JPH01502610A - Continuous speech recognition system - Google Patents

Abstract

Method and arrangement for speech recognition for use in a speech recognition system, wherein a grammar model may be stored in memory, the grammar model being composed of nodes connected by arcs, each arc generally having an associated template prestored in memory and respective originating and terminating grammar nodes, and wherein groups of input frames may be compared to templates to generate similarity measure parameters. The invention includes structuring the grammar model such that arcs may be looped within the structure. For implementing this type of structure, the invention includes providing first parameter storage and second parameter storage for one or more nodes including a terminating node for a selected arc, determining similarity measure parameters for the originating node for the selected arc, determining similarity measure parameters for the arc using a template associated with the arc and the originating node's similarity measure parameters. Further, similarity measure parameters are determined at the terminating node and are stored in the first parameter storage for the terminating node. The first parameter storage contents are transferred to the second parameter storage for the terminating node, and the groups of input frames are recognized using the parameter storage contents associated with the terminating node.

Description

[Detailed description of the invention] continuous speech recognition system Background of the invention TECHNICAL FIELD The present invention relates to speech recognition systems, and more particularly, to the endpoint (e) of a spoken word (word). This invention relates to the recognition of speech whose nd, polnt) are not predetermined.

isolated from a given grammar ((rammar) for a known speaker). It has been known for a long time that humans can recognize words that are set up.

Memorize grammar words in advance as individual templates and use them for each test. Let templates represent sound patterns for words using grammar. isolated When a word is spoken, the system converts the word into each individual template representing the grammar. Compare with. This method is generally used as the exact word template matching method (who!e-wo rd tesplaie matching). give good results Most recognition systems use complete word template matching using dynamic programming. is adopted to detect non-linearity between spoken words and pre-memorized templates. It deals with changes in the shape time scale.

Although this method is effective for recognizing isolated words, it is not suitable for many practical applications. requires recognition of consecutive words. In continuous word recognition, the number of words in a phrase is There is no need to limit, the body of the word at the beginning can be determined before the end of the phrase. However, in isolated word recognition, a delimiter is used to identify the beginning and end of an input pattern. Recognition is done one word at a time using deliciters. Furthermore, the series The continuous speech recognition system recognizes incoming patterns from other recognizable patterns, background noise, and breathing. It must be distinguished from the noise coming from the speaker, such as background noise, while isolated recognition is Usually cannot tolerate other recognizable patterns at the beginning or end of words.

IEEE) Analysis, Acoustics, Speech and Signal Processing, vol, ASSP -27, No, 8. pp, sgg ~595 (December 1979) H, 5 ``Two-level DP matching method - Connected word recognition'' by akoe Dynamic Programming Based on Pattern Matching Algorithm for Complete Words The template matching method has been extended to handle connected word recognition. this theory Sentences are created using two passes ( two-pass) dynamic programming algorithm. first In the path, each text is matched against each possible part of the input pattern. Generates a score indicating the similarity between templates. In the second pass, this score is to find the best template sequence that corresponds to the entire input pattern.

This expanded method has obvious drawbacks. One drawback of this technique is that the required is the amount of computing time. Depending on your specific design requirements, this limit may make expensive high-speed Sessa may be unreasonably needed.

Another disadvantage of this method is that the end point of the input pattern must be determined in advance; The entire turn must be stored in the system until an exact template match occurs. This means that it must be done. If the input pattern is quite long, the recognition response The gap actually gets worse. In addition, the error in detecting the end point can seriously degrade the performance of the recognizer. let Furthermore, the memory required to store this information can be prohibitive. .

IEEE) Analysis, Acoustics, Speech and Signal Processing, vol, ASSP -27, No, 6. pp, 588-595 (December 1979), Brown, J, 5pohrers P, Rochschild, “Partial traceback” by Baker, J. Dynamic programming allows you to program inputs of arbitrary length without predetermining the end point. A technique considering continuous speech recognition of turns is described. This is a partial trace It is done using a technique called backing. For partial traceback, the recognizer words recognized before the completion of a complete input pattern without sacrificing performance. is output. However, the described partial traceback technique It seems to be a burden on the user and difficult to implement.

Therefore, it is easy to implement and works effectively and cheaply in real time. There is a need for a continuous speech recognition system that can perform continuous speech recognition.

Purpose and outline of the invention The purpose of the present invention is to implement it for real-time applications and to recognize continuous speech using inexpensive hardware. It is an object of the present invention to provide a speech recognition mechanism and method that can perform speech recognition.

Still another object of the present invention is to effectively manage speech recognition memory during the recognition process. An object of the present invention is to provide a speech recognition mechanism and method that can perform speech recognition.

Still another object of the present invention is to provide a speech recognition mechanism and method that provides a grammar with loops. It is to provide.

Briefly, the present invention provides a speech recognition method and mechanism for use in a speech recognition system. , the grammar model is stored in memory in advance, and this Grammar models are pre-stored in memory and associated templates. ate) and their respective originating nodes. ) and a terminal node. ), and the input frame group is a template. is compared with the plate to generate a similarity measure parameter. The present invention has an arc structure. Contains operations to construct a grammar model such as drawing a loop within. This kind of structure To achieve this, the present invention provides one or more The first parameter storage device and the second parameter storage device for the node of setting the parameters related to the arc at the end node for that arc. The similarity measure parameter is calculated using the sample template and the similarity measure parameter determined earlier. It includes being able to understand. Furthermore, the similarity parameter determined at the end node is is stored in the first parameter storage for the parameter. First parameter storage device The contents of the input frame are moved to the second parameter store for the destination node and Groups are recognized using the contents of the parameter store for the destination nodes.

Brief description of the drawing Features of the invention believed to be novel are included in the claims with specificity. This is an indication. The invention, among other objects and advantages, comprises the following: Although best understood by referring to the description, the same Reference numbers indicate the same elements.

Figure 1 is a hardware block diagram of a speech recognition system according to the present invention. It is a diagram.

FIG. 2 shows recognition illustrating one aspect of a speech recognition system implemented in accordance with the present invention. This is a diagrammatic representation of a grammar model.

Figure 3 shows all possible paths (paths) according to the speech grammar model in Figure 1. s) is a diagrammatic representation of a phonetic grammar tree that enumerates the following.

FIG. 4 is a diagrammatic representation of a grammar model according to the invention.

Figures 5a, 5b and MSc illustrate the implementation of the recognition process according to the invention. 1 is a flowchart depicting a series of steps performed in a process.

FIG. 6 is a flowchart illustrating block 72 of FIG. 5C in more detail.

Figures 7a, 7b, 7c, and 7d combine block 44 of figure 5a. FIG. 3 is a flowchart showing layer details; FIG.

FIG. 8 is a series of grammar tree diagrams illustrating an example of "traceback" according to the present invention. Ru.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS Referring to FIG. 1, a speech recognition system that can be used to implement the present invention is shown. A block diagram of the system is shown.

The block diagram shows a template memory 10 containing pre-stored grammars. Contains. Typical pre-stored grammar formations can be found in Journal of the Acoustical Society of America, 6. g (5), November 1980 Nori, R, Rablner, J, G, Wilpon, “A Simplified Powerful Training Procedure for Trained Speakers, Isolated Words” Recognition System”. I on Acoustics, Speech, and Signal Processing EEE) Ranzakjon, vol, ASSP-31゜PP, 793-80 B (August 1983), B, A, Dautrich, L, R.

