CN111597798A - Method and system for improving identification accuracy of dynamic model - Google Patents
Method and system for improving identification accuracy of dynamic model Download PDFInfo
- Publication number
- CN111597798A CN111597798A CN202010326352.1A CN202010326352A CN111597798A CN 111597798 A CN111597798 A CN 111597798A CN 202010326352 A CN202010326352 A CN 202010326352A CN 111597798 A CN111597798 A CN 111597798A
- Authority
- CN
- China
- Prior art keywords
- vocabulary
- optimized
- weight
- false alarm
- alarm rate
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/237—Lexical tools
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/237—Lexical tools
- G06F40/247—Thesauruses; Synonyms
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L13/00—Speech synthesis; Text to speech systems
- G10L13/02—Methods for producing synthetic speech; Speech synthesisers
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The invention provides a method and a system for improving identification accuracy of a dynamic model, wherein the method comprises the following steps: firstly, sorting and summarizing word lists with poor recognition effects and submitting the word lists to a language model dynamic optimization platform to serve as word lists to be optimized; then, respectively obtaining a forward set and a reverse set, a reference generation efficiency and a false alarm rate aiming at the vocabulary to be optimized; carrying out default weight compiling, validation and testing on the vocabulary to be optimized, and calculating to obtain the validation rate and the false alarm rate of the vocabulary to be optimized after the testing; then, adjusting the vocabulary to be optimized by taking the false alarm rate as a reference to adjust the vocabulary entry weight, obtaining the vocabulary after adjusting the weight, and compiling and validating the vocabulary; and finally, circularly calling the comprehensive processing module and the word list compiling processing module, selecting a group of word list weights with the highest average effective rate of the word lists to be optimized within the allowable range of the false alarm rate as final weight output, and issuing the final weight output to the dynamic language model. The system comprises modules corresponding to the method steps.
Description
Technical Field
The invention provides a method and a system for improving the identification accuracy of a dynamic model, and belongs to the technical field of language identification.
Background
The current language model mainly comprises a static model, namely a general language model, and a dynamic model, namely a user-defined model. And when the ASR is recognized, packing the dynamic model and the static model, and sending the dynamic model and the static model to an ASR engine for recognition. The static model has a long online optimization release period, and some hot words cannot be identified in time, so that certain problems are brought to user identification. The problem can be solved by customizing and optimizing the dynamic model, but the problem that partial entries are not effective or are not recognized disorderly after being implemented exists. After the dynamic model implements the entries, the recognition is not effective or is disorderly recognized, so that the expected effect is not achieved after the implementation. For example, a brand name [ carrying technology ] is implemented, and when a user says [ carrying technology ] the recognition result is [ carrying technology ] and the term of carrying technology implemented is not valid.
Disclosure of Invention
The invention provides a method and a system for improving identification accuracy of a dynamic model, which are used for solving the problem that the dynamic model is not effective or disorderly identified after part of entries are implemented in the prior art, and adopt the following technical scheme:
a system for enhancing dynamic model identification accuracy by dynamically lexical weighting, the system comprising:
the import module is used for sorting and summarizing word lists with poor recognition effects and submitting the word lists to a language model dynamic optimization platform;
the comprehensive processing module is used for respectively utilizing a TTS voice synthesis method and a screening and removing method to obtain a forward set and a reverse set aiming at the vocabulary to be optimized, and utilizing the forward set and the reverse set to obtain the reference generation efficiency and the false alarm rate of the vocabulary to be optimized; carrying out default weight compiling, validation and testing on the vocabulary to be optimized, and calculating to obtain the validation rate and the false alarm rate of the vocabulary to be optimized after the testing;
the vocabulary compiling processing module is used for adjusting the vocabulary to be optimized by taking the false alarm rate as a reference to adjust the vocabulary entry weight, obtaining the vocabulary after the weight is adjusted and compiling and validating the vocabulary;
and the final weight acquisition module is used for circularly calling the comprehensive processing module and the word list compiling processing module, selecting a group of word list weights with the highest average effective rate of the word lists to be optimized within the allowable range of the false alarm rate as final weights to be output and issuing the final weights to the dynamic language model.
Further, the import module includes:
the to-be-optimized word list obtaining module is used for screening out word lists with poor recognition effects and sorting and summarizing the word lists;
and the vocabulary import module is used for importing the sorted and summarized vocabularies to the language model dynamic optimization platform and taking the vocabularies imported to the language model dynamic optimization platform as the vocabularies to be optimized.
