WO2017012243A1 - Voice recognition method and apparatus, terminal device and storage medium - Google Patents

Abstract

A voice recognition method and apparatus, a terminal device and a storage medium. The method comprises: obtaining pronunciation information according to voice information recognition (21); loading a language model score query tree according to score query tree information, and querying the language model score query tree to determine a probability score of a character recognition result matching the pronunciation information, wherein the score query tree information comprises a plurality of nodes corresponding to characters, and each node at least comprises a storage position offset between a current node and a sub-node (22); and selecting a character recognition result as a final recognition result according to the probability score (13). By means of the method and apparatus, a language model score query tree is directly loaded according to a storage location offset between a current node and a sub node during starting, such that a starting time is greatly shortened.

Description

Speech recognition method, device, terminal device and storage medium

This patent application claims priority from the Chinese patent application filed on July 20, 2015, the application number is 201510427908.5, the applicant is Baidu Online Network Technology (Beijing) Co., Ltd., and the invention name is "voice recognition method and device". The full text of the application is hereby incorporated by reference.

Technical field

The embodiments of the present invention relate to the field of voice recognition technologies, and in particular, to a voice recognition method, device, terminal device, and storage medium.

Background technique

In the field of embedded speech recognition, speech recognition results are determined by two parts: acoustic model and language model. The language model plays an important role. For example, when the pronunciations of “Bei Daihe” and “Be brought by the river” are similar, the scores of the acoustic models are almost the same. At this time, the language model is needed to further decide which words are used in the language. of. That is to say, the language model solves the problem of evaluating the natural language order in speech recognition.

As shown in FIG. 1 , the voice recognition method provided in the prior art mainly includes the following steps:

S11. The language model resource is read from the hard disk, and the resource is stored in a node manner;

Each node corresponds to one word, and each node is composed of node information (including the corresponding word or word, child information, such as the word corresponding to the child node and the number of children), probability list (ProbList) (storage probability), and fallback. The probability list (BackOff) consists of three parts; as shown in the following table 1:

Table I

S12. Construct a multi-cross check tree according to the read language model resource;

The process of constructing a tree is specifically: after the language model resource is loaded into the cache, the storage address of the node changes, so each node only knows which word its own child node is, and does not know its storage address, so it needs to be based on The child node information recorded in each node queries the storage address of its child node one by one and adds it to the parent node to establish a score tree.

S13. Loading an acoustic model and other speech recognition resources;

S14. Receive input voice information, and decode using a Viterbi algorithm;

S15. In the process of decoding, the acoustic model is used for speech recognition, the pronunciation information is obtained, and the multi-cross check tree of the language model is searched according to the pronunciation information for scoring;

S16. Obtain a recognition result of the language model;

S17. Output the recognition result and release the resource.

However, the existing speech recognition method needs to dynamically load the language model resources after reading the language model resources, and construct a multi-cross check tree, which is a waste of time and leads to low recognition efficiency.

Summary of the invention

The embodiment of the invention provides a voice recognition method, device, terminal device and storage medium, which can greatly shorten the startup time.

In a first aspect, an embodiment of the present invention provides a voice recognition method, including:

Acquiring the pronunciation information according to the voice information;

And loading the language model to check the tree according to the query tree information, and querying the language model to determine a probability score of the text recognition result that matches the pronunciation information; wherein the query tree information includes a plurality of nodes corresponding to the text, Each node includes at least a storage location offset between the current node and the child node;

The character recognition result is selected based on the probability score as the final recognition result.

In a second aspect, an embodiment of the present invention further provides a voice recognition apparatus, including:

a pronunciation information obtaining module, configured to identify the pronunciation information according to the voice information;

a probability score query module, configured to load a language model check score tree according to the check score tree information, and query the language model check score tree to determine a probability score of a text recognition result that matches the pronunciation information; wherein the check score tree information includes a plurality of nodes corresponding to the text, each node including at least a storage location offset between the current node and the child node;

A text recognition module is configured to select a text recognition result according to the probability score as a final recognition result.

In a third aspect, the embodiment of the present invention further provides a terminal device for implementing voice recognition, including:

One or more processors;

Memory

One or more modules, the one or more modules being stored in the memory, and when executed by the one or more processors, performing the following operations:

Acquiring the pronunciation information according to the voice information;

And loading the language model to check the tree according to the query tree information, and querying the language model to determine a probability score of the text recognition result that matches the pronunciation information; wherein the query tree information includes a plurality of nodes corresponding to the text, Each node includes at least a storage location between the current node and the child node Offset;

The character recognition result is selected based on the probability score as the final recognition result.

In a fourth aspect, an embodiment of the present invention further provides a non-volatile computer storage medium, where the computer storage medium stores one or more modules, when the one or more modules are executed by a device that performs a voice recognition method. When the device is caused to perform the following operations:

Acquiring the pronunciation information according to the voice information;

And loading the language model to check the tree according to the query tree information, and querying the language model to determine a probability score of the text recognition result that matches the pronunciation information; wherein the query tree information includes a plurality of nodes corresponding to the text, Each node includes at least a storage location offset between the current node and the child node;

The character recognition result is selected based on the probability score as the final recognition result.

The technical solution of the embodiment of the present invention directly stores the language model check tree according to the storage location offset between the current node and the child node, and does not need to dynamically construct the language model to check the tree at startup, thereby greatly shortening the startup time.

