial dictionary that can be referred to at high speed and that is frequently used; And a second dictionary storage means for storing remaining incompletely used partial dictionaries that cannot be referred to. 2. The dictionary storage unit further includes a storage medium storage unit that stores storage medium information indicating whether the partial dictionary is stored in the first dictionary storage unit or the second dictionary storage unit. The speech recognition system according to claim 1, wherein: 3. The first dictionary storage means and the second dictionary storage means, corresponding to each node, wherein a partial dictionary to be connected next is either the first dictionary storage means or the second dictionary storage means. 2. The speech recognition system according to claim 1, wherein a partial dictionary including recording medium information indicating whether the information is stored in the storage unit is stored. 4. The first dictionary storage unit and the second dictionary storage unit store, as a group, partial dictionaries having a strong dependency on the same recording medium, and the model matching unit stores a certain partial dictionary. 4. The speech recognition system according to claim 3, wherein when referring to, the group including the partial dictionary is read out collectively and a collation process is performed. 5. The dictionary storage unit further includes a group storage unit configured to store a group including the partial dictionary transferred collectively when referring to the partial dictionary, and the model matching unit includes the group storage unit. 5. The speech recognition system according to claim 4, wherein the matching processing is performed by individually referring to the partial dictionaries in the group stored in the means. 6. The dictionary storage unit further includes a transfer amount measurement unit that measures an amount of a partial dictionary read from the second dictionary storage unit to the comparison data storage unit, and the model comparison unit includes the transfer unit. 5. The speech recognition system according to claim 4, wherein the amount measured by the amount measuring means is compared with a prescribed amount, and control is performed so that a partial dictionary having the prescribed amount or more is not read. 7. The first dictionary storage means stores, as a new group, a partial dictionary composed only of a head portion of a group to be stored in the second dictionary storage means, wherein the second dictionary storage means Storing a partial dictionary composed of the remainder of the leading part as a new group, wherein the model matching unit reads a group including the partial dictionary collectively when referring to a certain partial dictionary, and performs a matching process. 5. The speech recognition system according to claim 4, wherein the speech recognition is performed. 8. A speech signal is input, acoustic analysis is performed, and the speech signal is converted into a time series of feature vectors.
In a voice recognition method of performing a matching process with reference to a plurality of partial dictionaries obtained by dividing a standard model and a word dictionary to be recognized, and outputting a recognition result, in the matching process, first, a high-speed reference and a usage frequency A first partial dictionary reading step of reading a partial dictionary from a first dictionary storage unit that stores a high partial dictionary; and in the collation processing, next, a remaining partial dictionary that cannot be referred to at high speed and is used less frequently Reading a partial dictionary from the second dictionary storage means or the first dictionary storage means;
And a sub-dictionary reading step. 9. In the reading step of the second partial dictionary,
Next, the first dictionary storage unit or the second dictionary storage unit is based on recording medium information indicating which of the first dictionary storage unit and the second dictionary storage unit the partial dictionary to be connected to is stored. 9. The speech recognition method according to claim 8, wherein a partial dictionary is read out of the dictionary. 10. The second partial dictionary reading step includes, when referring to a partial dictionary, the partial dictionary from a second dictionary storage unit that stores a partial dictionary having a strong dependence as a group. The speech recognition method according to claim 9, wherein the groups are read out collectively. 11. The second partial dictionary reading step includes, when referring to a certain partial dictionary, reading out a group including the partial dictionary from the second dictionary storage unit and storing the group in the group storage unit. The voice recognition method according to claim 10, wherein 12. A sound analysis procedure for inputting a sound signal, performing sound analysis, converting the sound signal into a time series of feature vectors, and outputting the time series, a standard model storage area for storing a standard model to be recognized, and a plurality of divided word dictionaries. A dictionary storage area for storing partial dictionaries, a collation data storage area for storing collation data as a work area in the collation processing, and a feature vector from the acoustic analysis procedure with reference to the standard model and the word dictionary. A computer-readable recording medium storing a speech recognition program including a model matching procedure for performing a matching process and outputting a recognition result, wherein the dictionary storage area stores a partial dictionary that can be referred to at high speed and that is frequently used. The first dictionary storage area and the second dictionary storage area for storing the remaining infrequently used partial dictionaries that cannot be referenced at high speed. A computer-readable recording medium having recorded thereon a voice recognition program. 13. The dictionary storage area further includes a storage medium storage area for storing storage medium information indicating whether the partial dictionary is stored in the first dictionary storage area or the second dictionary storage area. A computer-readable recording medium on which the voice recognition program according to claim 12 is recorded. 14. The first dictionary storage area and the second dictionary storage area correspond to each node, and a partial dictionary to be connected next is either the first dictionary storage area or the second dictionary storage area. 13. A computer-readable recording medium storing a speech recognition program according to claim 12, wherein a partial dictionary including recording medium information indicating whether the speech recognition program is stored is stored. 15. The first dictionary storage area and the second dictionary storage area store, as a group, partial dictionaries having a strong dependency relationship on the same recording medium, and the model collation procedure includes: 15. The computer-readable recording medium storing a speech recognition program according to claim 14, wherein a group including the partial dictionary is read out at a time when the reference is made, and a collation process is performed. 16. The dictionary storage area further includes a group storage area for storing a group including the partial dictionary transferred collectively when the partial dictionary is referred to, and wherein the model collation procedure includes the step of: 16. The computer-readable recording medium according to claim 15, wherein the collation processing is performed by individually referring to the partial dictionaries in the group stored in the area. 17. The method according to claim 17, further comprising: a transfer amount measuring step of measuring an amount of the partial dictionary read from the second dictionary storage area to the comparison data storage area; 16. A computer-readable recording medium storing a speech recognition program according to claim 15, wherein a control is performed so as not to read the partial dictionary having the prescribed amount or more. 18. The first dictionary storage area stores, as a new group, a partial dictionary composed only of a leading part of a group to be stored in the second dictionary storage area, And storing a partial dictionary composed of the remainder from which the leading part is omitted as a new group. In the model matching procedure, when referring to a certain partial dictionary, the group including the partial dictionary is read out collectively, and the matching process is performed. 16. A computer-readable recording medium on which the voice recognition program according to claim 15 is recorded.