Further, the integrated processing module comprises:
the forward set acquisition module is used for synthesizing a plurality of sentence patterns which are set in advance by the vocabulary to be optimized and the language model dynamic optimization platform by using a TTS (text to speech) speech synthesis method to generate a forward set;
the reverse set acquisition module is used for forming a reverse set by utilizing the fixed phrase tone test set;
the basic reference acquisition module is used for carrying out test calculation on the forward set and the adjusted reverse set to obtain the generation efficiency and the false alarm rate of entries in the vocabulary to be optimized when the optimization implementation is not carried out, and taking the generation efficiency and the false alarm rate when the optimization implementation is not carried out as the basic generation efficiency and the basic false alarm rate;
the weight compiling and validating module is used for compiling and validating each entry in the vocabulary to be optimized by default weight;
the generating efficiency obtaining module is used for testing the forward set and the adjusted reverse set and calculating to obtain the effective rate of each entry in the vocabulary to be optimized after the test;
and the false alarm rate acquisition module is used for testing the forward set and the adjusted reverse set and calculating to obtain the false alarm rate of each entry in the vocabulary to be optimized after the test.
Further, the forward set acquisition module comprises:
the sentence pattern generation module is used for judging the category of the vocabulary to be optimized by utilizing the language dynamic model optimization platform and setting a plurality of sentence patterns according to the category of the vocabulary to be optimized;
the combination module is used for combining the sentence patterns and the word list to be optimized to obtain a combination unit;
and the forward set generating module is used for carrying out voice synthesis on the combination unit by using a TTS voice synthesis method and taking the synthesized voice as a forward set.
Further, the reverse set acquisition module comprises:
the voice elimination module is used for eliminating the voice containing the vocabulary to be optimized in the short voice test set;
and the reverse set generation module is used for forming a reverse set by the phrase sound test set after the speech containing the vocabulary to be optimized is eliminated.
Further, the vocabulary compiling processing module comprises:
the weight adjusting module is used for adjusting the weight of the entry to be optimized and lower than the reference generation efficiency by taking the false alarm rate as a reference to obtain a word list after the weight is adjusted;
and the vocabulary compiling and validating module is used for compiling and validating the vocabulary after the weight adjustment.
Further, the weight adjusting module comprises:
the weight increasing module is used for continuously increasing the weight of the entry with the effective rate lower than the reference generation efficiency in the vocabulary to be optimized within the allowable range of the false alarm rate by taking the false alarm rate as the reference, and the weight increasing range is [0,2 ];
the false alarm rate judging module is used for judging whether the false alarm rate of the entry with the increased weight exceeds an allowable range or not, and if the false alarm rate exceeds the allowable range, the weight reducing module is started;
the weight reduction module is used for continuously reducing the weight of the entries on the basis of the historical weight formed after the weights of the entries are increased until the false alarm rate of the entries is recovered to an allowable range;
and the vocabulary compiling and acquiring module is used for acquiring vocabulary compiling corresponding to the entry after the weight is adjusted.
A method according to any one of claims 1 to 7, the method comprising:
sorting and summarizing word lists with poor recognition effects, and submitting the word lists to a language model dynamic optimization platform;
respectively utilizing a TTS speech synthesis method and a screening and removing method to obtain a forward set and a reverse set aiming at a vocabulary to be optimized, and utilizing the forward set and the reverse set to obtain the reference generation efficiency and the false alarm rate of the vocabulary to be optimized; carrying out default weight compiling, validation and testing on the vocabulary to be optimized, and calculating to obtain the validation rate and the false alarm rate of the vocabulary to be optimized after the testing;
adjusting a vocabulary to be optimized by taking the false alarm rate as a reference to adjust the vocabulary entry weight, obtaining the vocabulary after adjusting the weight, and compiling and validating the vocabulary;
and circularly calling the comprehensive processing module and the word list compiling processing module, selecting a group of word list weights with the highest average effective rate of the word lists to be optimized within the allowable range of the false alarm rate as final weight output, and issuing the final weight output to the dynamic language model.