Rablners and T. B. Martin, ``Field of isolated word recognition''. ``On the effects of varying the router bank parameters''. An audio processor 12, such as the audio segments). Each frame is A time segment of the input audio, typically of LPG or filter bank data. Expressed in form. Frames from the acoustic processor are passed to recognizer 14 .

The recognizer 14 selects words from a grammar previously stored in the template memory 10. access the template and convert each input frame from the sound processor 12 into a word template. process using segments of the template. Such techniques are used in many speech recognition systems. It is unique to the stem and can be referred to as "template processing."

The recognizer 14 has a link table memory 16 and a node table memory 19. Access the two tables bidirectionally. The link memory table 16 is 5 used to store related arrays. The node table memory 19 is Used to store parameters related to the legal model. on these tables This will be further explained below along with the grammar model.

The recognizer 14 includes two processors: a recognition processor 18 and a link tray. This can be realized using the Subac processor 20. recognition processor 1 8 includes template matching, grammar, control, and link traceback ψ processes. handles all communications with the server 20; Link Traceback - Processor 20 is used to maintain memory for the link contention table.

This feature allows you to record possible template matches while inputting continuous audio, storing link information in the link table memory 16; 16 space for other information, and if the input audio is This includes outputting recognition results as the process progresses. Recognition processor 18 and link Functions with raceback processor 20 can be combined into one processor or Alternatively, the recognition processor 18 can be separated as shown, thereby making the recognition processor 18 IEEE International Conference on Science, Speech and Signal Processing A:i Proceedings pp, 899- 902 (1982) J, Br1dles M, Brown.

and R. Chamberlain, “Algorithms for Connected Word Recognition” in detail. It can be implemented as described. Link traceback process When used in accordance with the present invention, a processor such as the Motorola MC6801 This can be realized using a similar 8-bit processor.

Grammar modeling Now referring to Figure 2, we can see all the possible units that the system can recognize. A simplified recognition grammar model showing word sequences is shown. This model is The grammar provided is for illustrative purposes and is severely limited from what is generally needed. Therefore, it is said to be ``simplified''. Figure 2 shows six possibilities, each consisting of two words. There are word strings (further explained in Figure 3). Typical speech recognition system In the system, the grammar model can be used to generate more possible word strings, each containing a Park) can be provided. The topology of the grammar model is , a network of interconnected arcs, each having a source node and a destination node. It is stored further in memory as a archive. Each arc has a corresponding one in template memory There may also be one or more pointers to templates to be used.

Modeling the grammar bus In Figure 3, each of the six possible word strings from Figure 2 is plotted in a tree diagram. are listed. Three possible first words, rONEJ (one), rTWOJ ( ) Wu), and rTHREEJ (three) exist. For each possible first word Two possible second words rFOURJ (four) and rFIVEJ (five) ) is followed. During template matching, i.e. if the input frame While being compared to the stored word templates, the recognition processor selects possible Recognize the terminus. Possible "word endings" are when a sequence of input frames is a word template. If there is a possibility of a match, it will be found. The identified word templates are link information stored in the input link table and the set of inputs being processed. The cumulative distance and the measure of similarity between the force frame and the template leading to the node. is added to the tree diagram through For example, the grammar strings in Figures 2 and 3 It is possible that the sequence of input frames matches the word rTWOJ. Once identified as possible, “TWO” becomes the first node, i.e. It is added to the tree diagram from step 24. Figure 3 shows another frame being input and processed. , indicating that the word rONEJ is a possible match. Therefore This is then also added to the tree diagram at node 24. Then the word rTHREEJ is added to the tree diagram, then rFOURJ is added to node 26, and then the same rF IVEJ is added to the same node, and so on. This is for each possible case. Continue to add these to the tree diagram as matching templates are identified. It will be done.

The term "tree node" or a reference to a node in a tree diagram is a link - Interchangeable with the word record (link (ed) record) I decide to use it. In general, link records specify connections in a tree diagram It is a data collection (data 5et) stored in memory, and it is - Contains the identification of a node and its relationship to previous nodes in the tree topology.

A new entry, or link Records are added to the link table corresponding to the links in the representative tree diagram. . Very often, a series of frames (representing a state), typically in the form of a state diagram, and As the input frame is processed, the word template, represented as There are multiple possibilities for the end of a word. Possible end of word detected each time the corresponding template is added to the tree as a new link. . Furthermore, each state of each template represents the storage processed through the current input frame. The product distance and the tree link from which decoding begins for that template. Record a link pointer pointing to a link record in the corresponding link table are doing. For further information on template matching, please refer to Acoustics, Audio , and Proceedings of the IEEE International Conference on Signal Processing, pp. 899-902. 19F12 J, Br1dle, M, Brown and RoChambe You can refer to rlain's "Algorithm for Connecting Word Recognition".

Unfortunately, large grammars cause problems from constantly appending templates to the tree. arise. First, this delays recognition response time. Sequence of input frames The longer the sequence, the more the operator can use the words the system recognizes. The amount of time you have to wait for processing increases.

Second, templates are constantly being added for linking tree diagram information. Requires a large amount of memory. The grammar model is complex and for each possible word If several possible end-of-word frames match, the link table is The memory required for bulls (trees) grows at a very rapid rate. As it gets bigger If the ratio is too large, the memory requirements become impractical.

The third problem has to do with actually modeling the grammar. Applies to continuous audio It is desirable to be able to recognize indefinitely long connected word sequences when is necessary.

A model structure with such a word sequence is called a "non-separable fixed length model". Make it. As mentioned above, the grammar model in Figure 2 is somewhat simplified. audio For those skilled in the art of certification, it is not possible to represent an infinitely long sequence as a model in memory. It should be clear that it is impractical to try to do so. Moreover, the model Computing segments is impractical for real-time processing applications. These reasons Therefore, the present invention allows infinite-length models to be created with as few computational and memory requirements as possible. It is designed to be accommodated without having to do so.

infinite length model Now referring to Figure 4, an infinitely long sequence can be stored in memory and in computation time. A diagrammatic representation of a particular grammatical model that can be accommodated with as little as possible is drawn. There is. Those skilled in the art will understand that a given grammar requires specific modeling criteria. It should be understood. Therefore, the example shown in Figure 4 accommodates the practically necessary grammar. It should be obvious that it can be modified to do so.

The grammar can recognize any number from 0 to 9, and any of them can be designated as a silent period. (period or 5 ilence) may or may not be separated. stomach. In this model, the input to the recognition process ends with the word rstopJ. Ta For example, the sequence r0-l-1-9-8-0-4-stopJ is optionally It can be recognized by the dash that represents the sound.

Two characteristics of this type of grammar model include null arcs (null arcs) between nodes. c) and loops within the grammar model. invalid arc A network is essentially a virtual connection between two nodes. This will cans or words without duplicating the arcs representing these words in the grammar model. can be recognized. For example, at node 27, disable the word rstopJ If you want to recognize without using an arc, you must first detect silence. Must be. If you do not need silence before the word rstopJ, use the grammar model The arc representing rstopJ must also be modified to exit from node 27. No. Therefore, an invalid arc allows a characteristic that ends at the starting node of the invalid arc to A given arc is replicated from its starting node to the terminal node of the invalid arc. It disappears.