DRAWINGS

1 is a schematic flow chart of a voice recognition method provided by the prior art;

2A is a schematic flowchart of a voice recognition method according to Embodiment 1 of the present invention;

2B is a schematic structural diagram of a first check molecular tree in a speech recognition method according to Embodiment 1 of the present invention;

2C is a schematic structural diagram of a second molecular check tree in a speech recognition method according to Embodiment 1 of the present invention;

2D is a schematic structural diagram of a third check molecular tree in a speech recognition method according to Embodiment 1 of the present invention;

2E is a schematic diagram of a fourth molecular tree structure in a speech recognition method according to Embodiment 1 of the present invention;

3 is a schematic structural diagram of a voice recognition apparatus according to Embodiment 2 of the present invention;

FIG. 4 is a schematic structural diagram of a terminal device for implementing voice recognition according to Embodiment 3 of the present invention.

detailed description

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. It should also be noted that, for ease of description, only some, but not all, of the structures related to the present invention are shown in the drawings.

The execution body of the voice recognition method provided by the embodiment of the present invention may be a voice recognition device provided by an embodiment of the present invention, or a terminal device (for example, a smart phone, a tablet computer, etc.) integrated with the voice recognition device, and the voice recognition The device can be implemented in hardware or software.

Embodiment 1

2A is a schematic flowchart of a voice recognition method according to Embodiment 1 of the present invention. As shown in FIG. 2A, the method specifically includes:

S21. Obtaining pronunciation information according to the voice information;

Specifically, the user can input voice information in the voice recognition device provided by the embodiment of the present invention. For example, a voice recording button can be set in the input field of the voice recognition device, and the user clicks on the above. The voice recording button can start the recording function and record the user's speech to obtain the voice information. Then, the voice information is identified and processed by the pre-loaded acoustic model and the voice recognition resource, so that the required pronunciation information can be obtained. For example, if the voice that the user wants to input is "Beidaihe", the pronunciation information obtainable by the above recognition process is "beidaihe".

S22. The language model is searched according to the query tree information, and the language model check tree is queried to determine a probability score of the text recognition result that matches the pronunciation information. The check tree information includes multiple texts corresponding to the text. a node, each node including at least a storage location offset between the current node and the child node;

The check tree information is similar to the language model resource, and includes a plurality of nodes corresponding to the text, wherein each node includes at least a storage location offset between the current node and the child node. In addition, the check tree may further include a storage probability (ProbList, that is, a probability of occurrence of the current node) of each node, a probability of returning the current node (BackOff), and a number of corresponding child nodes. The child node is the child node. The parent node and the child node are the combination of words that will appear at the same time. For example, the "Beijing" and "Beijing" nodes are the parent nodes of the "North" node. The storage location offset is specifically the distance between the storage locations of the nodes and the child nodes.

Specifically, the query tree information of the language model is directly written into the language model resource, so that the initialization of the tree tree is not required to dynamically construct the language model, but the pointer information of the dynamically constructed tree is regarded as an offset. The quantity is written into the language model resource, that is, the language model check tree is constructed offline in advance. The storage location offset between the current node and the child node is directly written into the language model resource, and the query tree information is as shown in Table 2 below. When the load is started, the offline language model search tree is directly loaded according to the check tree information.

Table II

When the check tree information is loaded into the cache, the storage relative distance between the nodes does not change. Therefore, the storage locations of other nodes can be determined based on the storage locations of the initial nodes and the offsets with other nodes.

According to the information in Table 2 above, the voice model check tree can be loaded onto the line, and the check tree is queried according to the pronunciation information. For example, the pronunciation information obtained by the above step S21 is “beidaihe”, and first, in the root node (RootProbList) of the check tree, the text node corresponding to “he” in the pronunciation information “beidaihe” is searched for, and includes a plurality of, for example, “charges”. , "drink", "river", etc., as shown in FIG. 2B, and then query the text node corresponding to "dai" in the child nodes of the text nodes corresponding to each "he", and also include a plurality of, for example, "bands" and "wearing "," "wait", etc., query the ProbList of the text node corresponding to each "dai" and the probability score corresponding to "daihe" in BackOff, for example, the results shown in Table 3 below are obtained:

Table 3

ProbList BackOff Dai He 1% 98% Belt load 1% 98% Waiting for the load 1% 98% Daihe 99% 1%

With river 80% 5% Waiting for the river 60% 30% Wear and drink 1% 99% Drink with 1% 98% Waiting to drink 70% 7%

Then, through the above Table 3, the backoff probability BackOff of the text nodes under the "dot", "band" and "to" of the "charge" corresponding to the "charge" is higher than 60%, which is higher, and the ProbList is lower than 60%. If it is low, the subtree corresponding to "荷" is returned. Similarly, the ProbList of the text nodes under the "river" corresponding to "river" are higher than 60%, and the BackOff is lower than 60%, then the subtree corresponding to "river" remains. . The ProbList of the text node under the "children" corresponding to "drink" is higher than 60%, and the BackOff is lower than 60%, then the corresponding child node "drink" is reserved, and the corresponding child node "drinks" The ProbList of Dai and "Band" are both lower than 60%, and BackOff is higher than 60%. The corresponding child nodes "Dai" and "Band" of "Drink" are returned. The final result obtained by the above selection process is shown in the two subtrees shown in Figures 2C and 2D.

On the basis of 2C and the subtree shown in FIG. 2D, according to "bei" in "beidaihe" in the pronunciation information, the child nodes of each character node corresponding to "dai" are again queried, and also include a plurality of, for example, " Being "," "north" and "back".