Further, the specific process of the method comprises the following steps:
step 1, sorting and summarizing word lists with poor recognition effects, and submitting the word lists to a language model dynamic optimization platform to serve as word lists to be optimized;
step 2, importing the vocabulary to be optimized into a dynamic model implementation platform;
step 3, setting a plurality of sentence patterns in advance by using a language model dynamic optimization platform according to the category of the vocabulary to be optimized, combining the vocabulary to be optimized and the sentence patterns, and synthesizing voice as a forward set through TTS;
step 4, eliminating the voice containing the vocabulary to be optimized in the short voice test set, and forming a reverse set by the phrase voice test set after the voice containing the vocabulary to be optimized is eliminated;
step 5, testing and calculating the forward set and the adjusted reverse set to obtain the generation efficiency and the false alarm rate of the entries in the vocabulary to be optimized when the entries are not optimized, and taking the generation efficiency and the false alarm rate as the reference generation efficiency and the false alarm rate;
step 6, compiling and validating each entry in the vocabulary to be optimized by default weight, wherein the default weight is 0;
step 7, testing the forward set and the adjusted reverse set, and calculating to obtain the effective rate and the false alarm rate of each entry in the vocabulary to be optimized after the test;
step 8, giving priority to false alarm rate, and continuously increasing the weight of the entry with poor generation efficiency within the allowable range of the false alarm rate, wherein the weight range is [0,2 ]; if the false alarm rate exceeds the allowable range, continuously reducing the entry weight on the basis of the entry historical weight until the false alarm rate is restored to the allowable range, and finally obtaining a vocabulary after the weight is adjusted;
step 9, compiling and validating the vocabulary after the weight is adjusted, which is obtained in the step 8;
and step 10, circularly calling the step 7, the step 8 and the step 9 for a plurality of times, selecting a group of vocabulary weights with the highest average effective rate of the to-be-optimized vocabulary as final weights within the allowable range of the false alarm rate, and outputting the final weights to the dynamic language model.
The invention has the beneficial effects that:
the method and the system for improving the identification accuracy of the dynamic model add the weight adjustable function to the dynamic model so as to solve the problem that entries are not effective or are not common. Under the condition of ensuring that the reverse set is not spurious, namely the false alarm rate is controlled to be in accordance with a set range, the weight of each entry is continuously improved, the entry generation efficiency is improved, and the recognition result is optimal.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
fig. 2 is a schematic structural diagram of the system of the invention.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.
The embodiment of the invention provides a method and a system for improving the identification accuracy of a dynamic model, aiming at solving the problems that the dynamic model has no effect or disorderly identification after part of entries are implemented in the prior art,
a method for improving accuracy of dynamic model identification by dynamically adjusting weights according to word expressions, as shown in fig. 1, the method comprising:
firstly, sorting and summarizing word lists with poor recognition effects, and submitting the word lists to a language model dynamic optimization platform;
then, respectively utilizing a TTS speech synthesis method and a screening and removing method to obtain a forward set and a reverse set aiming at the vocabulary to be optimized, and utilizing the forward set and the reverse set to obtain the reference generation efficiency and the false alarm rate of the vocabulary to be optimized; carrying out default weight compiling, validation and testing on the vocabulary to be optimized, and calculating to obtain the validation rate and the false alarm rate of the vocabulary to be optimized after the testing;
then, adjusting the vocabulary to be optimized by taking the false alarm rate as a reference to adjust the vocabulary entry weight, obtaining the vocabulary after adjusting the weight, and compiling and validating the vocabulary;
and finally, circularly calling the comprehensive processing module and the word list compiling processing module, selecting a group of word list weights with the highest average effective rate of the word lists to be optimized within the allowable range of the false alarm rate as final weight output, and issuing the final weight output to the dynamic language model.
The method comprises the following specific processes:
step 1, sorting and summarizing word lists with poor recognition effects, and submitting the word lists to a language model dynamic optimization platform to serve as word lists to be optimized;
step 2, importing the vocabulary to be optimized into a dynamic model implementation platform;
step 3, setting a plurality of sentence patterns in advance by using a language model dynamic optimization platform according to the category of the vocabulary to be optimized, combining the vocabulary to be optimized and the sentence patterns, and synthesizing voice as a forward set through TTS;
step 4, eliminating the voice containing the vocabulary to be optimized in the short voice test set, and forming a reverse set by the phrase voice test set after the voice containing the vocabulary to be optimized is eliminated;
step 5, testing and calculating the forward set and the adjusted reverse set to obtain the generation efficiency and the false alarm rate of the entries in the vocabulary to be optimized when the entries are not optimized, and taking the generation efficiency and the false alarm rate as the reference generation efficiency and the false alarm rate;
step 6, compiling and validating each entry in the vocabulary to be optimized by default weight, wherein the default weight is 0;
step 7, testing the forward set and the adjusted reverse set, and calculating to obtain the effective rate and the false alarm rate of each entry in the vocabulary to be optimized after the test;
step 8, giving priority to false alarm rate, and continuously increasing the weight of the entry with poor generation efficiency within the allowable range of the false alarm rate, wherein the weight range is [0,2 ]; if the false alarm rate exceeds the allowable range, continuously reducing the entry weight on the basis of the entry historical weight until the false alarm rate is restored to the allowable range, and finally obtaining a vocabulary after the weight is adjusted;
step 9, compiling and validating the vocabulary after the weight is adjusted, which is obtained in the step 8;
and step 10, circularly calling the step 7, the step 8 and the step 9 for a plurality of times, selecting a group of vocabulary weights with the highest average effective rate of the to-be-optimized vocabulary as final weights within the allowable range of the false alarm rate, and outputting the final weights to the dynamic language model.