The second advantage that this type of grammar model provides is the loop itself. Allow loops By doing this, you can avoid duplicating template expressions in the grammar model. I can do it.

For example, we have to recognize the word sequence r　0-O-0-1-3-9J If not, recognize the progression through the word model without duplicating ``0'' three times within the model. can do.

To accommodate this kind of word model, the present invention introduces special parameters. . These parameters include, for each node, the current cumulative distance, the beginning of its cumulative distance, The current link as a point, the previously accumulated distance, and the starting point of the previously accumulated distance. Contains previous links. These parameters are stored in the node table , can be constructed as follows.

node table The node table contains input frame arcs (input rra■ Used to temporarily store information on each node when processing . Arcs leading to nodes can be invalid arcs and have associated templates. It can also be an arc.

Two sets of entries in the table are assigned to each node. 1 II The entries in i are the previously accumulated distances and their associated links. This information is temporary used to determine the cumulative distance for the "best" arc to the node. can do.

The second set of entries are the current cumulative distances and their associated links. Once on the node Once all leading arcs have been processed for a given input frame, the previous parameter' is copied to the 'current' parameter. That is, from the two columns on the left information is copied into the two columns on the right, which gives the best cumulative distance leading to the linode, and A link to the starting arc is maintained. Also, the cumulative distance in the previous column is and become inactive. For the node table, see Figure 5a and Figure 5b. This will be further explained using a recognition flowchart.

link table The link table is configured to fit the tree network shape similar to Figure 2. It represents all possible word sequences under consideration. word sequence In fact, strings with possible word endings detected during the template matching period are This is a combined template. Setting up the network in this way These links are distinct parts of any more likely word sequence. can be analyzed. This analysis process can be traced back (traceback). k). If used properly, tracebacks can make things no longer obvious. An efficient way to free link records that are no longer part of the sequence being monitored. method becomes possible.

Each link in a tree diagram or node-style tree connection stores several types of information. I have to keep it. This information is stored in memory L of link table 16 in FIG. -ACT.

Stored in L-FWRD, L-BACK, L-WORD, and L-PTR arrays be done. In this example, each array is 255 bytes long with a 256-byte buffer. It is placed one byte past the boundary to allow efficient access. It has become. Corresponding elements from each array constitute a "link record". Ru. Linked records are chained into a two-linked list. 1 A list of free link records, that is, free records available for additional links. It has a code space. The second list is the list you have configured and is currently in use. Contains a record of links used. These lists are for L-PTR arrays. are chained together by i. In this case - one entry in the PTR returns the next linked record in the table from the set or free list. As shown, each record contains one byte from each of the five arrays. parable For example, for a given link record in the configuration list, the L-PTR array's correspondence the next record in the configuration list if the byte to be will be present in the second byte of all five arrays. L-PTR array The "0" entry in defines the end of the linked list.

The L-BACK array and L-WORD array contain actual link information.

L-BACK is the previous link in the decoding path, i.e. the previous node in the tree diagram. The L-WORD has a pointer pointing to the current link, while the L-WORD has a pointer pointing to the current link. It has symbols to represent the coded words. For example, in Figure 3, the tree ・After adding the word “FOURJ” to node 26, L-WORD becomes the word “rFOU” with an 8-bit symbol representing RJ, and L-BACK corresponds to tree node 26. A pointer to the linked record will be provided. Two other arrays , L-ACT and L-FWRD are “traceback” (possible) through the decoding path. possible word sequences). L-ACT is a possible match is used to indicate the decoding path still under consideration, and L-FWRD is is used to point to the successor node within, i.e. the inverse of L-BACK.

Paths that may be recognized in the future (active paths) ) )). This will be referred to as an active link record.

Link records show traceback information and can be used to reach that state. You can see the arc passing through the word model used. The traceback is You can also remove useless information from Lee. This means that the information is redundant in memory. This is necessary to prevent it from accumulating. Tracebacks are clearly recognized is used to output words that are common to all active paths. You can also The L-BACK entry in the link table is the previous link in the table. It points to the entry, which corresponds to the previously connected node within the tree. death Therefore, the traceback is the point where all paths meet through the tree diagram, i.e. In the process of tracing back to the tree node, It is said that there is. The concept of tracing all paths back to the point where they meet is well within the skill of the art. well known to those who

This is because the general explanation of "traceback" is the same as "partial traceback" and "partial traceback" mentioned above. This is because it is done with reference to ``and dynamic programming''.

The arrays described above are listed below for reference in the description of subsequent figures.

L-PTR: 255 baots. Each byte is set in a tree diagram (table) list as a pointer to the previous link record appended as a function of time to Available. Also used to chain free linked records in free lists. be done.

L-BACK: 25'5 bytes. Each byte represents the previous link record in the tree diagram. It can be used as a pointer.

L-WORD: 255 bytes. Each byte corresponds to the current link record , can be used as a symbol to indicate a potentially recognized word.

L-ACT: 255 bytes. Each byte has the current link record active (used during traceback).

L-FWRD: 255 bytes. Each byte is a subsequent valid link record in the tree diagram. (used during traceback).

In addition to the array described above, another five pointers are used.

They are as follows.

HEAD: First in the configuration list chained by the L-PTR array , a 1-byte pointer to the most recently added link record. .

FREE First list of free list chained by two L-PTR arrays A 1-byte pointer to the link record.

PTR: 1-byte pointer to the current tree node being processed.

TMPI and TMP2: 1-byte temporary ( temporary) pointer.

Structurally, assuming a table with only 10 entries in the configuration list , these arrays can be arranged as follows.

Record number L-PTRL-BACK L-1jORD L-ACT L-F WRD8 5 5” 9 0 0 The entries in the table above are shown in a tree diagram in Appendix A. HEAD is record # number) 7, and “0” in the L-PTR entry of record #3 indicates the last record in the list. It should be noted that the code is shown. FREE is not shown.

L-PTR allows you to simply delete records in the configuration list and records from the free list. Take it out and specify its L-PTR entry 1: HE AD record, and It can be inserted by updating AD and FREE. traceback If you delete a record from a table's configuration list during link the set list to the free list and use the L-PTR entry to link the set list to the free list. You can reorder the table by linking across the removed records. It will be available for free. Entries for L-ACT and L-FWRD are Used only during backup, otherwise always reset to O.

Recognition flow chart Referring now to FIGS. 5a through 5c, the recognition flowchart is shown in accordance with the present invention. is shown. The flowchart in Figure 5a shows the link table and its associated pointers. We begin at block 30 by resetting. L-FRW for reset procedure Setting each byte of D and L-ACT equal to 0, setting the HEAD pointer to set equal to 1 to indicate the beginning of the configuration list, and L-PTR(1) and L-PTR (255) are set equal to 0 to create the configuration list and free read respectively. This includes indicating the end of the strike.

Additionally, the template state memory, which is typically stored in more memory, is be made inactive. Therefore, the first record is records 2 to 255 use L-PTR entries. Chained to form a free list. Here HEAD is the beginning of the settings list. FREE points to the beginning of the free list (link record #2).

At block 32 of Figure 5a, the recognition grammar model initializes the node table. Initialized by 2 and 2. This corresponds to the cumulative distance before the initial, period node, and the low cumulative This can be done by arbitrarily assigning a distance measure to indicate the starting point of the grammar.

The link pointer before the initial node points to the first entry in the link table. is set to 1 accordingly.