The problist corresponding to the text node corresponding to each "bei" and the probability score corresponding to "beidaihe" in BackOff are obtained, for example, the results shown in Table 4 below are obtained:

Table 4

ProbList BackOff Back wearing river 30% 70%

Daihe 60% 15% Beidaihe 99% 1% Strap river 5% 90% Brought to the river 1% 95% North Belt River 30% 91% Beibeihe 1% 90% Being treated by the river 1% 98% Back to the river 2% 90% Waiting to drink 1% 99% North to drink 1% 91% Being to be drunk 1% 97%

Then, through the above Table 4, the ProbList of the text nodes under the "Bei Daihe" and "Be Daihe" corresponding to the "Dai" are higher than 60%, and the BackOff is lower than 60%, then the "Dai" corresponding The child nodes "Bei Daihe" and "Waihehe" are reserved, and the BackOff of the text node of the "children wearing" corresponding to "Dai" is higher than 60%, and the ProbList is lower than 60%, corresponding to "Dai" The child node "back wearing the river" was returned. In the same way, it can be concluded that the subtrees corresponding to "to" and "band" are returned. The results finally obtained by the above selection process are shown in Fig. 2E.

S23. Select a character recognition result according to the probability score as a final recognition result.

Similarly, taking the above step S23 as an example, the final recognition result is “Bei Daihe” and “Beaded River”, and the corresponding probability scores are 99% and 60%, respectively, and the score recognition result can be placed according to the score. In front of the display, the low-scoring text recognition results are displayed on the principle, and they are returned to The user, that is, "Bei Daihe" and "Daihe" will be returned to the user at the same time for the user to choose. It is also possible to return only the highest score to the user, that is, return "Bei Daihe" to the user.

In this embodiment, it is not necessary to dynamically construct a language model check tree at startup, and the language model check tree is pre-recorded by the storage location offset between the current node and the child node, and the load location may be directly stored according to the storage location between the nodes. The offset loads the lookup tree into the cache, eliminating the need for dynamic builds, which greatly reduces startup time.

Exemplarily, for a faster query, the tree may be searched according to the language model loaded by the pronunciation information query, and the following steps are added before determining the probability score of the text recognition result matching the pronunciation information:

Querying a probability score of a text recognition result matching the pronunciation information in a common word sequence stored in the cache and/or a text recognition result of the recorded history query according to the pronunciation information;

If there is no probability score of the character recognition result matching the pronunciation information in the cache, an operation of performing a query in the language model check tree is triggered.

The common word sequence includes some words and hot words that people often visit in people's lives, for example, including the name of a tourist attraction, the place name of a municipality directly under the central government, the name of a network celebrity, the name of a song, etc. The common vocabulary is placed in the cache, which can greatly improve the query efficiency.

The text recognition result of the historical query can also be recorded in the cache. When the user inputs the same voice information again, it can be directly returned to the user from the cache, which also saves the query time.

Since the language model resources currently used are relatively large in size, even a tailored language model occupies a large amount of memory, and the clipping also affects the query efficiency. To further save memory, the present embodiment of the present invention The language model looks at the tree for conversion and converts it into a more memory-saving language. Word model. Specifically, before the probability score of the character recognition result matching the pronunciation information is queried in the common word sequence in the cache according to the pronunciation information, the following operations are further added to form a common word sequence:

Forming a first array by using a probability that a probability score of a single character in the language model is higher than a set threshold and a probability score thereof;

Forming, by the language model, a combination of at least two characters of a text group having a probability score higher than a set of text values and a probability score thereof to form a second array;

The first array and the second array are stored as the common word sequence.

Specifically, a part of the single text or all the single characters included in the root node in the language model checking tree and the corresponding probability scores are stored in the form of an array. The parent node of the language model and the text combination corresponding to each child node and its probability score are also stored in the form of an array. In this embodiment, by setting a set threshold value, a single character and a combination of characters with low probability included in the language model check tree can be removed to improve query efficiency.

For example, the text included in the root node of the language model scoring tree includes “North”, “Beijing”, “River”, “Hang”, “Drink”, etc., and the corresponding storage probabilities are P1, P2, P3, respectively. P4, P5, can be realized by two-dimensional array, and its storage form is as shown in Table 5 below:

Table 5

Text north Beijing River Charge drink ...... Probability of occurrence P1 P2 P3 P4 P5 ......

Similarly, for the language model, the parent node and the character combinations corresponding to each child node and their probability scores may be stored in a two-dimensional array, for example, as shown in Table 6 below, which is a combination of binary characters:

Table 6

Text combination Beijing Nanjing Daihe Flower lotus Waiting to drink ...... Probability of occurrence P6 P7 P8 P9 P10 ......

In subsequent queries, you can query directly from the above array.

Exemplarily, in order to further improve query efficiency, the target area to be queried may be quickly located by using a positioning table, and the following steps may be further added after storing the first array and the second array as the common word sequence:

And dividing a plurality of text combinations in the second array into an ordered sequence array and an unordered sequence array according to a predetermined rule, wherein the ordered sequence array includes at least two sub-arrays, and each sub-array stores the same eigenvalue Multiple text combinations;

Generating a probability score in the array of unordered sequences, and a starting position and/or a ending position, and a feature value, a starting position, and/or a ending position of each sub-array in a positioning table;

The predetermined rule may be set according to a specific scenario, and different division rules are used for different scenarios, so that a better text recognition result can be better and faster matched. Specifically, the identifier value corresponding to the first character in the binary character combination may be right shifted by the first specified number of digits and the identifier value corresponding to the second character is shifted to the left by the second specified digit as the feature value K, and the feature value is obtained. A binary text combination in which the number of binary text combinations of K is greater than or equal to a preset value is classified into an ordered sequence array; a binary combination of characters whose binary value combination of the feature value K is smaller than a preset value Classified as an array of unordered sequences.