The principle and the technical effect of the scheme are as follows: the problem that entries are not effective or are not common is solved by adding a weight adjustable function in the dynamic model. Under the condition of ensuring that the reverse set is not spurious, namely the false alarm rate is controlled to be in accordance with a set range, the weight of each entry is continuously improved, the entry generation efficiency is improved, and the recognition result is optimal.
A system for enhancing dynamic model identification accuracy by dynamically lexical weighting, as shown in fig. 2, the system comprising:
the import module is used for sorting and summarizing word lists with poor recognition effects and submitting the word lists to a language model dynamic optimization platform;
the comprehensive processing module is used for respectively utilizing a TTS voice synthesis method and a screening and removing method to obtain a forward set and a reverse set aiming at the vocabulary to be optimized, and utilizing the forward set and the reverse set to obtain the reference generation efficiency and the false alarm rate of the vocabulary to be optimized; carrying out default weight compiling, validation and testing on the vocabulary to be optimized, and calculating to obtain the validation rate and the false alarm rate of the vocabulary to be optimized after the testing;
the vocabulary compiling processing module is used for adjusting the vocabulary to be optimized by taking the false alarm rate as a reference to adjust the vocabulary entry weight, obtaining the vocabulary after the weight is adjusted and compiling and validating the vocabulary;
and the final weight acquisition module is used for circularly calling the comprehensive processing module and the word list compiling processing module, selecting a group of word list weights with the highest average effective rate of the word lists to be optimized within the allowable range of the false alarm rate as final weights to be output and issuing the final weights to the dynamic language model.
The working principle of the technical scheme is as follows: obtaining a vocabulary to be optimized by utilizing implementation entries in a language model, and importing the vocabulary to be optimized to a dynamic model implementation platform; then, a forward set and an adjusted reverse set aiming at the entry information to be optimized are obtained through a TTS speech synthesis method and an automatic screening and removing method, and the reference generation efficiency and the false alarm rate of the vocabulary to be optimized are obtained by utilizing the forward set and the adjusted reverse set; carrying out default weight compiling, validation and testing on the vocabulary to be optimized, and calculating to obtain the validation rate and the false alarm rate of the vocabulary to be optimized after the testing; then, the weight of the entry which is lower than the reference generation efficiency is adjusted by taking the false alarm rate as a reference, and the vocabulary compiling after the weight adjustment is obtained and is carried out for effectiveness; and finally, circularly calling the comprehensive processing module and the word list compiling processing module, selecting a group of word list weights with the highest average effective rate of the word lists to be optimized within the allowable range of the false alarm rate as final weight output, and issuing the final weight output to the dynamic language model.
The technical effects of the scheme are as follows: the problem that entries are not effective or are not common is solved by adding a weight adjustable function in the dynamic model. Under the condition of ensuring that the reverse set is not spurious, namely the false alarm rate is controlled to be in accordance with a set range, the weight of each entry is continuously improved, the entry generation efficiency is improved, and the recognition result is optimal.
In an embodiment of the present invention, the import module includes:
the to-be-optimized word list obtaining module is used for screening out word lists with poor recognition effects and sorting and summarizing the word lists;
and the vocabulary import module is used for importing the sorted and summarized vocabularies to the language model dynamic optimization platform and taking the vocabularies imported to the language model dynamic optimization platform as the vocabularies to be optimized.
The working principle of the scheme is as follows: obtaining a vocabulary to be optimized by using an implementation entry in the language model through a vocabulary to be optimized obtaining module; and then, importing the vocabulary to be optimized to a dynamic model implementation platform through an information import module. In the operation process of the to-be-optimized vocabulary obtaining module, the implementation vocabulary entries in the language model are screened through the screening and identifying module, the vocabulary entries which are not effective or are identified in disorder are identified, and then the vocabulary entries which are identified by the screening and identifying module and are not effective or are identified in disorder are sorted through the vocabulary entry sorting module to obtain the to-be-optimized vocabulary.