All other nodes in the grammar model are initialized to inactive. grammar no The node is imported by setting the cumulative distance before the node table equal to infinity. Can be set to active, which makes it possible for other nodes to It can be shown that it does not exist.

At block 34, a traceback counter is initialized to ten. trace The back counter periodically indicates that a traceback process should be performed. used for. In this example, the traceback is processed when 10 input frames are processed. It is done every time it is done.

At block 36, the next input frame is sent to the system for template matching as described above. input into the system. All remaining steps in this flowchart process the current input frame. related to reason.

The traceback counter is decremented in block 38. t) to indicate that the frame has just been input.

At block 39, the node table is updated before processing the input frame.

Previous node parameters ('parameters' are the cumulative distance and where it started) (pointing to the link indicated) is copied to the current node parameter in the node table. It will be done. Additionally, all previously accumulated distances become inactive.

Test to see if all nodes of the grammar model have been processed in block 40 will be held. In other words, has the input frame been processed for the entire grammar model? It's about whether or not. If all nodes of the grammar model have been processed , flow proceeds to block 42 where the traceback counter registers the traceback. Check if the frame indicates what to do. If so, then 7a to 7d, which will be described later. A traceback is then performed. Following the traceback, block 36 Before processing the input frame, the traceback counter is reset at block 46. will be played.

If all the nodes in the grammar model have not been processed, the flow is as shown in Figure 5a. The process proceeds from block 40 to block 50 of 5b@. At block 50, the recognition grammar Model processing continues to the next node. Nodes have been processed for this frame. If not, "next node" is any node that is not the terminal node for the invalid arc. It can be a code.

In particular, the order of nodes for processing is such that the node where the invalid arc begins is relative to the invalid arc. must be processed before the node that is the final node . This means that the node cumulative distance and links are the same for the starting node of the template. This is to ensure that these templates are updated before being processed.

Have all templates ending in this node been processed in block 52? A test will be conducted to confirm whether this is the case.

Then each template immediately before each node in the grammar model before proceeding to other nodes Recognized to be processed. All of the templates ending in 20 nodes are processed. If this occurs, flow proceeds to block 68 of FIG. 5c, discussed below. All Tees If the sample plate has not yet been processed, flow continues to block 54 to process this node. Check whether the next arc to be processed by the code is an invalid arc. The next arc is invalid If so, flow proceeds to block 55.

At block 55, the cumulative distance and the links leading to that node are entered into the node table. The stored previous cumulative distance and the link to the starting node of the invalid arc are will be cut.

At this point, we can summarize the processing of the recognition grammar model shown in the recognition flow diagram so far. would be useful. Referring again to Figure 2, the initial node 22 of the grammar model is set to active and the corresponding link table entry may All grammar buses (branches of the tree) are initialized to indicate the criteria on which they occur.

Each time an input frame is processed, the grammar model returns one node at a time to the starting node. Proceed from to the end node. Furthermore, for each node in the grammar model, at that node Each template that ends processes one template at a time, as explained below. do. Therefore, for each input frame, each node is processed, and for each node Then, each template ending at that node is processed.

As indicated at block 56, regardless of whether a traceback is required, The next template for the current arc will be processed at block 58 or block 60. It will be done. If the template matches in any block, the current input frame, template, and the current cumulative distance and the arch from which the node in the node table starts. The cumulative distance and all states of the template based on the link pointer of the The link pointer for the link is updated. The current cumulative distance and links have been calculated time (previous frame) and the time they are used (current frame) Since the processing time for the previous frame is delayed by one frame, this information is not processed for the previous frame. It must be utilized by template processing as if it were. In other words , template processing is performed on the current cumulative frame before performing template processing on the current frame. The product distance and current link will be used to finish processing the previous frame. current If there is a possibility that it is a word end for this template in the current frame, Cumulative distances and link pointers corresponding to possible word ends are generated. Above Connected word recognition algorithm.

Flow proceeds from block 56 to block 60, where the traceback follows the template described above. In addition to rate processing, if you indicate that you are processing this frame, its template The link record pointed to by the link pointer for each active state in the All corresponding L-ACT entries are set to non-zero. "Active" A template state is one that has a finite cumulative distance.

Flow then proceeds to block 62 where the current tenbray distance is The best template (arc) processed earlier that ends in a node that Infinite if this is the first template processed for this node ) will be tested to see if it is better than the current cumulative distance corresponding to this exam The result is that the current input frame's template match is the word end of the word template. is true only if it indicates that it is possible that As mentioned earlier, word-final The sequence of input frames is stored in template memory if there is a possibility of an edge. Indicates that it corresponds to, or matches, the stored word template.

If the template has no end-of-word possibility for the current input frame, then The cumulative distance of association becomes infinite.

The most recently processed template (arc) stored for that node If it is found that there is no cumulative distance better than the current 'Am distance, the flow goes to block 52. , where another template ending at that node is processed.

The cumulative distance for the most recently processed arc so far processed for that node. If the information is found to be the best, flow proceeds to block 64 where this information is record information. At block 64, the cumulative distance and distance corresponding to the arc processed above are determined. The link pointer is the current cumulative distance to the node in the grammar model in the node table. Recorded as separated and current links. In addition, the word number representing the template, Or a symbol is recorded. The word number will indicate the word if it is later confirmed as recognized. Recorded for subsequent output. From block 64, the flow is as described above. Then proceed to block 52.

Block 54 indicates that all arcs ending at that node have been processed. If so, flow continues to block 68 of FIG. 5C. In Figure 5c, blocks 68 to 74 Check whether a linked record should be added to the tree until , the linked record is added to the tree through the linked array. Ru.

At block 68, a test is performed to see if the node in the grammar model is active. It will be done. The only way a node can be active is when that node At least one word template processed for is in the current input frame On the other hand, there is a possibility that it is the end of a word.

If the node is inactive, flow continues to block 40 of FIG. 5a. , find other nodes to process on the current frame. Otherwise, the flow is Proceed to block 70.

At block 70, the best arc ending at that node is equipped with a corresponding word template. A test will be conducted to confirm whether the In some cases, like the silent template. It may be a different type of template, such as a template, or it may be an invalid arc.

In this case, flow continues to block 40 of Figure 5a. The silent template is typical Since there is no need to output silence that is recognized as such, it is not added to the tree. Not possible. If the best arc ending at that node represents a word template The LINK subroutine (Figure 6) is called to read the link record. Add to diagram. the link pointer corresponding to that template's link record and the word number representing the template are sent to the LINK subroutine. Sent.

After a link record is added, a new link point is added, as described below. Inter is returned from LINK.

At block 74, the current link pointer for the node of the grammar model is set to LINK. Set to the link pointer sent from.

Following block 74, flow continues to block 40 of Figure 5a, where all grammar trees are check if all nodes have been processed for the current input frame .

Adding links to the tree Referring now to Figure 6, as explained earlier, this subroutine Add code to the tree diagram as specified by the link array. subroutine The parameter sent to is the word number corresponding to the tree node added from there. and a link pointer.