For example, for a binary combination of words, Equation 1 can be used to calculate an ordered sequence array and an unordered sequence array. First, the eigenvalue K of the binary combination is calculated:

K=M1>>3+M2<<13

Wherein, the first designated digit has a value of 3, the second designated digit has a value of 13, and “>>” is a right shift symbol. "<<" is the left shift symbol, M1 is the identifier value corresponding to the first text, and M2 is the identifier value corresponding to the second text.

The binary text combination whose number of binary character combinations whose feature value is K is greater than or equal to the preset value is classified into an ordered sequence array; the number of binary character combinations whose feature value is K is less than the preset value The metacharacter combination is classified as an unordered sequence array.

The identification value is a value that can uniquely identify the text. For example, when the characters are identified by ASCII code, the ASCII code value of the text is the identification value. After the identification value of the character is shifted to the left and right, the feature value K is calculated, which is equivalent to classifying each character combination according to the feature value K, and grouping the characters having the same feature value K into one group. If the number of combinations of texts in the group is too small, it is not necessary to set the group.

It will be understood by those skilled in the art that the preset rule may also be other formulas, not limited to left shift and right shift, and is not limited to the specific number of bits of the above shift.

For the ternary character combination, the eigenvalue K of the binary combination can be calculated by the above formula, and then the eigenvalue K is shifted to the right by the first specified number of digits and the identification value corresponding to the third character is shifted to the left by the second specified number of digits. And as the feature value T; the ternary character combination whose number of ternary characters whose feature value is T is greater than or equal to the preset value is classified into an ordered sequence array; the number of ternary characters combined with the feature value T A ternary text combination that is less than a preset value is classified as an unordered sequence array.

For example, an ordered sequence array and an unordered sequence array can be obtained by combining the above formula 1 and the following formula 2. First, the eigenvalue K of the binary combination is obtained by using the formula 1, and then the eigenvalue T of the ternary combination is obtained by using the formula 2

T=K>>3+M3<<13

The first designated digit has a value of 3, the second specified digit has a value of 13, K is a feature value corresponding to the combination of M1 and M2, and M3 is an identifier value corresponding to the third character;

The ternary character combination whose ternary character combination whose feature value is T is greater than or equal to the preset value is classified into an ordered sequence array; the number of ternary character combinations whose feature value is T is less than the preset value of three The metacharacter combination is classified as an unordered sequence array.

The ordered sequence array can be divided into multiple sub-subarrays according to the feature value, and each sub-array stores a combination of characters having the same feature value. For the eigenvalues of the binary combination, it is calculated by the above formula 1. For the ternary combination, it can be calculated by combining formula 1 and formula 2.

The following is a detailed description of the construction process of the positioning table.

First, according to the voice check tree, all the text combinations are counted, including binary text combination, ternary text combination and n-gram combination, where n is a natural number greater than 3. The more common ones are binary text combinations and ternary text combinations. For the binary combination, the eigenvalue is calculated by using the above formula 1. For the ternary combination, the eigenvalue is calculated by using the above formula 1 and formula 2. For example, a combination of characters with a eigenvalue of K1 is calculated including "Beijing", "Tianjin", "Bei Daihe", "Baidu", and "Sohu", and the combination of characters with a eigenvalue of K2 is calculated to include "Hangzhou good" and "taken River, "milk" and "yoghurt", the text combination with the eigenvalue K3 is calculated, including "Suzhou", and the text combination with the eigenvalue K4 is calculated, including "summer heat" and "wearing the river", and the eigenvalue is calculated. For the text combination of K3, including "Suzhou", the number of text combinations whose statistical feature value is K1 is 5, the number of text combinations whose statistical feature value is K2 is 4, and the number of text combinations whose statistical feature value is K3 is 1. The number of text combinations whose statistical feature value is K4 is 2. If the preset value is set to 3, the number of text combinations with the same feature value exceeds 3, which is classified as an ordered sequence array. Otherwise, it is classified as none. The ordered sequence array, the resulting ordered sequence array is represented in the form of a list, as shown in Table 7 below, where the ordered sequence array also contains the probability of occurrence of the feature value and each combination of words, which can be directly from the language. Type check points to obtain the tree:

Table 7

The resulting array of unordered sequences is represented in the form of a list, as shown in Table VIII below, where the array of unordered sequences also contains the probability of occurrence of eigenvalues and combinations of words, which can be obtained directly from the language model check tree:

Table eight

Since the number of combinations of characters included in the ordered sequence array is relatively large, and the search is not convenient, the text combination in the ordered sequence array is further divided, and the sub-array is divided according to the feature value, for example, in Table VII. The eigenvalues can be divided into one sub-array and divided into two sub-arrays. As shown in the following table IX:

Table 9

For an unordered sequence array, since it contains fewer text combinations, it is not necessary to group them.

Finally, a positioning table is constructed according to each sub-array divided by the above table 9 and the unordered sequence array shown in the above table 8. The feature values corresponding to the sub-array and the starting storage location, and the feature values corresponding to the combination of characters in the unordered sequence array and their probability scores are placed in the positioning table. For example, the obtained positioning table is as shown in Table 10 below.