The technical effects of the scheme are as follows: the accuracy of obtaining the vocabulary to be optimized is improved, and omission of ineffective or disordered identification is effectively prevented.
In one embodiment of the present invention, the integrated processing module includes:
the forward set acquisition module is used for synthesizing a plurality of sentence patterns which are set in advance by the vocabulary to be optimized and the language model dynamic optimization platform by using a TTS (text to speech) speech synthesis method to generate a forward set;
the reverse set acquisition module is used for forming a reverse set by utilizing the fixed phrase tone test set;
the basic reference acquisition module is used for carrying out test calculation on the forward set and the adjusted reverse set to obtain the generation efficiency and the false alarm rate of entries in the vocabulary to be optimized when the optimization implementation is not carried out, and taking the generation efficiency and the false alarm rate when the optimization implementation is not carried out as the basic generation efficiency and the basic false alarm rate;
the weight compiling and validating module is used for compiling and validating each entry in the vocabulary to be optimized by default weight;
the generating efficiency obtaining module is used for testing the forward set and the adjusted reverse set and calculating to obtain the effective rate of each entry in the vocabulary to be optimized after the test;
and the false alarm rate acquisition module is used for testing the forward set and the adjusted reverse set and calculating to obtain the false alarm rate of each entry in the vocabulary to be optimized after the test.
Wherein the forward set acquisition module comprises:
the sentence pattern generation module is used for judging the category of the vocabulary to be optimized by utilizing the language dynamic model optimization platform and setting a plurality of sentence patterns according to the category of the vocabulary to be optimized;
the combination module is used for combining the sentence patterns and the word list to be optimized to obtain a combination unit;
and the forward set generating module is used for carrying out voice synthesis on the combination unit by using a TTS voice synthesis method and taking the synthesized voice as a forward set.
The reverse set acquisition module comprises:
the voice elimination module is used for eliminating the voice containing the vocabulary to be optimized in the short voice test set;
and the reverse set generation module is used for forming a reverse set by the phrase sound test set after the speech containing the vocabulary to be optimized is eliminated.
The working principle of the technical scheme is as follows: synthesizing the entry information to be optimized, which is led into the dynamic model real-time platform, into a plurality of forward sets by using a forward set acquisition module in a TTS (text to speech) speech synthesis mode; then, processing the reverse set in the dynamic model implementation platform by adopting a screening and removing method through the adjusted reverse set acquisition module to obtain an adjusted reverse set; then, a reference acquisition module is used for carrying out test calculation on the forward set and the adjusted reverse set to obtain reference generation efficiency and false alarm rate of entries in the vocabulary to be optimized when the entries are not optimized; then, a weight compiling and validating module is used for carrying out default weight compiling and validating on each entry in the vocabulary to be optimized; testing the forward set and the adjusted reverse set by using an effectiveness rate acquisition module, and calculating to obtain the effectiveness rate of each entry in the vocabulary to be optimized after the test; and finally, testing the forward set and the adjusted reverse set by using an error rate acquisition module, and calculating to obtain the error rate of each entry in the vocabulary to be optimized after the test. In the process of obtaining the forward set, judging the category of a vocabulary to be optimized through a language dynamic model optimization platform, setting a plurality of sentence patterns according to the category of the vocabulary to be optimized, combining the sentence patterns and the vocabulary to be optimized to obtain a combination unit, carrying out voice synthesis on the combination unit by using a TTS voice synthesis method, and taking the synthesized voice as the forward set; and the reverse set is used for verifying whether a spurious phenomenon occurs after the entry weight is implemented, the reverse set is a batch of fixed phrase sound test set, and the voice containing the entry to be implemented is removed according to the labeled text before the reverse set is run.
The technical effect of the technical scheme is as follows: the reference generating efficiency and the false alarm rate are measured and calculated through the forward set and the adjusted reverse set, the reference generating efficiency and the false alarm rate, the generating efficiency and the false alarm rate in the current test and the matching degree of the vocabulary to be optimized are improved, and data information with high accuracy and matching degree is provided for the subsequent weight adjusting process.
In an embodiment of the present invention, the vocabulary compiling processing module includes:
the weight adjusting module is used for adjusting the weight of the entry to be optimized and lower than the reference generation efficiency by taking the false alarm rate as a reference to obtain a word list after the weight is adjusted;
and the vocabulary compiling and validating module is used for compiling and validating the vocabulary after the weight adjustment.