At block 78, a test is performed to determine if a free link record exists. It will be done. This is done by comparing FREE to 0. FREE equals 0 If so, the free link competition record no longer exists. As explained above, link ・Records in the array are chained together in the L-PTR array and are free It consists of a link record and a setting link record. Free link/ The record allows another linked record to be added to the tree diagram. Therefore If no free link records exist, all link records are used. has been reset, flow continues to block 80 where an error is reported and the system is reset. will be played. This step in block 80 causes the link table to overflow. It should be noted that they should only be used to protect against abnormal conditions that may cause Ru. Under normal conditions, the invention uses a link table of appropriate length to perform free linking. I'm trying not to run out of records.

If there is one or more free link records, flow continues to block 82 where the next The available link records are retrieved from the free list and set to HEAD and FREE. is inserted at the top of the configuration list, i.e. at the beginning, by updating the pointer of It will be done. FREE points to the index of the next free record. set and HEAD points to the link meeting record that was just added. Set. New rHE A DJ link record L-JPTR is new Chain the new record to the settings list. set to point.

At block 86, the HEAD selects the link record just added to the configuration list. code, so the word number sent to this subroutine is the L-WORD array. recorded as a new record. Also, the link sent to this subroutine The link pointer is recorded in the L-BACK array as a link record.

Block 88 checks whether a traceback is required for the current input frame A test will be conducted. If necessary, flow continues to block 90 to add new The linked record that was added is marked active. This is the L-AC of that record This is done by setting the T array equal to one. for the current input frame. If traceback is not required, the subroutine ends and the flow continues as shown in Figure 5C. Return to block 74.

tracing back through tree Referring now to Figures 7a through 7d, the traceback subroutine , block 44 of FIG. 5a is shown in detail. Traceback The subroutine searches through the tree diagram for words that have been identified as possible matches. ambiguit, ambiguous about the uniqueness of the match Check whether y) exists. Words that are uniquely identified are recognized by the recognition system as 2 It is output as a 2 word. Furthermore, the traceback subroutine is dead All linked records, i.e., are no longer considered as possible matches. records that are not available in the free list, freeing up memory for future linked records. make it possible to do so.

Before entering the traceback, L-ACT checks all active resources as described above. flag is set or flagged for the link record. trace At the beginning of the back, active link records are still being considered through the tree. Represents the end of all paths. The basic concept of traceback is to end of active path (first marked by L-ACT array) ) to ``tray-spack'' to see where all active paths meet. It's about finding out. The parts of the tree that are common to all active paths are The exact partial path can be represented and output as known. traceback During the process, the L-FWRD array moves the partial path in the forward direction (towards the end of the tree). Used to chain. Each partial bus as these partial buses are formed. The elementary nodes of are made active via the L-ACT array. Partial bus ( Previously marked as active from the current node (using L-BACK information) If you try to expand to a node, one or more possible paths originate from this previous node. The partial buses from both nodes are deleted (forward pointer Φ chain (L- FWRD) is reset to 0). All nodes marked as active are It is processed in this way. The order of node processing is This is the opposite order of the times they were added to the table. This order is set list is specific to the structure of The last node processed is the root of the tree (roo t)-node. the forward chain (partial bus) leaving that node at this point represents a distinct partial bus, and the corresponding recognized word is output. training The reset back procedure also causes the L-FWRD and L-ACT arrays to reset to zero. Immediately after completion, it “cleans up” itself. All link records that are not on the active contention path are explicitly Returned to the free list along with link records on the partial path.

Before explaining the traceback subroutine in detail, it may be helpful to go through an illustrated example. Probably one. Referring to Figure 8, in Figure A the tree is shown before the traceback. active link records, i.e. active word links Links that appear are marked with a thick dot on the right side of the link. 1st traceback The step is the most recent active link record added at just the right time. node, in this case the node marked 25.

A test is performed to see if there is an active node immediately before this node in the tree. be exposed. If it exists, it will be routed through node 25 or node 21 in this example. There is an ambiguity as to which path returns to node 21, such that Ru. When an ambiguity occurs, the forward pointer to the ambiguous node is Removes the forward pointer that was inserted, if it exists. This is for each subsequent link - Done by inserting a 0 into the L-FWRD array for the record.

In this example, there are no forward pointers for any nodes. That is, L-FWRD-0 It is.

The next most recently added active link record is identified (node 2 4). The preceding link of node 24 is also node 21, and the processing of node 25 described above is performed. Processed similarly.

The next most recently added active node is node 23. to this node Since there is no active predecessor node, the traceback amount process is is active and the preceding node is equal to the node currently being processed, node 23. The forward direction pointer (L-FWRD) of node 19 is recorded. Diagram B is a node Figure A is shown after processing 23 and the forward pointer added to node 19. is drawn with a thick line. After each node is processed, it becomes the active node. The L-ACT entry indicating that the Therefore, the tree Diagram B no longer depicts nodes 23, 24, and 25 as active.

Node 22 comprises the active node, the next most recently added entry. It is li. It has no active predecessor nodes. Therefore, at node 23 As we did for node 18, the forward pointer is set equal to node 22. and node 18 is marked active.

Ru. Diagram C shows the tree after processing node 22.

Node 21, as shown in the diagram, is active because its predecessor node is not active. It is processed in the same way as code 22.

Node 20 is the next active node to be processed. Precedent node 20 is active, indicating ambiguity. Ambiguity arises and forward pointers to the nodes of ambiguity, in this case nodes 18 and 20. has its forward pointer chain unraveled.

In this example, only node 18 has a forward pointer. Indicated by a thick line after node 18 As before, the forward pointer for node 18 is set equal to node 22. It is. Therefore, in Figure E, we set the forward pointer of node 18 equal to 0. By doing so, the thick lines are removed.

Node 19 is the next most recently added entry with an active node. It is li. Since its predecessor is active, the ambiguous The forward pointer chain for both nodes 16 and 19 is removed due to the There must be.

Node 18 is processed next. Its predecessor node, 16, is active, but Since neither node 16 or 18 has a forward pointer, the No action is taken other than removing the thick dots that indicate activity.

The next active node is node 16, whose predecessor nodes are active do not have. In this case, the traceback process makes the predecessor node active. , the predecessor node, node 13, is equal to the currently processed node, node 16 , record the forward pointer to . Figure G shows the tree after processing node 16. Figure A is shown.

Node 13 is treated similarly to node 16. Therefore, in Figure H, the tree is Only node 11 is shown marked as active; the order for nodes 11 and 13 is Only the direction pointer remains.

Once the traceback process has reached the root of the tree with no other nodes ahead of it, When the node (node 11) is reached, the word chained through the forward pointer Output as a recognized word. This has a chained forward pointer Continue searching for link records in L-FWRD starting from the root node. This is done by As shown in Figure 1, r eight J and r "Output a link record that records 1veJ. Additionally, in Figure H, a link record representing the link between nodes 11 and 16 is set. removed from the fixed list, as shown by the L-PTR array and the FREE pointer. will be linked to the free list. The new root node of the tree at this point is This is node 16. The remaining tree is shown in Figure I, which represents another input frame. is processed, i.e., flow returns to block 46 of Figure 5a of the recognition flow diagram. Used when

The traceback flowchart will now be explained in detail with reference to Figures 7a to 7d. do. In Figure 7a, the most recently added active from the link table is Find link records. At block 94, the first record of the link table is the only record in the configuration record list. It will be done. This is the index of the L-PTR array instructed by HE ADD. This is done by looking for Kusu. As mentioned above, HEAD is the most recently added Contains an index of linked records. L- corresponding to HEAD If the PTR entry is equal to O, the chain is terminated and there are no other entries in the table. There are no records. In this case, flow proceeds to block 96 where the corresponding L-AC T entry is set to inactive.