Table ten

In addition, when constructing the positioning table, the corresponding feature value can be directly used as the subscript of the array, that is, the subscript corresponding to the subarray 1 is K1, the subscript corresponding to the subarray 2 is K2, and the subscript array 1 corresponds to the lower subscript. Marked as K3, the subscript corresponding to the unordered sequence array 2 is K4, then the subscript corresponding to each array is directly stored in the positioning table, and the obtained positioning table is as shown in Table 11 below:

Table XI

Correspondingly, when querying, the language model is searched according to the search tree information, and the language model check tree is queried to determine a probability score of the text recognition result matching the pronunciation information. include:

Querying the positioning table according to the pronunciation information and the corresponding feature value, and determining a sub-array matching the pronunciation information;

The fast sub-array is used to query the matched sub-array to determine a probability score of the text recognition result that matches the pronunciation information.

For example, for the text recognition result corresponding to the pronunciation information “beidaihe”, all the text combinations corresponding to the pronunciations “bei”, “dai”, “he” are obtained from the single characters stored in the first array, and then According to the positioning table query, the probability scores of each word combination of "beidaihe" are obtained. For example, taking the positioning table corresponding to Table 10 as an example, if the probability of P (by | river) is to be checked, the eigenvalue of the text combination "taken river" is calculated by using the above formula 1 and formula 2, and then according to the characteristic value. K2 queries the positioning table of Table 10 above, and knows that the corresponding query range is sub-array 2, according to the start and stop bits of the sub-array 2 recorded in the positioning table, the query is returned in the sub-array 2, and a fast query algorithm can be used (for example, Dichotomy) The search is performed to obtain a probability score of P (by | river) of P15. For example, to check the probability of P (by | Daihe), the eigenvalue of the text combination "Waihe" is calculated by using the above formula 1 and formula 2, and the positioning table of the above table 10 is queried according to the feature value K4. If the corresponding query result is recorded in the positioning table, the probability score of P (by | Daihe) is directly obtained by querying P17. Similarly, the probability score of P (North | Daihe) is P8, and the probability scores of all combinations of words pronounced “beidaihe” are compared. The combination of words is sorted according to the probability score, and the combination of the previous texts is returned to user.

In the above embodiments, the language model is searched according to the query tree information, and the language model check tree is queried to determine a probability score of the text recognition result that matches the pronunciation information, wherein the check tree information includes a text corresponding to the text. Multiple nodes, each node including at least the current node and subsection The storage position offset between the points can be obtained according to the probability score. In the embodiment of the present invention, the language model is directly loaded according to the storage location offset between the current node and the child node, so that the startup time is greatly shortened.

In addition, the foregoing embodiments further construct a positioning table to initially locate the approximate location of the text combination to be queried, and further use a quick query algorithm to accurately find and determine a probability score of the text recognition result that matches the pronunciation information, thereby further improving the query. effectiveness.

Embodiment 2

3 is a schematic structural diagram of a voice recognition apparatus according to Embodiment 2 of the present invention, as shown in FIG. 3, specifically including: a pronunciation information acquisition module 31, a probability score query module 32, and a character recognition module 33;

The pronunciation information obtaining module 31 is configured to identify the pronunciation information according to the voice information;

The probability score query module 32 is configured to load a language model check tree according to the query tree information, and query the language model check tree to determine a probability score of a text recognition result that matches the pronunciation information; wherein the score tree information is Include a plurality of nodes corresponding to the text, each node including at least a storage location offset between the current node and the child node;

The character recognition module 33 is configured to select a character recognition result according to the probability score as a final recognition result.

The voice recognition device according to the embodiment of the present invention is used to perform the voice recognition method described in the foregoing embodiments, and the technical principle and the generated technical effect are similar, and are not described here.

Exemplarily, based on the foregoing embodiment, the device further includes: a cache query module 34 and a trigger module 35;

The cache query module 34 is configured to: after the probability score query module 32 loads the language model check tree according to the query tree information, query the language model check tree to determine the probability score of the text recognition result that matches the pronunciation information before And querying, according to the pronunciation information, a common word sequence stored in the cache and/or a text recognition result of the recorded historical query, a probability score of the text recognition result matching the pronunciation information;

The triggering module 35 is configured to trigger an operation of performing a query in the language model checking tree if the cache query module 34 does not have a probability score of a text recognition result matching the pronunciation information in the cache.

Exemplarily, the device further includes: a first array forming module 36, a second array forming module 37, and a storage module 38;

The first array forming module 36 is configured to: before the cache query module 34 queries the probability score of the text recognition result matching the pronunciation information in the common word sequence in the cache according to the pronunciation information, the language is The probability score of a single character appearing in the model check tree is higher than the single character of the set threshold and its probability score, forming a first array;

The second array forming module 37 is configured to form a second array by combining a combination of a probability score of a combination of characters composed of at least two characters in the language model check tree above a set threshold value and a probability score thereof;

The storage module 38 is configured to store the first array and the second array as the common word sequence.

Exemplarily, the device further includes: an array decomposition module 39 and a positioning table construction module 310;

The array decomposition module 39 is configured to: after the storage module 38 stores the first array and the second array as the common word sequence, the plurality of texts in the second array according to a predetermined rule The word combination is divided into an ordered sequence array and an unordered sequence array, wherein the ordered sequence array includes at least two sub-arrays, and each sub-array stores a plurality of character combinations of the same feature value;

The positioning table structure module 310 is configured to store the probability score in the array of the unordered sequence, and the starting position and/or the ending position, and the feature value, the starting position, and/or the ending position of each sub-array in the positioning. In the table;

Correspondingly, the cache query module 34 is specifically configured to:

Querying the positioning table according to the pronunciation information and the corresponding feature value, determining a sub-array matching the pronunciation information; and querying the matched sub-array by using a fast search algorithm to determine a text recognition result that matches the pronunciation information. Probability score.