Wherein the weight adjusting module comprises:
the weight increasing module is used for continuously increasing the weight of the entry with the effective rate lower than the reference generation efficiency in the vocabulary to be optimized within the allowable range of the false alarm rate by taking the false alarm rate as the reference, and the weight increasing range is [0,2 ];
the false alarm rate judging module is used for judging whether the false alarm rate of the entry with the increased weight exceeds an allowable range or not, and if the false alarm rate exceeds the allowable range, the weight reducing module is started;
the weight reduction module is used for continuously reducing the weight of the entries on the basis of the historical weight formed after the weights of the entries are increased until the false alarm rate of the entries is recovered to an allowable range;
and the vocabulary compiling and acquiring module is used for acquiring vocabulary compiling corresponding to the entry after the weight is adjusted.
The working principle of the technical scheme is as follows: the vocabulary compiling processing module adjusts the weight of the vocabulary entry with the false alarm rate as the reference through the weight adjusting module, and obtains the vocabulary compiling after the weight adjustment; and then, the vocabulary compiling validation module validates the vocabulary compiling obtained by the weight adjusting module. In the operation process of the weight adjusting module, the weight increasing module takes the false alarm rate as a reference, and the weight of the entry with the generation efficiency lower than the reference generation value is continuously increased within the allowable range of the false alarm rate; then, a false alarm rate judging module is used for judging whether the false alarm rate of the entry exceeds an allowable range, and if the false alarm rate exceeds the allowable range, a weight reducing module is started; and finally, continuously reducing the weight of the entries by a weight reduction module on the basis of the historical weight formed after the weights of the entries are increased until the false alarm rate of the entries is restored to an allowable range, and finally obtaining vocabulary compilation corresponding to the entries after the weights are adjusted.
The technical effects of the scheme are as follows: the problem that entries are not effective or are not common is solved by adding a weight adjustable function in the dynamic model. Under the condition of ensuring that the reverse set is not spurious, namely the false alarm rate is controlled to be in accordance with a set range, the weight of each entry is continuously improved, the entry generation efficiency is improved, and the recognition result is optimal.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.
Claims (9)
1. A system for enhancing dynamic model identification accuracy by dynamically lexical weighting, the system comprising:
the import module is used for sorting and summarizing word lists with poor recognition effects and submitting the word lists to a language model dynamic optimization platform;
the comprehensive processing module is used for respectively utilizing a TTS voice synthesis method and a screening and removing method to obtain a forward set and a reverse set aiming at the vocabulary to be optimized, and utilizing the forward set and the reverse set to obtain the reference generation efficiency and the false alarm rate of the vocabulary to be optimized; carrying out default weight compiling, validation and testing on the vocabulary to be optimized, and calculating to obtain the validation rate and the false alarm rate of the vocabulary to be optimized after the testing;
the vocabulary compiling processing module is used for adjusting the vocabulary to be optimized by taking the false alarm rate as a reference to adjust the vocabulary entry weight, obtaining the vocabulary after the weight is adjusted and compiling and validating the vocabulary;
and the final weight acquisition module is used for circularly calling the comprehensive processing module and the word list compiling processing module, selecting a group of word list weights with the highest average effective rate of the word lists to be optimized within the allowable range of the false alarm rate as final weights to be output and issuing the final weights to the dynamic language model.
2. The system of claim 1, wherein the import module comprises:
the to-be-optimized word list obtaining module is used for screening out word lists with poor recognition effects and sorting and summarizing the word lists;
and the vocabulary import module is used for importing the sorted and summarized vocabularies to the language model dynamic optimization platform and taking the vocabularies imported to the language model dynamic optimization platform as the vocabularies to be optimized.
3. The system of claim 1, wherein the integrated processing module comprises:
the forward set acquisition module is used for synthesizing a plurality of sentence patterns which are set in advance by the vocabulary to be optimized and the language model dynamic optimization platform by using a TTS (text to speech) speech synthesis method to generate a forward set;
the reverse set acquisition module is used for forming a reverse set by utilizing the fixed phrase tone test set;
the basic reference acquisition module is used for carrying out test calculation on the forward set and the adjusted reverse set to obtain the generation efficiency and the false alarm rate of entries in the vocabulary to be optimized when the optimization implementation is not carried out, and taking the generation efficiency and the false alarm rate when the optimization implementation is not carried out as the basic generation efficiency and the basic false alarm rate;
the weight compiling and validating module is used for compiling and validating each entry in the vocabulary to be optimized by default weight;
the generating efficiency obtaining module is used for testing the forward set and the adjusted reverse set and calculating to obtain the effective rate of each entry in the vocabulary to be optimized after the test;
and the false alarm rate acquisition module is used for testing the forward set and the adjusted reverse set and calculating to obtain the false alarm rate of each entry in the vocabulary to be optimized after the test.