From block 96, the subroutine returns to block 46 of FIG. 5a of the recognition flow diagram. .

If another link record exists in the table, block 98 A test is performed to see if the link record for is active. The corresponding A link record is active if its L-ACT entry is not equal to 0. Ru.

If the link record is active, flow continues to block 100. Block check 100 indicates that the link record has already been taken into account. , set the linked record to inactive. Next, the flow is block 11 Go to 0 and store the record pointed to by HEAD in temporary pointer PTH. be done. From block 110, flow continues to block 120, as will be described. nothing.

Once the first link record returns to inactive, the flow continues at block 9 8, proceed to block 112. At block 112, the L-PTR Move around until you find a link record, and then move around until you find an active link record. Then, the green beans and sour beans for that active record are stored in the PTR. Ru. At block 114, the active indicator is cleared and at block 10 0, the processing for that link record is shown.

At block 116, the inner Link records found to be active are returned to the free list for future use. be done.

At block 118, a test is performed to see if there are more links in the link table. be exposed. This test is the same as the test performed in block 94 above.

If there are no more linked records in the configuration list, the subroutine is recognized. The flow diagram returns to block 46 of FIG. 5a.

Block 120 indicates that the link (node) preceding the current link is inactive. A test will be conducted to confirm whether the

This looks for the back pointer of the current link record and its corresponding L-AC This is done by looking for T entries. If the predecessor node is active, the flow proceeds to block 124 of FIG. 7b to handle the ambiguity problem discussed above. . If the predecessor node is inactive, flow continues to block 142 of Figure 7c. nothing.

Referring now to Figure 7b, the steps here are There may be more than one link record originating from the node under consideration. Handle ambiguity when indicating that This state is an active node Occurs when the node preceding it is also active. link record data Three temporary pointers (TMPI, TMP2 and and PTR). The step in block 121 is a forward pointer chain. Remove from previous link. This step includes blocks 124.126.128, and 130 are included. Figure 7b shows the node where the PTR is currently being processed. , or enter with a link record pointed to.

The L-BACK entry corresponding to the link Q record is currently being processed as described above. Points to the node immediately before the node being displayed. L related to link record -FWRD entries point only to possible descendant link records.

At block 124, a pointer to the node immediately preceding the current active node is Stored in MPI. At block 126, the predecessor node's The indicated descendant link record is stored in TMP2.

At block 128, the node indicated by TMP2 determines whether the actual L-FWRD entry is A test is performed to verify that the be exposed. If the node pointed to by TMP2 has a forward pointer (L-F if WRD is not equal to 0), flow continues to block 130 where the The forward pointer that points to is removed. At block 130, the contents of TMP2 are also TMP I, which causes the current node reference to be temporarily directed by TMP2. , and then, starting at block 126, the above steps are performed. For subsequent nodes, in the forward chain, as indicated by block 128, Iterate until a node is found that does not have a forward pointer indicating the end of the forward chemole. returned.

The steps in block 122 start with the current link, as indicated by PTR. Remove directional pointer chain. At block 132, the current link record - Pointer PTR is stored in TMPI. At block 134, the link record is The L-FWRD entry for the code is stored in TMP2. block 1 36, for this link record as done in block 128 above. A test is made to see if a forward pointer exists. to this link record If a forward pointer exists for the forward pointer, flow continues to block 138 where the forward The forward pointer is removed, and the descendant node removes its forward pointer as well. Therefore, it is stored in TMPI. Starting at block 134, the above steps Repeats until all forward pointers chained from the row node are removed. It will be done. From block 136, flow then proceeds to block 144 of FIG. Process active link records.

Referring back to block 120 of FIG. 7a, the node preceding the current node is If so, flow continues to block 142 of Figure 7c. this I will explain about it.

In block 142 of Figure 7c, the predecessor node (link record) is inactive. 20 nodes are set active, and in that order A direction pointer is set to point to the current link record. Block 14 4, starting from the current link record, the next active link record is found, the PTR is set to point to this record, and a new current link record is created. The table is searched until a code is found. At block 146, the block All inactive records encountered during step 144 are stored in the L-P It is returned to the free list by modifying the appropriate entry in the TR array.

At block 148, the new node is set to inactive, as shown in FIG. 7a. In the same way as was done in blocks 100 and 114, the node is already taken into account. Indicates that the

Block 152 determines whether this new link record is the last in the chain. Tests will be conducted to confirm this. If last, all link records are processed. 7d, flow continues to block 156 of Figure 7d to write the traceback. Outputs the words recognized during the checking process.

If this new link record is not the last in the chain, the flow continues processing Therefore, proceed to block 120 of FIG. 7a.

Referring now to FIG. 7d, block 156 sets the index of the current link record. The box, which is the root node of the tree, is stored in TMPl. Bu At lock 158, the node represented by the current node's forward pointer ( link record) is stored in TMP2.

For example, referring to Figure H in Figure 8, TMPl includes 11 (node 11). Therefore, TMP2 includes 13 (node 13).

At block 160, the forward point that is a descendant from the node stored in the TMPI is A test is performed to check if the interface exists. This compares the contents of TMP2 and 0. This is done by comparing. If so, flow continues to block 162 where the The forward pointer stored in TMP2 is removed and the current node is changed to the current node stored in TMP2. the next node in the forward chain, as indicated by the forward pointer of the be lifted up to the board. At block 164, associated with the current link record The word is output as a recognized word. Starting at block 158, the above A step is a process in which each link record in a forward pointer chain is It is repeated until the related word is included as a word. in block 160 During the step, a descendant link record is found that does not have a forward pointer. However, in this case, flow proceeds to block 168 where the flow is transferred between TMPL and PTRO. All dead link records are returned to the free list as shown . Additionally, block 168 determines whether the L-PTR array is currently pointed to by the PTR. Set the L-PTR entry for the new base node to O. It is updated by 0 in this record is the root of the tree. Indicates the end of the link record settings list. At block 168, the tracebar The check is complete and flow continues to block 46 of Figure 5a.

The present invention therefore provides a new and improved system and method for continuous speech recognition. and provide. As illustrated by the above-described flowcharts, the present invention is efficiently stylized. Easily implement real-time verification using a simple and inexpensive 8-bit processor be able to. The present invention further provides a method for minimizing the number of links when processing an input frame. Provides an excellent memory management method that only requires storing records.