Exemplarily, the array decomposition module 39 is specifically configured to:

And the identification value corresponding to the first character in the binary character combination is shifted right by the first specified number of digits and the identification value corresponding to the second character is shifted to the left by the second specified number of digits as the feature value K;

The binary text combination whose number of binary character combinations whose feature value is K is greater than or equal to the preset value is classified into an ordered sequence array; the number of binary character combinations whose feature value is K is less than the preset value The metacharacter combination is classified as an unordered sequence array.

Exemplarily, the array decomposition module 39 is specifically configured to:

Transmitting the feature value K to the right of the first specified number of digits and the identification value corresponding to the third character is shifted to the left by the second specified number of digits as the feature value T;

The ternary character combination whose ternary character combination whose feature value is T is greater than or equal to the preset value is classified into an ordered sequence array; the number of ternary character combinations whose feature value is T is less than the preset value of three The metacharacter combination is classified as an unordered sequence array.

The voice recognition device described in each of the above embodiments is also used to perform the voice described in the above embodiments. The identification method, the technical principle and the generated technical effect are similar, and will not be described here.

Embodiment 3

FIG. 4 is a schematic diagram of a hardware structure of a terminal device for implementing voice recognition according to Embodiment 3 of the present invention, where the terminal device includes one or more processors 41, a memory 42, one or more modules, and the one or more The module (for example, the pronunciation information acquisition module 31, the probability score query module 32, the character recognition module 33, the cache query module 34, the trigger module 35, the first array forming module 36, and the second array in the voice recognition device shown in FIG. The forming module 37, the storage module 38, the array decomposition module 39, and the positioning table construction module 310) are stored in the memory 42; in FIG. 4, a processor 41 is taken as an example; the processor 41 and the memory 42 in the terminal device can pass Bus or other way of connection, in Figure 4 by way of a bus connection.

When executed by the one or more processors 41, the following operations are performed:

Acquiring the pronunciation information according to the voice information;

And loading the language model to check the tree according to the query tree information, and querying the language model to determine a probability score of the text recognition result that matches the pronunciation information; wherein the query tree information includes a plurality of nodes corresponding to the text, Each node includes at least a storage location offset between the current node and the child node;

The character recognition result is selected based on the probability score as the final recognition result.

The foregoing terminal device can perform the methods provided in Embodiment 1 and Embodiment 2 of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

Exemplarily, the processor 41, after querying the loaded language model according to the pronunciation information, and determining the probability score of the character recognition result that matches the pronunciation information, according to the pronunciation information. Querying a probability score of a text recognition result matching the pronunciation information in a common word sequence stored in the cache and/or a text recognition result of the recorded history query; if there is no matching in the cache with the pronunciation information The probability score of the text recognition result triggers the operation of querying in the language model check tree.

Exemplarily, the processor 41 displays the single character in the language model check tree before querying the probability score of the character recognition result matching the pronunciation information in the common word sequence in the cache according to the pronunciation information. The probability score of the single character and its probability score higher than the set threshold value, forming a first array; and the probability score of the combination of at least two characters in the language model check tree is higher than the set threshold The combination of words and their probability scores form a second array; the first array and the second array are stored as the sequence of common words.

Exemplarily, after the processor 41 stores the first array and the second array as the common word sequence, the plurality of text combinations in the second array are divided into an ordered sequence array according to a predetermined rule. And an array of unordered sequences, the ordered sequence array comprising at least two sub-arrays, each of the sub-arrays storing a plurality of combinations of characters having the same eigenvalue; a probability score in the array of the unordered sequences, and a starting position And/or the termination position, and the feature value, the starting position, and/or the ending position of each sub-array are stored in the positioning table; and the positioning table is queried according to the pronunciation information and the corresponding feature value, and the pronunciation information is determined Matching sub-arrays; querying the matched sub-array using a fast lookup algorithm to determine a probability score of the text recognition result that matches the pronunciation information.

Exemplarily, the processor 41 shifts the identifier value corresponding to the first character in the binary character combination to the left of the first specified number of digits and the identifier value corresponding to the second character to the left of the second specified digit as the feature value. K; classify the binary combination of the number of binary combinations with the eigenvalue K greater than or equal to the preset value as an ordered sequence array; the number of binary combination of eigenvalues K is less than the preset value Binary text The combination is classified as an array of unordered sequences.

Exemplarily, the processor 41 shifts the feature value K right by the first specified number of digits and the identifier value corresponding to the third character to the left of the second specified number of digits as the feature value T; The ternary character combination whose number of ternary characters is greater than or equal to the preset value is classified into an ordered sequence array; the ternary combination of the number of ternary characters whose eigenvalue is T is less than the preset value is classified as none An array of ordered sequences.

Embodiment 4

An embodiment of the present invention further provides a non-volatile computer storage medium, where the computer storage medium stores one or more modules, when the one or more modules are executed by a device that performs a voice recognition method, The device performs the following operations:

Acquiring the pronunciation information according to the voice information;

And loading the language model to check the tree according to the query tree information, and querying the language model to determine a probability score of the text recognition result that matches the pronunciation information; wherein the query tree information includes a plurality of nodes corresponding to the text, Each node includes at least a storage location offset between the current node and the child node;

The character recognition result is selected based on the probability score as the final recognition result.

When the module stored in the storage medium is executed by the device, before the score model of the language model that is loaded according to the pronunciation information is queried, and the probability score of the text recognition result matching the pronunciation information is determined, the method preferably includes:

Querying a probability score of a text recognition result matching the pronunciation information in a common word sequence stored in the cache and/or a text recognition result of the recorded history query according to the pronunciation information;

If there is no probability score of the text recognition result matching the pronunciation information in the cache, Then triggering the operation of querying in the language model check tree.

When the module stored in the storage medium is executed by the device, before the probability score of the character recognition result matching the pronunciation information is queried in the common word sequence in the cache according to the pronunciation information, the method preferably includes:

Forming a first array by using a probability that a probability score of a single character in the language model is higher than a set threshold and a probability score thereof;

Forming, by the language model, a combination of at least two characters of a text group having a probability score higher than a set of text values and a probability score thereof to form a second array;

The first array and the second array are stored as the common word sequence.

When the module stored in the storage medium is executed by the device, after storing the first array and the second array as the common word sequence, the method preferably includes:

And dividing a plurality of text combinations in the second array into an ordered sequence array and an unordered sequence array according to a predetermined rule, wherein the ordered sequence array includes at least two sub-arrays, and each sub-array stores the same eigenvalue Multiple text combinations;

Generating a probability score in the array of unordered sequences, and a starting position and/or a ending position, and a feature value, a starting position, and/or a ending position of each sub-array in a positioning table;

Correspondingly, the language model is searched according to the query tree information, and the probability score of the text recognition result that matches the pronunciation information is determined by querying the language model check tree:

Querying the positioning table according to the pronunciation information and the corresponding feature value, and determining a sub-array matching the pronunciation information;

The fast sub-array is used to query the matched sub-array to determine a probability score of the text recognition result that matches the pronunciation information.

When the module stored in the storage medium is executed by the device, dividing the binary text combination in the second array into an ordered sequence array and an unordered sequence array according to a predetermined rule is preferably:

And the identification value corresponding to the first character in the binary character combination is shifted right by the first specified number of digits and the identification value corresponding to the second character is shifted to the left by the second specified number of digits as the feature value K;

Combining a binary combination of two characters of a binary combination having a feature value of K greater than or equal to a preset value into an ordered sequence array;

The binary text combination whose number of binary character combinations whose feature value is K is smaller than the preset value is classified into an unordered sequence array.

When the module stored in the storage medium is executed by the device, the ternary character combination in the second array is divided into an ordered sequence array and an unordered sequence array according to a predetermined rule:

Transmitting the feature value K to the right of the first specified number of digits and the identification value corresponding to the third character is shifted to the left by the second specified number of digits as the feature value T;

The ternary character combination whose number of ternary characters whose feature value is T is greater than or equal to the preset value is classified into an ordered sequence array;

A ternary character combination in which the number of ternary characters whose feature value is T is less than a preset value is classified into an unordered sequence array.

Note that the above are only the preferred embodiments of the present invention and the technical principles applied thereto. Those skilled in the art will appreciate that the present invention is not limited to the specific embodiments described herein, and that various modifications, changes and substitutions may be made without departing from the scope of the invention. Therefore, the present invention has been described in detail by the above embodiments, but the present invention is not limited to the above embodiments, and other equivalent embodiments may be included without departing from the inventive concept. The scope is determined by the scope of the appended claims.

Claims (14)