4. The system of claim 3, wherein the forward set acquisition module comprises:
the sentence pattern generation module is used for judging the category of the vocabulary to be optimized by utilizing the language dynamic model optimization platform and setting a plurality of sentence patterns according to the category of the vocabulary to be optimized;
the combination module is used for combining the sentence patterns and the word list to be optimized to obtain a combination unit;
and the forward set generating module is used for carrying out voice synthesis on the combination unit by using a TTS voice synthesis method and taking the synthesized voice as a forward set.
5. The system of claim 3, wherein the reverse set acquisition module comprises:
the voice elimination module is used for eliminating the voice containing the vocabulary to be optimized in the short voice test set;
and the reverse set generation module is used for forming a reverse set by the phrase sound test set after the speech containing the vocabulary to be optimized is eliminated.
6. The system of claim 1, wherein the vocabulary compiling processing module comprises:
the weight adjusting module is used for adjusting the weight of the entry to be optimized and lower than the reference generation efficiency by taking the false alarm rate as a reference to obtain a word list after the weight is adjusted;
and the vocabulary compiling and validating module is used for compiling and validating the vocabulary after the weight adjustment.
7. The system of claim 6, wherein the weight adjustment module comprises:
the weight increasing module is used for continuously increasing the weight of the entry with the effective rate lower than the reference generation efficiency in the vocabulary to be optimized within the allowable range of the false alarm rate by taking the false alarm rate as the reference, and the weight increasing range is [0,2 ];
the false alarm rate judging module is used for judging whether the false alarm rate of the entry with the increased weight exceeds an allowable range or not, and if the false alarm rate exceeds the allowable range, the weight reducing module is started;
the weight reduction module is used for continuously reducing the weight of the entries on the basis of the historical weight formed after the weights of the entries are increased until the false alarm rate of the entries is recovered to an allowable range;
and the vocabulary compiling and acquiring module is used for acquiring vocabulary compiling corresponding to the entry after the weight is adjusted.
8. A method for improving the accuracy of dynamic model identification by dynamically adjusting weights by vocabulary, which is characterized by comprising the following steps:
sorting and summarizing word lists with poor recognition effects, and submitting the word lists to a language model dynamic optimization platform;
respectively utilizing a TTS speech synthesis method and a screening and removing method to obtain a forward set and a reverse set aiming at a vocabulary to be optimized, and utilizing the forward set and the reverse set to obtain the reference generation efficiency and the false alarm rate of the vocabulary to be optimized; carrying out default weight compiling, validation and testing on the vocabulary to be optimized, and calculating to obtain the validation rate and the false alarm rate of the vocabulary to be optimized after the testing;
adjusting a vocabulary to be optimized by taking the false alarm rate as a reference to adjust the vocabulary entry weight, obtaining the vocabulary after adjusting the weight, and compiling and validating the vocabulary;
and circularly calling the comprehensive processing module and the word list compiling processing module, selecting a group of word list weights with the highest average effective rate of the word lists to be optimized within the allowable range of the false alarm rate as final weight output, and issuing the final weight output to the dynamic language model.