Although the invention has been illustrated and described with reference to particularly preferred embodiments, those skilled in the art will appreciate that the invention Various modifications and changes may be made to the invention described above without departing from its spirit and scope. It will be understood that this can be done.

international search report

Claims (68)

[Claims] 1. Storing in memory a grammar model consisting of nodes connected by arcs and the arc has at least one associated template stored in memory. Each has a starting node and an ending node, and stores each group of input frames. Similarity measure parameters can be generated by comparing the audio to the template In a recognition system, one or more devices for recognizing each group of input frames. A first predetermined arc collection consisting of arcs and their respective nodes is a grammar model. A means of configuring the grammar model to have loops within the model and a given arc the first and second parameters for one or more nodes, including the terminal node for means for providing data storage; at least one first similarity measure parameter for the starting node of the predetermined arc; means for identifying a meter assembly; and at least one first meter for said predetermined arc. 2 of the similarity measure parameter set for one or more templates related to the given arc. template and the first similarity measure parameter set for the starting node; a means of checking using a third similarity measure parameter at the terminal node for the given arc; the at least one second class identified for the predetermined arc; A means for confirming using a similarity measure parameter set and a method for confirming the terminal node. said identified in said terminal node in said first parameter storage means; means for storing a third similarity measure parameter set; and a means for storing a third similarity measure parameter set; selecting the contents of the storage device as said second parameter storage means for said terminal node; means for selectively transferring; each group of input frames is stored in said parameter storage means associated with said terminal node; each of the input frames characterized by comprising: means for recognizing using the contents of the input frame; A device that recognizes groups. 2. Each similarity measure parameter set includes a cumulative distance measure and link information. 2. An apparatus for recognizing each group of input frames as claimed in claim 1. 3. the terminal node is the terminal node for one of the arcs comprising the loop; An apparatus for recognizing each group of input frames as claimed in claim 1. 4. The method according to claim 1, further comprising one or more invalid arcs within the grammar model. A device that recognizes each group of input frames. 5. Additionally, check the similarity measure parameter set at the starting node for invalid arcs. and the similarity measure parameter ascertained at the starting node for the invalid arc. the first parameter storage device for the terminal node of the invalid arc; each group of input frames as claimed in claim 4, further comprising means for selectively storing each group of input frames at a location. Recognizing device. 6. Furthermore, the node in the grammar model is defined as the starting node for the invalid arc. Provides a means to process the node so that it is processed before the terminal node for the invalid arc. 5. An apparatus for recognizing each group of input frames according to claim 4. 7. Storing in memory a grammar model consisting of nodes connected by arcs and the arc has at least one associated template stored in memory. Each has a starting and ending node, and stores each group of input frames. The similarity measure parameter can be generated by comparing the template In a voice recognition system, a method for recognizing each group of input frames, the method comprising: A first predetermined arc collection consisting of the above arcs and their respective nodes is a grammar The steps of configuring a grammar model to have a loop in the model and the first and second parameters for one or more nodes, including the terminal node for the providing data storage; at least one first similarity measure parameter for the starting node of the given arc; determining a meter aggregate; at least one second similarity measure parameter set for the predetermined arc; for one or more templates associated with the given arc and the starting node. determining the first similarity measure parameter set using the first similarity measure parameter set; a third similarity measure parameter at the terminal node for the predetermined arc; the at least one second class determined for the predetermined arc; determining the final node using the similarity measure parameter set; The third similarity measure parameter determined by applying the third similarity measure parameter to the terminal node storing in a first parameter storage means; and selecting the contents of the storage device as said second parameter storage means for said terminal node; selectively transferring the data; the contents of said parameter storage device relating each group of input frames to said terminal node; each group of input frames, characterized in that the step of recognizing the input frames using How to recognize. 8. Each similarity measure parameter set includes a cumulative distance measure and link information. 8. A method for recognizing each group of input frames as claimed in claim 7. 9. the terminal node is the terminal node for one of the arcs with the loop; 8. A method for recognizing each group of input frames as claimed in claim 7. 10. Furthermore, two or more nodes are associated with at least one invalid archive in the grammar model. 8. The method for recognizing each group of input frames according to claim 7, comprising the step of: Law. 11. Furthermore, the similarity measure parameter set at the starting node for invalid arcs the class determined at the starting node for the invalid arc; the similarity measure parameter set to the first parameter for the terminal node of the invalid arc; 11. The input file of claim 10, further comprising the step of selectively storing in a parameter storage device. How to recognize each group of frames. 12. Further, the node in the grammar model is set as the starting node for invalid arcs. A step of processing such that is processed before the terminal node for the invalid arc. 11. A method for recognizing each group of input frames as claimed in claim 10. 13. The input frame is matched to a pre-stored template representing audio. Templates under consideration as templates that may be processed and recognized as are recorded separately as link records in the link network, and The link record generally comprises ancestor and descendant link records. In speech recognition systems, means for providing a temporary pointer to the link record; label a given one of said link records with said temporary pointer; possible recognized templates corresponding to said one link record with means for tracing back through said network while connecting to said network; Link record with two or more possible descendant link records a means for verifying the linked record; and a means for verifying the linked record; means for deleting the re-pointer; the linked record still labeled with the temporary pointer means for outputting data corresponding to; A device for recognizing voice patterns, comprising: 14. Periodically alert the system after a predetermined number of input frames have been processed 14. A device for recognizing speech patterns as claimed in claim 13, comprising means. 15. Means for configuring a network according to a predetermined grammar model topology 14. The apparatus for recognizing speech patterns according to claim 13. 16. For at least one of said templates, a possible word end is Once recognized, the contract must have a means of attaching a link record to the network. A device for recognizing the voice pattern according to claim 13. 17. Provides a means to indicate the most recently added link record in the network. 14. The device for recognizing speech patterns according to claim 13, wherein 18. Provides a means for indicating that a link record has no ancestor link record. 14. An apparatus for recognizing speech patterns according to claim 13. 19. The link record is defined as a free link record and a set link record. 14. The speech pattern according to claim 13, further comprising means for storing the speech pattern in a table consisting of A device that recognizes 20. Indicates the beginning of a set link record and the beginning of a free link record 20. Apparatus for recognizing speech patterns according to claim 19, comprising means. 21. A means of indicating the end of a set link record and the end of a free link record. 20. An apparatus for recognizing speech patterns according to claim 19. 22. For link records that are descendants from the confirmed link record. 14. The voice of claim 13, further comprising means for deleting a temporary pointer that A device that recognizes patterns. 23. The input frame is matched to a pre-stored template representing audio. Templates under consideration as templates that may be processed and recognized as are recorded separately as link records in the link network, and The link record generally comprises ancestor and descendant link records. In speech recognition systems, The link record is divided into a symbol, each representing a template, and a link record. a sequence indicator that represents the relative time the code was stored, and a network a first pointer to the link record from which each of the links is derived; , a second temporary pointer; and means for storing the second temporary pointer; trace back through the network using data aggregations with means for identifying at least one known link record; means for outputting the clearly recognized link record; 1. A device for recognizing voice patterns, comprising: 24. Warn the system periodically after a predetermined number of input frames have been processed 24. Apparatus for recognizing speech patterns according to claim 23, comprising means. 25. A means of constructing a network according to a predetermined grammar model topology 24. The apparatus for recognizing speech patterns according to claim 23. 26. possible end-of-word for at least one of said templates. Claims that have a means of attaching link records to the network after recognition 24. A device for recognizing a voice pattern according to item 23. 27. Provides a means to indicate the most recently added link record in the network. 24. The device for recognizing speech patterns according to claim 23, wherein 28. Provides a means for indicating that a link record has no ancestor link record. 24. An apparatus for recognizing speech patterns as claimed in claim 23, comprising: 29. The said link record is a free link record and a set link record. 24. The audio pattern according to claim 23, further comprising means for storing the audio pattern in a table comprising: A device that recognizes 30. Indicates the beginning of a set link record and the beginning of a free link record 30. Apparatus for recognizing speech patterns according to claim 29, comprising means. 31. A means of indicating the end of a set link record and the end of a free link record. 30. An apparatus for recognizing speech patterns according to claim 29. 32. For link records that are descendants from the confirmed link record. 24. The voice of claim 23, further comprising means for deleting a temporary pointer that A device that recognizes patterns. 33. The input frame is matched to a pre-stored template representing the audio. Templates under consideration as templates that may be processed and recognized for The route is recorded separately in the link network as a link record. , said link record generally comprises ancestor and descendant link records. In a speech recognition system, The link record is divided into a symbol, each representing a template, and a link record. a sequence indicator that represents the relative time the code was stored, and a network and a pointer to the link record derived from it. a means of storing the indexed data in a table as an indexed data collection; The table consists of a free record space and a set record space, The link record is stored in the configuration record space and Traceback through the network using indexed data collections , one or more link records whose corresponding templates are explicitly known a means for identifying the code; outputting data representing said specifically recognized link record; A record is retrieved from said configuration record space, thereby retrieving said retrieved record. The linked record continues to have free record space to store the linked record. a means of making it so that A device for recognizing voice patterns, comprising: 34. A way to periodically alert the system after processing a predetermined number of input frames. 34. An apparatus for recognizing speech patterns as claimed in claim 33, comprising steps. 35. A means of constructing a network according to a predetermined grammar model topology An apparatus for recognizing a speech pattern according to claim 33, comprising: 36. A possible word terminator for at least one of said templates is Once recognized, the contract must have a means of attaching a link record to the network. 34. A device for recognizing a voice pattern according to claim 33. 37. Provides a means to indicate the most recently added link record in the network. 34. The device for recognizing speech patterns according to claim 33, wherein 38. Provides a means for indicating that a link record has no ancestor link record. 34. An apparatus for recognizing speech patterns as claimed in claim 33, comprising: 39. Possible to have more than one descendant link record that may be recognized. 34. The sound of claim 33, comprising means for determining possible link records. A device that recognizes voice patterns. 40. In said configuration record space, the corresponding template is clearly recognized. select a specific link record that is not 34. A voice pattern according to claim 33, further comprising means for returning to an original record space. A device that recognizes 41. The input frame is matched to a pre-stored template representing audio. Templates under consideration as templates that may be processed and recognized as are recorded separately as link records in the link network, and The link record generally comprises ancestor and descendant link records. In speech recognition systems, providing a temporary pointer to the link record; labeling a predetermined one of the link records with the temporary pointer; While tracing back through the network, the one link record connecting a potentially recognized template corresponding to the code; of the possibility of having more than one descendant link record that could be recognized. determining a link record; delete the temporary pointer corresponding to the determined link record; and the link still labeled with the temporary pointer. ・A step of outputting data corresponding to the record; A method for recognizing speech patterns, comprising: 42. Periodically alert the system after a predetermined number of input frames have been processed 42. The method of recognizing speech patterns of claim 41, comprising the steps of: 43. A step to construct a network according to a predetermined grammar model topology. 42. The method of recognizing speech patterns of claim 41, comprising: 44. A possible word terminator for at least one of said templates is After being recognized, it includes the step of attaching a link record to the network. 42. The method of recognizing speech patterns according to claim 41. 45. Step to show the most recently attached link record in the network 42. The method of recognizing speech patterns of claim 41, comprising: 46. A step indicating that the link record has no descendant link records. 42. The method of claim 41, wherein the method comprises: 47. The said link record is a free link record and a set link record. 42. The audio parameters according to claim 41, further comprising the step of: storing the audio parameters in a table comprising: How to recognize turns. 48. Indicates the beginning of a set link record and the beginning of a free link record 48. A method of recognizing speech patterns as claimed in claim 47, comprising the steps of: 49. A step indicating the end of a set link record and the end of a free link record. 48. The method of recognizing speech patterns of claim 47, comprising: 50. for link records that are descendants from the determined link record. 42. The method of claim 41, further comprising the step of deleting a temporary pointer that How to recognize speech patterns. 51. The input frame is matched to a pre-stored template representing audio. Templates under consideration as templates that may be processed and recognized as are recorded separately as link records in the link network, and The link record generally comprises ancestor and descendant link records. In speech recognition systems, Each of the link records includes a symbol representing a template and a link record. a sequence indicator that represents the relative time the code was stored, and a network A first pointer to the link record from which each in the link is derived. and a second temporary pointer. and the step of storing it as the indexed data collection, including the second temporary pointer; traceback through the network using identifying one link record; and the step of identifying one link record; ・A step to output a record, A method for recognizing speech patterns, comprising: 52. Periodically alert the system after a predetermined number of input frames have been processed 52. The method of recognizing speech patterns of claim 51, comprising the steps of: 53. A step to construct a network according to a predetermined grammar model topology. 52. The method of recognizing speech patterns of claim 51, comprising: 54. possible word endings for at least one of said templates; includes the step of attaching a link record to the network after the 52. The method of recognizing speech patterns according to claim 51. 55. Step to show the most recently attached link record in the network 52. The method of recognizing speech patterns of claim 51, comprising: 56. A step indicating that the link record has no ancestor link record. 52. The method of claim 51, wherein the method comprises: 57. The link record is defined as a free link record and a set link record. 52. The audio pattern according to claim 51, further comprising the step of storing the audio pattern in a table consisting of: How to recognize the zone. 58. Indicates the beginning of a set link record and the beginning of a free link record 58. The method of recognizing speech patterns of claim 57, comprising the steps of: 59. A step indicating the end of a set link record and the end of a free link record. 58. The method of recognizing speech patterns of claim 57, comprising: 60. for link records that are descendants of the determined link record. 52. The audio of claim 51, comprising the step of deleting the temporary pointer. How to recognize patterns. 61. The input frame is matched to a pre-stored template representing audio. Templates under consideration as templates that may be processed and recognized as are recorded separately as link records in the link network, and The link record generally comprises ancestor and descendant link records. In speech recognition systems, Each of the link records includes a symbol representing a template and a link record. a sequence indicator that represents the relative time the code was stored, and a network Each link in the record contains a pointer to the link record from which it was derived. free record space and configured record space as indexed data collections. The data stored in the configuration record space of the table consisting of the code space and step, tracebar through the network using the indexed data collection. one or more link records for which the corresponding template is clearly known. identifying the code; outputting data representing said specifically recognized link record; A record is retrieved from said configuration record space, thereby retrieving said retrieved record. Free record space for linked records that the linked record subsequently stores a voice pattern characterized by comprising steps for making it paced; How to recognize. 62. Periodically alert the system after a predetermined number of input frames have been processed 62. The method of recognizing speech patterns of claim 61, comprising the steps of: 63. A step to construct a network according to a predetermined grammar model topology. 62. The method of recognizing speech patterns of claim 61, comprising: 64. A possible word terminator for at least one of said templates is After being recognized, it includes the step of attaching a link record to the network. 62. A method for recognizing speech patterns according to claim 61. 65. Step to show the most recently attached link record in the network 62. The method of recognizing speech patterns of claim 61, comprising: 66. A step indicating that the link record has no ancestor link record. 62. The method of claim 61, wherein the method comprises: 67. Possible to have more than one descendant link record that may be recognized. 62. The method of claim 61, further comprising the step of determining possible link records. How to recognize speech patterns. 68. If the corresponding template is not output in the configuration record space above. select a specific link record, and add the specific link record to the free record. 62. The voice pattern of claim 61, comprising the step of returning the audio pattern to a code space. How to recognize.