A speech recognition method, comprising: Acquiring the pronunciation information according to the voice information; And loading the language model to check the tree according to the query tree information, and querying the language model to determine a probability score of the text recognition result that matches the pronunciation information; wherein the query tree information includes a plurality of nodes corresponding to the text, Each node includes at least a storage location offset between the current node and the child node; The character recognition result is selected based on the probability score as the final recognition result. The method according to claim 1, wherein before the querying the loaded language model based on the pronunciation information to determine the probability score of the text recognition result matching the pronunciation information, the method further comprises: Querying a probability score of a text recognition result matching the pronunciation information in a common word sequence stored in the cache and/or a text recognition result of the recorded history query according to the pronunciation information; If there is no probability score of the character recognition result matching the pronunciation information in the cache, an operation of performing a query in the language model check tree is triggered. The method according to claim 2, further comprising: before querying the probability score of the text recognition result matching the pronunciation information in the common word sequence in the cache according to the pronunciation information, further comprising: Forming a first array by using a probability that a probability score of a single character in the language model is higher than a set threshold and a probability score thereof; Forming, by the language model, a combination of at least two characters of a text group having a probability score higher than a set of text values and a probability score thereof to form a second array; The first array and the second array are stored as the common word sequence. The method according to claim 3, wherein after the storing the first array and the second array as the sequence of common words, the method further comprises: And dividing a plurality of text combinations in the second array into an ordered sequence array and an unordered sequence array according to a predetermined rule, wherein the ordered sequence array includes at least two sub-arrays, and each sub-array stores the same eigenvalue Multiple text combinations; Generating a probability score in the array of unordered sequences, and a starting position and/or a ending position, and a feature value, a starting position, and/or a ending position of each sub-array in a positioning table; Correspondingly, the language model is searched according to the query tree information, and the probability score of the text recognition result that is matched by the language model check tree to determine the pronunciation information is: Querying the positioning table according to the pronunciation information and the corresponding feature value, and determining a sub-array matching the pronunciation information; The fast sub-array is used to query the matched sub-array to determine a probability score of the text recognition result that matches the pronunciation information. The method according to claim 4, wherein dividing the binary text combination in the second array into an ordered sequence array and an unordered sequence array according to a predetermined rule comprises: And the identification value corresponding to the first character in the binary character combination is shifted right by the first specified number of digits and the identification value corresponding to the second character is shifted to the left by the second specified number of digits as the feature value K; Combining a binary combination of two characters of a binary combination having a feature value of K greater than or equal to a preset value into an ordered sequence array; The binary text combination whose number of binary character combinations whose feature value is K is smaller than the preset value is classified into an unordered sequence array. The method of claim 5 wherein said second number is based on a predetermined rule The ternary text combination in the group is divided into an ordered sequence array and an unordered sequence array including: Transmitting the feature value K to the right of the first specified number of digits and the identification value corresponding to the third character is shifted to the left by the second specified number of digits as the feature value T; The ternary character combination whose number of ternary characters whose feature value is T is greater than or equal to the preset value is classified into an ordered sequence array; A ternary character combination in which the number of ternary characters whose feature value is T is less than a preset value is classified into an unordered sequence array. A speech recognition device, comprising: a pronunciation information obtaining module, configured to identify the pronunciation information according to the voice information; a probability score query module, configured to load a language model check score tree according to the check score tree information, and query the language model check score tree to determine a probability score of a text recognition result that matches the pronunciation information; wherein the check score tree information includes a plurality of nodes corresponding to the text, each node including at least a storage location offset between the current node and the child node; A text recognition module is configured to select a text recognition result according to the probability score as a final recognition result. The device according to claim 7, wherein the device further comprises: a cache query module, configured to: after the probability score query module loads the language model check tree according to the check tree information, query the language model check tree to determine a probability score of the text recognition result that matches the pronunciation information, according to the The pronunciation information of the common pronunciation word stored in the cache and/or the text recognition result of the recorded historical query is used to query the probability score of the text recognition result matching the pronunciation information; a triggering module, if the cache query module does not exist in the cache and the pronunciation information The probability score of the assigned text recognition result triggers the operation of querying in the language model check tree. The device according to claim 8, wherein the device further comprises: a first array forming module, configured to check the language model before the cache query module searches for a probability score of a text recognition result matching the pronunciation information in a common word sequence in the cache according to the pronunciation information. The probability that a single character appears in a single character above a set threshold and its probability score forms a first array; a second array forming module, configured to form a second array by combining a combination of a probability score of a combination of at least two characters in the language model scoring tree and a probability score higher than a set threshold; and a probability score; And a storage module, configured to store the first array and the second array as the common word sequence. The device according to claim 9, wherein the device further comprises: An array decomposition module, configured to: after the storage module stores the first array and the second array as the common word sequence, divide the plurality of text combinations in the second array into ordered according to a predetermined rule An array of sequences and an array of unordered sequences, the ordered sequence array comprising at least two sub-arrays, each of which stores a plurality of combinations of characters having the same eigenvalue; a positioning table module for storing a probability score in the array of unordered sequences, and a starting position and/or a ending position, and a feature value, a starting position, and/or a ending position of each sub-array in the positioning table ; Correspondingly, the cache query module is specifically configured to: Querying the positioning table according to the pronunciation information and the corresponding feature value, determining a sub-array matching the pronunciation information; and querying the matched sub-array by using a fast search algorithm to determine a text recognition result that matches the pronunciation information. Probability score. The apparatus according to claim 10, wherein said array decomposition module is specifically used to: And the identification value corresponding to the first character in the binary character combination is shifted right by the first specified number of digits and the identification value corresponding to the second character is shifted to the left by the second specified number of digits as the feature value K; The binary text combination whose number of binary character combinations whose feature value is K is greater than or equal to the preset value is classified into an ordered sequence array; the number of binary character combinations whose feature value is K is less than the preset value The metacharacter combination is classified as an unordered sequence array. The apparatus according to claim 11, wherein the array decomposition module is specifically configured to: Transmitting the feature value K to the right of the first specified number of digits and the identification value corresponding to the third character is shifted to the left by the second specified number of digits as the feature value T; The ternary character combination whose ternary character combination whose feature value is T is greater than or equal to the preset value is classified into an ordered sequence array; the number of ternary character combinations whose feature value is T is less than the preset value of three The metacharacter combination is classified as an unordered sequence array. A terminal device for implementing voice recognition, comprising: One or more processors; Memory One or more modules, the one or more modules being stored in the memory, and when executed by the one or more processors, performing the following operations: Acquiring the pronunciation information according to the voice information; And loading the language model to check the tree according to the query tree information, and querying the language model to determine a probability score of the text recognition result that matches the pronunciation information; wherein the query tree information includes a plurality of nodes corresponding to the text, Each node includes at least a storage location between the current node and the child node Offset; The character recognition result is selected based on the probability score as the final recognition result. A non-volatile computer storage medium storing one or more modules, wherein when the one or more modules are executed by a device that performs a voice recognition method, the device is caused Do the following: Acquiring the pronunciation information according to the voice information; And loading the language model to check the tree according to the query tree information, and querying the language model to determine a probability score of the text recognition result that matches the pronunciation information; wherein the query tree information includes a plurality of nodes corresponding to the text, Each node includes at least a storage location offset between the current node and the child node; The character recognition result is selected based on the probability score as the final recognition result.

Images