9. The method according to claim 1, wherein the specific process of the method comprises:
step 1, sorting and summarizing word lists with poor recognition effects, and submitting the word lists to a language model dynamic optimization platform to serve as word lists to be optimized;
step 2, importing the vocabulary to be optimized into a dynamic model implementation platform;
step 3, setting a plurality of sentence patterns in advance by using a language model dynamic optimization platform according to the category of the vocabulary to be optimized, combining the vocabulary to be optimized and the sentence patterns, and synthesizing voice as a forward set through TTS;
step 4, eliminating the voice containing the vocabulary to be optimized in the short voice test set, and forming a reverse set by the phrase voice test set after the voice containing the vocabulary to be optimized is eliminated;
step 5, testing and calculating the forward set and the adjusted reverse set to obtain the generation efficiency and the false alarm rate of the entries in the vocabulary to be optimized when the entries are not optimized, and taking the generation efficiency and the false alarm rate as the reference generation efficiency and the false alarm rate;
step 6, compiling and validating each entry in the vocabulary to be optimized by default weight, wherein the default weight is 0;
step 7, testing the forward set and the adjusted reverse set, and calculating to obtain the effective rate and the false alarm rate of each entry in the vocabulary to be optimized after the test;
step 8, giving priority to false alarm rate, and continuously increasing the weight of the entry with poor generation efficiency within the allowable range of the false alarm rate, wherein the weight range is [0,2 ]; if the false alarm rate exceeds the allowable range, continuously reducing the entry weight on the basis of the entry historical weight until the false alarm rate is restored to the allowable range, and finally obtaining a vocabulary after the weight is adjusted;
step 9, compiling and validating the vocabulary after the weight is adjusted, which is obtained in the step 8;
and step 10, circularly calling the step 7, the step 8 and the step 9 for a plurality of times, selecting a group of vocabulary weights with the highest average effective rate of the to-be-optimized vocabulary as final weights within the allowable range of the false alarm rate, and outputting the final weights to the dynamic language model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010326352.1A CN111597798B (en) | 2020-04-23 | 2020-04-23 | Method and system for improving dynamic model identification accuracy |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010326352.1A CN111597798B (en) | 2020-04-23 | 2020-04-23 | Method and system for improving dynamic model identification accuracy |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111597798A true CN111597798A (en) | 2020-08-28 |
| CN111597798B CN111597798B (en) | 2023-05-26 |
Family
ID=72187647
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010326352.1A Active CN111597798B (en) | 2020-04-23 | 2020-04-23 | Method and system for improving dynamic model identification accuracy |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111597798B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20020111806A1 (en) * | 2001-02-13 | 2002-08-15 | International Business Machines Corporation | Dynamic language model mixtures with history-based buckets |
| US20170004823A1 (en) * | 2015-06-30 | 2017-01-05 | Ibm | Testing words in a pronunciation lexicon |
| CN109523991A (en) * | 2017-09-15 | 2019-03-26 | 阿里巴巴集团控股有限公司 | Method and device, the equipment of speech recognition |
-
2020
- 2020-04-23 CN CN202010326352.1A patent/CN111597798B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20020111806A1 (en) * | 2001-02-13 | 2002-08-15 | International Business Machines Corporation | Dynamic language model mixtures with history-based buckets |
| US20170004823A1 (en) * | 2015-06-30 | 2017-01-05 | Ibm | Testing words in a pronunciation lexicon |
| CN109523991A (en) * | 2017-09-15 | 2019-03-26 | 阿里巴巴集团控股有限公司 | Method and device, the equipment of speech recognition |
Non-Patent Citations (1)
| Title |
|---|
| 肖镜辉;刘秉权;王晓龙;: "面向汉语建模的自适应词表生成算法" * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111597798B (en) | 2023-05-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN1248192C (en) | Semi-monitoring speaker self-adaption | |
| US7756710B2 (en) | Method and apparatus for error correction in speech recognition applications | |
| CN1196104C (en) | Speech processing | |
| US8494853B1 (en) | Methods and systems for providing speech recognition systems based on speech recordings logs | |
| EP1679694A1 (en) | Improving error prediction in spoken dialog systems | |
| CN111916108B (en) | Voice evaluation method and device | |
| CN102385859A (en) | Method and system for synthesizing parameter voice | |
| CN101727904A (en) | Voice translation method and device | |
| WO2009040382A1 (en) | Method and system for identifying information related to a good | |
| CN102074234A (en) | Voice variation model building device and method as well as voice recognition system and method | |
| CN1819017A (en) | Method for extracting feature vectors for speech recognition | |
| CN112489655B (en) | Method, system and storage medium for correcting voice recognition text error in specific field | |
| CN110083711A (en) | A kind of phonetic transcriptions of Chinese characters conversion method and converting system | |
| WO2023114064A1 (en) | Adaptation and training of neural speech synthesis | |
| CN116665676A (en) | Semantic recognition method for intelligent voice outbound system | |
| CN111597798A (en) | Method and system for improving identification accuracy of dynamic model | |
| CN1153127C (en) | Intelligent common spoken Chinese phonetic input method and dictation machine | |
| CN1342969A (en) | Method for recogniting voice | |
| US10402492B1 (en) | Processing natural language grammar | |
| CN113744718A (en) | Voice text output method and device, storage medium and electronic device | |
| CN1825430A (en) | Speech synthetic method and apparatus capable of regulating rhythm and session system | |
| CN116579317B (en) | Method and system for automatically generating publications based on AI content | |
| CN111428860B (en) | Method and device for reducing power consumption of time delay neural network model | |
| CN113870875A (en) | Tone feature extraction method, device, computer equipment and storage medium | |
| CN102117615B (en) | Device, method and system for generating utterance verification critical value |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |