CN101464896A - Voice fuzzy retrieval method and apparatus - Google Patents

Abstract

The invention discloses a speech fuzzy retrieval method and device, wherein the method includes the following steps: using a preset acoustic model and a language model to perform speech recognition on an acquired speech signal to obtain a recognition result; using a preset index table according to the The recognition result is searched in the preset text entry library to obtain the primary selection item; the string fuzzy matching is performed on the primary selection item and the recognition result, and the entry whose matching degree is within the preset matching degree threshold range is selected as the Select the items, and record the matching positions of each item at the same time; calculate the posterior probability between the selected item matching part of the text and the speech signal, and finally use the posterior probability and the matching ratio obtained by the matching position to select several items as Search results for speech signals. By adopting the present invention, text entries matching the speech signal can be quickly and accurately retrieved based on the speech signal on a massive text entry database.

Description

Speech fuzzy retrieval method and device

technical field

The invention relates to the fields of speech recognition and retrieval, in particular to a speech fuzzy retrieval method and device.

Background technique

As a branch of multimedia retrieval technology, speech fuzzy retrieval is different from traditional text retrieval and audio retrieval. Retrieval, that is, how to retrieve text information related to its content in the text library according to a piece of voice signal submitted by the user.

Speech recognition technology can convert speech signals into text content. If the converted text is used and the text retrieval method is used for reference, the retrieval of audio from the text library can be realized. However, speech recognition technology cannot be 100% accurate, especially for spoken language For speech, the recognition accuracy rate is usually lower than 90%. It is conceivable that if non-accurate text is used to retrieve a massive text entry library, the retrieval results will be even more inaccurate.

Contents of the invention

The invention provides a voice fuzzy retrieval method and device to solve the problem of inaccurate retrieval existing in the existing voice recognition technology.

For this reason, the embodiment of the present invention adopts following technical scheme:

A speech fuzzy retrieval method, comprising:

Use the preset acoustic model and language model to perform speech recognition on the acquired speech signal, and obtain the recognition result;

Using the preset index table to search in the preset text item library according to the recognition result to obtain the primary selection item;

performing string fuzzy matching on the preliminary selection item and the recognition result, selecting selected items whose matching degree is within a preset matching degree threshold range, and recording the matching position at the same time;

Calculate the posterior probability between the selected item matching part of the text and the speech signal, and use the posterior probability and the matching ratio obtained through the matching position to select several entries as the retrieval result of the speech signal.

The method also includes:

According to the text items to be retrieved, the index table is established with syllables, characters or words as index units for one-level or multi-level indexing.

The method also includes:

The language model is trained in whole or in part by using the preset text entry library.

in:

The form of the recognition result includes the most probable text string corresponding to the speech signal, the most probable various text strings corresponding to the speech signal, and the word graph corresponding to the speech signal.

The specific process of using the preset index table to search in the preset text entry library according to the recognition result to obtain the primary selection entry is as follows:

Using a preset index table to vote for each word/word in the recognition result, and selecting an entry whose vote count is higher than the preset vote count threshold as the primary selection entry;

Wherein, the voting refers to using the words/phrases in the recognition result to look up the index items of the index table, and after finding the index items, adding 1 to the number of votes for each item included in the index.

The matching algorithm of the fuzzy matching adopts a dynamic programming calculation method of editing distance between texts based on a confusion matrix, wherein the confusion matrix is obtained through training or preset, and optimizes the cost of replacement, insertion, and deletion.

A speech fuzzy retrieval device, comprising:

A voice signal acquiring unit, configured to acquire a voice signal;

The recognition unit is used to use the preset acoustic model and language model to perform speech recognition on the acquired speech signal to obtain a recognition result;

A retrieval unit, configured to use a preset index table to search in a preset text entry library according to the recognition result to obtain primary selection entries;

A fuzzy matching unit, configured to fuzzy character string match the primary selection item and the recognition result, select selected items whose matching degree is within a preset matching degree threshold range, and record the matching position;

The result determining unit is used to calculate the posterior probability between the matching part of the selected item and the speech signal, and select several entries as the retrieval result of the speech signal by using the posterior probability and the matching ratio obtained through the matching position.

The unit also includes:

The index table building unit is used to build the index table with syllables, words or words as index units according to the preset text entry library to be retrieved, and the index table is used for one-level or multi-level indexing.

The unit also includes:

A language model building unit, configured to use the preset text entry library to train part or all of the language model.

The retrieval unit includes:

The index voting subunit is used to vote each word/word in the recognition result by using a preset index table, wherein the voting refers to searching the index item of the index table with the word/word in the recognition result, and the query is reversed. After the index item, add 1 to the number of votes for each item included in the index;

The primary selection item selection subunit is configured to select an item whose vote count is higher than a preset vote count threshold as the primary selection item.

It can be seen that the present invention proposes a brand-new speech fuzzy retrieval mode, which overcomes incomplete correctness by applying steps such as relevant language model, index voting, character string fuzzy matching, selected items and posterior probability calculation of speech signals. The adverse effect of the speech recognition results on the text database retrieval has realized the rapid and accurate retrieval of the speech signal on the massive text entry database.

Description of drawings

Fig. 1 is the flowchart of speech fuzzy retrieval method of the present invention;

Fig. 2 is the flowchart of the method embodiment of the present invention;

Fig. 3 is a schematic structural diagram of the speech fuzzy retrieval device of the present invention.

Detailed ways

The voice fuzzy retrieval scheme provided by the present invention adds a suitable language model to improve the accuracy rate during recognition, performs character string fuzzy matching when using the recognition result as text retrieval to reduce the impact of recognition errors, and calculates candidate keywords as The posterior probability of audio content is verified, thereby greatly improving the accuracy and reliability of retrieval.

Referring to Fig. 1, it is the speech fuzzy retrieval method flowchart of the present invention, comprises the following steps:

S101: Perform speech recognition on the acquired speech signal by using a preset acoustic model and language model, and obtain a recognition result;

S102: Use the preset index table to search in the preset text item library according to the recognition result to obtain the primary selection item;

Wherein, the preset text entry library is generally a massive text entry library, including a large amount of text entry information.

S103: Perform character string fuzzy matching on the preliminary selection item and the recognition result, select selected items whose matching degree is within a preset matching degree threshold range, and record the matching position at the same time;

S104: Calculate the posterior probability between the matching part of the selected item and the speech signal, and use the posterior probability and the matching ratio obtained through the matching position to select several entries as the retrieval result of the speech signal.

The present invention will be described in detail below in conjunction with specific examples.

Referring to Fig. 2, it is a flow chart of a specific embodiment method for voice retrieval of a large amount of text entry database using voice fuzzy retrieval technology, including:

S201: Obtain a voice signal input by a user;

S202: Perform speech recognition on the acquired speech signal by using the pre-established acoustic model and language model, and obtain a recognition result;

S203: Use the preset index table to perform a quick search in the preset text item library according to the recognition result to obtain the primary selection item;

Before starting to build a speech fuzzy retrieval system, it is necessary to establish a suitable speech model and an index table of a massive text entry library in advance.

Because it is necessary to retrieve the text containing speech content in the massive text entry library, the speech content is very likely to exist in the massive text entry library, and it is a certain entry or a part of a certain entry. Therefore, according to the massive text entry library A language model trained for a corpus is an application-dependent language model that is better suited to retrieval tasks.

For the preset index table, it consists of two parts: the index item and the corresponding content of the index item. The index item of the index table in the present invention is a character or a word, and the corresponding content of the index item is the text containing the word or word in the massive text entry library. Usually, one index item corresponds to multiple texts. For example, the content corresponding to the index item "中" includes "Communist Party of China", "People's Republic of China", "Our Greater China" and so on.

Therefore, when performing speech recognition on the input speech in S202, the application-related language model trained in S203 can be added to improve the accuracy of recognition, and a recognition result with high accuracy can be obtained in S202.

The recognition result is the character expression form after the speech signal is decoded. The commonly used forms are: the most likely text string corresponding to the input speech signal (that is, there is only one recognition result, such as "People's Republic of China"), and the most likely N types Text strings (that is, multiple recognition results, such as 3 kinds of recognition results: "Communist Party of China", "People's Republic of China" and "Our Great China"), and word graphs corresponding to voice signals. Ring graphs represent all possible text strings, word graphs are the most efficient form of recognition results, and they contain the most abundant information.

In S203, for each character/phrase in the recognition result obtained in S202, index voting is performed using a preset index table. The so-called voting means that the word/phrase in the recognition result is used to look up the index items of the index table, and after querying the inverted index items, the number of votes for the corresponding text is increased by 1. For example, if the word "中" is included in the recognition result, the number of votes corresponding to all texts containing "中", such as the Communist Party of China", "People's Republic of China" and "Our Great China", will be increased by 1. The higher the number of votes The text, the higher the degree of matching with the recognition result. Keep the text with the number of votes higher than the threshold as the primary selection item.

S204: Perform character string fuzzy matching on the preliminary selection items and the recognition results, sort the matched items according to the matching degree from high to low, and keep only the selected items whose matching degree is within the matching degree threshold;

Since the speech recognition technology cannot guarantee a 100% correct rate, there are some errors in the recognition results. Moreover, the index table only records those words/words contained in the text, and does not have the position information of the words/words. Therefore, the indexed Preliminary items cannot be directly used as search results.

Therefore, using the string fuzzy matching technology, the matching degree between the primary selection item and the recognition result is obtained. Compared with string precision matching, fuzzy matching allows the substring to be different from the main string. The two main methods of character string fuzzy matching are the bit vector method and the filter method at present, and the present invention can adopt the existing method to carry out. The simplest fuzzy matching algorithm is based on the edit distance of dynamic programming. There are three kinds of errors in the matching: deletion, insertion and substitution. Each error can define different error costs according to the actual application. For the correct matching part, the error cost is usually defined as zero. In the present invention, both the recognition result and the text in the massive text entry database can be regarded as a certain character representation, and the substring is the recognition result, and the main string is the entry in the massive text entry database. The degree of matching is inversely proportional to the error cost. Since the voice signal input by the user may be a text fragment in a massive text entry library, string fuzzy matching not only gives the matching degree, but also gives the most likely matching position.

S205: Calculate the posterior probability of each qualified selected item being the input audio content; at the same time, record the matching position;

Since the selected entry obtained in step S204 is compared with the recognition result at the character level, and the recognition result itself contains certain errors, a high degree of matching does not necessarily mean that it is more likely to be the actual content of the speech. Therefore, in S205, the posterior probability of the selected item under the given speech signal condition is calculated. The posterior probability is a value between 0 and 1, and the sum of the posterior probabilities of all selected items is 1. The greater the posterior probability is, the greater the possibility is that the corresponding item is indeed speech content. Posterior probability refers to the probability of re-correction after obtaining the "result" information. For example, in the Bayesian formula, it is the "cause" in the problem of "finding the cause of the fruit". The prior probability and the posterior probability are inseparable. The calculation of the posterior probability is based on the prior probability. The calculation method of the posterior probability is a mature prior art, and no further description is given here.

S206: Using the posterior probability and the matching ratio obtained through the matching position, select several items as the retrieval result of the speech signal, and then end the process.

Among them, by weighting the posterior probability and matching ratio, the items with relatively high posterior probability and matching ratio can be finally selected as the retrieval result.

Corresponding to the above method, the present invention provides a speech fuzzy retrieval device, which can be realized by software, hardware or a combination of software and hardware.

Referring to Fig. 3, it is a schematic diagram of the internal structure of the device, including: a voice signal acquisition unit 300, an identification unit 301, a retrieval unit 302, a fuzzy matching unit 303 and a result determination unit 304, wherein:

A voice signal acquisition unit 300, configured to acquire a voice signal;

The recognition unit 301 is configured to use a preset acoustic model and a language model to perform speech recognition on the speech signal obtained by the speech signal obtaining unit 300, to obtain a recognition result;

The retrieval unit 302 is configured to use the preset index table to search in the preset text entry library according to the recognition result obtained by the recognition unit 301 to obtain the primary selection entry;

The fuzzy matching unit 303 is used to perform character string fuzzy matching on the preliminary selection items obtained by the retrieval unit 302 and the recognition results obtained by the recognition unit 301, select selected entries whose matching degree is within a preset matching degree threshold range, and record the matching Location;

The result determination unit 304 is used to calculate the posterior probability between the selected items matched by the fuzzy matching unit 303 and the speech signal, and select several entries as the speech signal by using the posterior probability and the matching ratio obtained through the matching position search results.

Preferably, the device also includes:

An index table building unit 305, configured to build an index table with syllables, characters or words as index units according to the preset text entries.

Preferably, the device also includes:

A language model building unit 306, configured to use the preset text entry library to train and obtain a language model.

Preferably, the retrieval unit 302 further includes:

An index voting subunit (not shown in the figure) is used to vote for each word/word in the recognition result using a preset index table, wherein the voting refers to searching the index with the word/word in the recognition result The index item of the table, after querying the inverted index item, the number of votes for each item included in the index is increased by 1;

The primary selection item selection subunit (not shown in the figure) is configured to select an item whose vote count is higher than a preset vote count threshold as the primary selection item.

For the implementation details of the device provided by the present invention, reference may be made to the method embodiments, which will not be repeated here.

It can be seen that the present invention proposes a brand-new speech fuzzy retrieval scheme, which overcomes the incomplete correctness by applying steps such as relevant language model, index voting, character string fuzzy matching, candidate text and posterior probability calculation of speech signal. The unfavorable impact of speech recognition results on text database retrieval enables fast and accurate retrieval of speech signals on massive text entry databases.

Those of ordinary skill in the art can understand that the process of realizing the method of the above-mentioned embodiment can be completed by the related hardware of the program instruction, and the described program can be stored in a readable storage medium, and the program executes the above-mentioned method when executed. Corresponding steps. The storage medium may be, for example: ROM/RAM, magnetic disk, optical disk, etc.

The above is only a preferred embodiment of the present invention, it should be pointed out that, for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (10)

1, a kind of voice fuzzy retrieval method is characterized in that, comprising:

Acoustic model that utilization is preset and language model carry out speech recognition to the voice signal that obtains, and obtain recognition result;

The concordance list that utilization is preset is retrieved in the textual entry storehouse of presetting according to described recognition result, obtains the primary election clauses and subclauses;

Described primary election clauses and subclauses and described recognition result are carried out the character string fuzzy matching, choose the selected clauses and subclauses of matching degree in the matching degree threshold range that presets, write down matched position simultaneously;

Calculate the posterior probability between selected entries match part text and described voice signal, utilize posterior probability and select the result for retrieval of several clauses and subclauses as voice signal by the matching ratio that described matched position obtains.

2, according to the described method of claim 1, it is characterized in that, also comprise: is that indexing units is set up described concordance list according to textual entry to be retrieved with syllable, word or speech, in order to carry out one or more levels index.

3, according to the described method of claim 2, it is characterized in that, also comprise: described language model is all or part of to utilize described textual entry storehouse training of presetting to obtain.

According to the described method of claim 1, it is characterized in that 4, the form of described recognition result comprises the most probable text strings of voice signal correspondence, the most possible kinds of words string of voice signal correspondence, and the speech figure of voice signal correspondence.

According to the described method of claim 1, it is characterized in that 5, the concordance list that described utilization is preset is retrieved the detailed process that obtains the primary election clauses and subclauses according to described recognition result and is in the textual entry storehouse of presetting:

The concordance list that utilization is preset is voted to each character/word in the recognition result, chooses votes and is higher than the clauses and subclauses of the votes threshold value that presets as described primary election clauses and subclauses;

Wherein, described ballot is meant the index entry of searching concordance list with the character/word in the recognition result, inquire index entry after, each clauses and subclauses votes that this index is included all adds 1.

6, according to the described method of claim 1, it is characterized in that, the matching algorithm of described fuzzy matching adopts based on editing distance dynamic programming computing method between the text of confusion matrix, wherein, described confusion matrix obtains or preestablishes by training, is optimized replacing, insert, delete cost.

7, a kind of voice fuzzy indexing unit is characterized in that, comprising:

The voice signal acquiring unit is used to obtain voice signal;

Recognition unit is used to utilize the acoustic model and the language model that preset that the voice signal that obtains is carried out speech recognition, obtains recognition result;

Retrieval unit is used for utilizing the concordance list that presets to retrieve in the textual entry storehouse of presetting according to described recognition result, obtains the primary election clauses and subclauses;

The fuzzy matching unit is used for described primary election clauses and subclauses and described recognition result are carried out the character string fuzzy matching, chooses the selected clauses and subclauses of matching degree in the matching degree threshold range that presets, and the record matched position;

Determining unit is used to calculate the compatible portion of selected clauses and subclauses and the posterior probability between described voice signal as a result, utilizes posterior probability and selects the result for retrieval of several clauses and subclauses as voice signal by the matching ratio that described matched position obtains.

8, according to the described device of claim 7, it is characterized in that, also comprise:

Concordance list is set up the unit, and being used for according to the textual entry storehouse of presetting to be retrieved is that indexing units is set up described concordance list with syllable, word or speech, and described concordance list is in order to carry out one or more levels index.

9, described according to Claim 8 device is characterized in that, also comprises:

Language model is set up the unit, is used to utilize described textual entry storehouse training of presetting to obtain the part or all of of described language model.

10, according to claim 7,8 or 9 described devices, it is characterized in that described retrieval unit comprises:

Index ballot subelement, be used for utilizing the concordance list that presets that each character/word of recognition result is voted, wherein, described ballot is meant the index entry of searching concordance list with the character/word in the recognition result, after inquiring about index entry, each clauses and subclauses votes that this index is included all adds 1;

The primary election clauses and subclauses are chosen subelement, are used to choose votes and are higher than the clauses and subclauses of the votes threshold value that presets as described primary election clauses and subclauses.

Images