CN110415705A - A kind of hot word recognition methods, system, device and storage medium - Google Patents
A kind of hot word recognition methods, system, device and storage medium Download PDFInfo
- Publication number
- CN110415705A CN110415705A CN201910706314.6A CN201910706314A CN110415705A CN 110415705 A CN110415705 A CN 110415705A CN 201910706314 A CN201910706314 A CN 201910706314A CN 110415705 A CN110415705 A CN 110415705A
- Authority
- CN
- China
- Prior art keywords
- word
- hot word
- hot
- recognition result
- current
- 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
- 238000000034 method Methods 0.000 title claims abstract description 38
- 239000000945 filler Substances 0.000 claims description 8
- 238000010276 construction Methods 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 4
- 239000000284 extract Substances 0.000 claims description 3
- 239000012634 fragment Substances 0.000 claims description 3
- 230000005055 memory storage Effects 0.000 claims description 3
- 230000001360 synchronised effect Effects 0.000 claims description 3
- 235000013399 edible fruits Nutrition 0.000 claims 1
- 238000001514 detection method Methods 0.000 description 6
- 238000012545 processing Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 238000004891 communication Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 238000012805 post-processing Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/063—Training
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/26—Speech to text systems
Landscapes
- Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Machine Translation (AREA)
Abstract
The present invention provides a kind of hot word recognition methods, system, device and storage mediums, to solve the problems, such as in the prior art can accidentally modify correct speech recognition result, hot word recognition methods is the following steps are included: step 1: audio user is sent into universal identification engine, speech recognition result is obtained, while obtaining sound recognition result WiCorresponding position and confidence level in audio;Step 2: audio user being sent into hot word detecting and alarm and carries out hot word retrieval, hot word W, the hot word corresponding audio position P and score S of highest scoring is obtained, is expressed as (W, P, S);Step 3: judging the score S of the hot word (W, P, S) of highest scoring, if score S is greater than given threshold value, replace speech recognition result W with hot word Wi~WjThe word of middle corresponding audio position executes step 4;Otherwise, terminate;Step 4: if the word in hot word appearance position and current recognition result has overlapping, the word before and after hot word being modified.
Description
Technical field
The present invention relates to technical field of voice recognition, and in particular to a kind of hot word recognition methods, system, device and storage are situated between
Matter.
Background technique
Speech recognition technology has become the major technique of current manual's intelligent use.Typical speech recognition technology be all according to
Rely specific vocabulary, that is to say, that can only identify the word within the scope of given vocabulary;If occurring the outer word of vocabulary in voice, usually
Recognition performance can be very poor, or even basic identification does not come out.In response to this problem, there has been proposed some solutions.Main method
Referred to as recognition result post-processing technology is analyzed by the text to recognition result, then uses language model or given
Hot word pronounce to correct recognition result.Such methods have a fatal defects, that is, often correct recognition result is missed
Modification.
Summary of the invention
In view of the above-mentioned problems, the present invention provides a kind of hot word recognition methods, system, device and storage medium, to solve
The problem of correct speech recognition result accidentally being modified in the prior art.
Its technical solution is such that a kind of hot word recognition methods, which comprises the following steps:
Step 1: audio user being sent into universal identification engine, obtains speech recognition result, sound recognition result is expressed as W1,
W2,...,Wn, wherein n is natural number, while obtaining sound recognition result WiCorresponding position and confidence level in audio, wherein 1
≤ i≤n, i are natural number;
Step 2: audio user being sent into hot word detecting and alarm and carries out hot word retrieval, obtains hot word W, the hot word of highest scoring
Corresponding audio position P and score S, is expressed as (W, P, S);
Step 3: judging the score S of the hot word (W, P, S) of highest scoring, if score S is greater than given threshold value, use hot word W
Replace speech recognition result WI~WjThe word of middle corresponding audio position executes step 4;Otherwise, terminate;
Step 4: if the word in hot word appearance position and current recognition result has overlapping, the word before and after hot word being repaired
Just.
It further, further include step 1.5 between step 1 and step 2: speech recognition result W if it existsI~Wj, i < j,
J is natural number, WI~WjAverage confidence be lower than given threshold value, then extract WI~WjCorresponding audio fragment executes step 2.
Further, step 1 is synchronous with step 2 executes.
Further, step 2 specifically comprises the following steps:
Step 2-1: according to hot word vocabulary, a filler word is added, filler word is configured to connect all acoustics and builds
Form unit, the parallel grammer of construction one identify network;
Step 2-2: the input sound bite of extraction is carried out using the viterbi algorithm of beam-search beam-search
Decoding search;
Step 2-3: backtracking obtains the hot word and the corresponding audio position of hot word of highest scoring;
Step 2-4: it by calculating average posterior probability to the corresponding speech frame of hot word, is exported as the hot word score.
Further, in step 2, it is obtained in grammer identification network using the posterior probability of universal identification acoustic model output
Point.
Further, in step 4, it includes initial position that the word in hot word appearance position and current recognition result, which has overlapping,
It is overlapping and end position overlapping.
Further, when the word in hot word appearance position and current recognition result have initial position it is overlapping when, step 4 tool
Body includes the following steps:
Step 4-1: determining the word in recognition result where hot word original position, calculates the word initial position and hot word
The alternate position spike of initial position;
Step 4-2: if alternate position spike is greater than the duration of a word, the word and hot word not overlapping part are selected from vocabulary
In, the similar word of word pronunciation is as candidate word;
Step 4-3: using preparatory trained language model, word and current before predicting given beginning of the sentence word to current word
After word under the conditions of word, the probability that each candidate word of current word occurs, and the score as the candidate word;
Step 4-4: it if the score of the candidate word of highest scoring is greater than given threshold value, is replaced with the candidate word current
Word;Otherwise keep current word constant;
When the word in hot word appearance position and current recognition result have end position it is overlapping when, step 4 specifically include as
Lower step:
Step 4.1: determine hot word end position place in recognition result word, calculate the end position and hot word of the word
End position alternate position spike;
Step 4.2: if alternate position spike is greater than the duration of a word, the word and hot word not overlapping part are selected from vocabulary
In, the similar word of word pronunciation is as candidate word;
Step 4.3: using preparatory trained language model, predicting after given sentence tail word to current word word and current
Before word under the conditions of word, the probability that each candidate word of current word occurs, and the score as the candidate word;
Step 4.4: if the score of the candidate word of highest scoring is greater than given threshold value, being replaced with the candidate word current
Word;Otherwise keep current word constant.
Further, between step 4-3 and step 4-5, and between step 4.3 and step 4.5, further include respectively
Following steps: increasing the acoustic confidence information of each word, then predicts the probability that each candidate word of current word occurs, and make
For the score of the candidate word.
Further, hot word detecting and alarm is configured to corresponding with User ID, uploads hot word and use when adding hot word simultaneously
Family ID inquires pronunciation dictionary, obtains the pronunciation of hot word and its to sound aligned phoneme sequence;Then grammer network is added in hot word;It generates
Hot word detects resource, and hot word correspondence is added to hot word detecting and alarm corresponding with User ID.
A kind of hot word identifying system characterized by comprising
Universal phonetic identifies engine, is configured to export the time location of speech recognition result and each word in audio
And confidence level;
Hot word detecting and alarm is configured to detect whether to export ID, audio position and its score of hot word there are hot word;
Hot word modified result module is configured to know using the voice of hot word replacement universal phonetic identification engine output
The word of corresponding position in other result;
Language model modified result module is configured to have friendship as the word in hot word appearance position and current recognition result
Word before and after hot word is modified when folded.
Further, further include hot word adding module, be configured to add hot word to the hot word detecting and alarm.
A kind of hot word identification device, characterized in that it comprises: including processor, memory and program;
Described program stores in the memory, and the processor calls the program of memory storage, above-mentioned to execute
Hot word recognition methods.
A kind of computer readable storage medium, which is characterized in that the computer readable storage medium is configured to store
Program, described program are configured to execute above-mentioned hot word recognition methods.
Hot word recognition methods of the invention identifies hot word using word detection scheme is waken up, hot word can user by making by oneself
Justice, identify hot word after hot word result is modified, in addition, also detect hot word and it is modified on the basis of, can be further
Other identification mistakes caused by modifying factor hot word identification mistake, amendment and hot word have those of overlapping or neighbouring low confidence
Word, hot word recognition methods of the invention, significantly reduces and accidentally corrects mistake, and user can be with the hot word of fast custom oneself, hot word
Detection accuracy is high, ensure that modified high-accuracy.
Detailed description of the invention
Fig. 1 is the flow chart of the hot word recognition methods of specific embodiment 1;
Fig. 2 is the system block diagram of the hot word identifying system of specific embodiment 1;
Fig. 3 is the flow chart of the hot word recognition methods of specific embodiment 2;
Fig. 4 is the system block diagram of the hot word identifying system of specific embodiment 2.
Specific embodiment
Specific embodiment 1: see Fig. 1, a kind of hot word recognition methods, comprising the following steps:
Step 1: audio user being sent into universal identification engine, obtains speech recognition result, sound recognition result is expressed as W1,
W2,...,Wn, wherein n is natural number, while obtaining sound recognition result WiCorresponding position and confidence level in audio, wherein 1
≤ i≤n, i are natural number;
Step 1.5: speech recognition result W if it existsI~Wj, i < j, j are natural number, WI~WjAverage confidence lower than giving
Determine threshold value, threshold value takes 0.5, then extracts WI~WjCorresponding audio fragment executes step 2;Otherwise, terminate;
Step 2: audio user being sent into hot word detecting and alarm and carries out hot word retrieval, obtains hot word W, the hot word of highest scoring
Corresponding audio position P and score S, is expressed as (W, P, S);
Step 3: judge the score S of the hot word (W, P, S) of highest scoring, if score S is greater than given threshold value, threshold value takes 0.5,
Then speech recognition result W is replaced with hot word WI~WjThe word of middle corresponding audio position executes step 4;Otherwise, terminate;
Step 4: if the word in hot word appearance position and current recognition result has overlapping, the word before and after hot word being repaired
Just.
Specifically, step 2 specifically comprises the following steps:
Step 2-1: according to hot word vocabulary, a filler word is added, filler word is configured to connect all acoustics and builds
Form unit, the parallel grammer of construction one identify network;
Step 2-2: the input sound bite of extraction is carried out using the viterbi algorithm of beam-search beam-search
Decoding search;
Step 2-3: backtracking obtains the hot word and the corresponding audio position of hot word of highest scoring;
Step 2-4: it by calculating average posterior probability to the corresponding speech frame of hot word, is exported as the hot word score.
In the present embodiment, in step 2, general using the posteriority of universal identification acoustic model output in grammer identification network
Rate score.
Specifically, in step 4, it includes initial position that the word in hot word appearance position and current recognition result, which has overlapping,
It is overlapping and end position overlapping.
When the word in hot word appearance position and current recognition result have initial position it is overlapping when, step 4 specifically include as
Lower step:
Step 4-1: determining the word in recognition result where hot word original position, calculates the word initial position and hot word
The alternate position spike of initial position;
Step 4-2: if alternate position spike is greater than the duration of a word, the word and hot word not overlapping part are selected from vocabulary
In, the similar word of word pronunciation is as candidate word;
Step 4-3: using preparatory trained language model, word and current before predicting given beginning of the sentence word to current word
After word under the conditions of word, the probability that each candidate word of current word occurs, and the score as the candidate word;
Increase the acoustic confidence information of each word, then predicts the probability that each candidate word of current word occurs, and make
For the score of the candidate word;
Step 4-4: if the score of the candidate word of highest scoring is greater than given threshold value, threshold value takes 0.5, then with the candidate word
Replace current word;Otherwise keep current word constant;
When the word in hot word appearance position and current recognition result have end position it is overlapping when, step 4 specifically include as
Lower step:
Step 4.1: determine hot word end position place in recognition result word, calculate the end position and hot word of the word
End position alternate position spike;
Step 4.2: if alternate position spike is greater than the duration of a word, the word and hot word not overlapping part are selected from vocabulary
In, the similar word of word pronunciation is as candidate word;
Step 4.3: using preparatory trained language model, predicting after given sentence tail word to current word word and current
Before word under the conditions of word, the probability that each candidate word of current word occurs, and the score as the candidate word;
Increase the acoustic confidence information of each word, then predicts the probability that each candidate word of current word occurs, and make
For the score of the candidate word
Step 4.4: if the score of the candidate word of highest scoring is greater than given threshold value, threshold value takes 0.5, then with the candidate word
Replace current word;Otherwise keep current word constant.
Specifically, in the present embodiment, the language model used in step 4 is recurrent neural network, specially LSTM/
GRU class language model.
In the present embodiment, hot word detecting and alarm is configured to corresponding with User ID to realize that user relies on, and adds hot word
When upload hot word and User ID simultaneously, inquire pronunciation dictionary, obtain the pronunciation of hot word and its to sound aligned phoneme sequence;Then hot word
Grammer network is added;It generating hot word and detects resource, hot word correspondence is added in hot word detecting and alarm corresponding with User ID, so that
Hot word can be conveniently added in hot word detecting and alarm.
Specifically, user can transmit a triple to system, system is told to add or delete given hot word.Three
Tuple is defined as follows: (ID, HotWord, OPT), wherein ID: mark user;HotWord: mark hot word;OPT: mark is dynamic
Make, OPT is defined as adding or deleting.
Hot word identifying system corresponding with specific embodiment 1 is shown in Fig. 2, comprising:
Universal phonetic identifies engine 1, is configured to export the when meta position of speech recognition result and each word in audio
It sets and confidence level;
Hot word detecting and alarm 2 is configured to detect whether to export ID, audio position and its score of hot word there are hot word;
Hot word modified result module 3 is configured to the voice using hot word replacement universal phonetic identification engine output
The word of corresponding position in recognition result;
Language model modified result module 4 is configured to have as the word in hot word appearance position and current recognition result
Word before and after hot word is modified when overlapping.
Further include hot word adding module 5, is configured to add hot word to hot word detecting and alarm.
In specific embodiment 1, knows otherwise as shown in Fig. 2, providing a kind of sequence, first do universal identification, according to
The confidence level of recognition result, then hot word detection is done, additional computing resource is not needed, system delay will increase.
Specific embodiment 2: see Fig. 2, a kind of hot word recognition methods, comprising the following steps:
Step 1: audio user being sent into universal identification engine, obtains speech recognition result, sound recognition result is expressed as W1,
W2,...,Wn, wherein n is natural number, while obtaining sound recognition result WiCorresponding position and confidence level in audio, wherein 1
≤ i≤n, i are natural number;
Step 2: audio user being sent into hot word detecting and alarm and carries out hot word retrieval, obtains hot word W, the hot word of highest scoring
Corresponding audio position P and score S, is expressed as (W, P, S);
Step 3: judge the score S of the hot word (W, P, S) of highest scoring, if score S is greater than given threshold value, threshold value takes 0.5,
Then speech recognition result W is replaced with hot word WI~WjThe word of middle corresponding audio position executes step 4;Otherwise, terminate;
Step 4: if the word in hot word appearance position and current recognition result has overlapping, the word before and after hot word being repaired
Just.
In the present embodiment, step 1 is synchronous with step 2 executes.
Specifically, step 2 specifically comprises the following steps:
Step 2-1: according to hot word vocabulary, a filler word is added, filler word is configured to connect all acoustics and builds
Form unit, the parallel grammer of construction one identify network;
Step 2-2: the input sound bite of extraction is carried out using the viterbi algorithm of beam-search beam-search
Decoding search;
Step 2-3: backtracking obtains the hot word and the corresponding audio position of hot word of highest scoring;
Step 2-4: it by calculating average posterior probability to the corresponding speech frame of hot word, is exported as the hot word score.
In the present embodiment, in step 2, the acoustic model in grammer identification network is CLDNN model.
Specifically, in step 4, it includes initial position that the word in hot word appearance position and current recognition result, which has overlapping,
It is overlapping and end position overlapping.
When the word in hot word appearance position and current recognition result have initial position it is overlapping when, step 4 specifically include as
Lower step:
Step 4-1: determining the word in recognition result where hot word original position, calculates the word initial position and hot word
The alternate position spike of initial position;
Step 4-2: if alternate position spike is greater than the duration of a word, the word and hot word not overlapping part are selected from vocabulary
In, the similar word of word pronunciation is as candidate word;
Step 4-3: using preparatory trained language model, word and current before predicting given beginning of the sentence word to current word
After word under the conditions of word, the probability that each candidate word of current word occurs, and the score as the candidate word;
Increase the acoustic confidence information of each word, then predicts the probability that each candidate word of current word occurs, and make
For the score of the candidate word;
Step 4-4: if the score of the candidate word of highest scoring is greater than given threshold value, threshold value takes 0.5, then with the candidate word
Replace current word;Otherwise keep current word constant;
When the word in hot word appearance position and current recognition result have end position it is overlapping when, step 4 specifically include as
Lower step:
Step 4.1: determine hot word end position place in recognition result word, calculate the end position and hot word of the word
End position alternate position spike;
Step 4.2: if alternate position spike is greater than the duration of a word, the word and hot word not overlapping part are selected from vocabulary
In, the similar word of word pronunciation is as candidate word;
Step 4.3: using preparatory trained language model, predicting after given sentence tail word to current word word and current
Before word under the conditions of word, the probability that each candidate word of current word occurs, and the score as the candidate word;
Increase the acoustic confidence information of each word, then predicts the probability that each candidate word of current word occurs, and make
For the score of the candidate word
Step 4.4: if the score of the candidate word of highest scoring is greater than given threshold value, threshold value takes 0.5, then with the candidate word
Replace current word;Otherwise keep current word constant.
Specifically, in the present embodiment, the language model used in step 4 is recurrent neural network, specially LSTM/
GRU class language model.
In the present embodiment, hot word detecting and alarm is configured to corresponding with User ID to realize that user relies on, and adds hot word
When upload hot word and User ID simultaneously, inquire pronunciation dictionary, obtain the pronunciation of hot word and its to sound aligned phoneme sequence;Then hot word
Grammer network is added;It generating hot word and detects resource, hot word correspondence is added in hot word detecting and alarm corresponding with User ID, so that
Hot word can be conveniently added in hot word detecting and alarm.
Specifically, user can transmit a triple to system, system is told to add or delete given hot word.Three
Tuple is defined as follows: (ID, HotWord, OPT), wherein ID: mark user;HotWord: mark hot word;OPT: mark is dynamic
Make, OPT is defined as adding or deleting.
Hot word identifying system corresponding with specific embodiment 2 is shown in Fig. 4, comprising:
Universal phonetic identifies engine 1, is configured to export the when meta position of speech recognition result and each word in audio
It sets and confidence level;
Hot word detecting and alarm 2 is configured to detect whether to export ID, audio position and its score of hot word there are hot word;
Hot word modified result module 3 is configured to the voice using hot word replacement universal phonetic identification engine output
The word of corresponding position in recognition result;
Language model modified result module 4 is configured to have as the word in hot word appearance position and current recognition result
Word before and after hot word is modified when overlapping.
Further include hot word adding module 5, is configured to add hot word to hot word detecting and alarm.
In specific embodiment 2, as shown in figure 4, providing the mode of parallelism recognition, universal identification and hot word detection are simultaneously
It does, it is desirable that have more rich computing resource, system delay is basically unchanged, and response speed is faster.
Hot word recognition methods of the invention identifies hot word using word detection scheme is waken up, hot word can user by making by oneself
Justice, identify hot word after hot word result is modified, in addition, also detect hot word and it is modified on the basis of, can be further
Other identification mistakes, hot word recognition methods of the invention caused by the identification of modifying factor hot word is wrong significantly reduce and accidentally correct mistake
Accidentally, user can be with the hot word of fast custom oneself, and hot word Detection accuracy is high, ensure that modified high-accuracy.
In an embodiment of the present invention, a kind of hot word identification device is additionally provided comprising: including processor, memory
And program;Program stores in memory, and processor calls the program of memory storage, to execute above-mentioned hot word identification side
Method.
In the realization of above-mentioned hot word identification device, directly or indirectly it is electrically connected between memory and processor, with
Realize the transmission or interaction of data.For example, these elements can pass through one or more of communication bus or signal between each other
Line, which is realized, to be electrically connected, and can such as be connected by bus.The computer for realizing data access control method is stored in memory
It executes instruction, including the software function module that at least one can be stored in memory in the form of software or firmware, processor
By running the software program and module that are stored in memory, thereby executing various function application and data processing.
Memory may be, but not limited to, random access memory (Random Access Memory, referred to as: RAM),
Read-only memory (Read Only Memory, referred to as: ROM), programmable read only memory (Programmable Read-Only
Memory, referred to as: PROM), erasable read-only memory (Erasable Programmable Read-Only Memory, letter
Claim: EPROM), electricallyerasable ROM (EEROM) (Electric Erasable Programmable Read-Only Memory,
Referred to as: EEPROM) etc..Wherein, memory is for storing program, and processor executes program after receiving and executing instruction.
Processor can be a kind of IC chip, the processing capacity with signal.Above-mentioned processor can be logical
With processor, including central processing unit (Central Processing Unit, referred to as: CPU), network processing unit (Network
Processor, referred to as: NP) etc..It may be implemented or execute disclosed each method, step and the logic in the embodiment of the present application
Block diagram.General processor can be microprocessor or the processor is also possible to any conventional processor etc..
In an embodiment of the present invention, a kind of computer readable storage medium, computer readable storage medium are additionally provided
It is configured to store program, program is configured to execute above-mentioned hot word recognition methods.
Those of ordinary skill in the art will appreciate that: realize that all or part of the steps of above-mentioned each method embodiment can lead to
The relevant hardware of program instruction is crossed to complete.Program above-mentioned can store in computer readable storage medium.The program exists
When being executed by processor, realization includes the steps that above-mentioned each method embodiment;And computer readable storage medium above-mentioned includes:
The various media that can store program code such as ROM, RAM, magnetic or disk, including some instructions are used so that one big number
Each embodiment or embodiment are executed according to transmission device (can be personal computer, server or the network equipment etc.)
Method described in certain parts.
Claims (12)
1. a kind of hot word recognition methods, which comprises the following steps:
Step 1: audio user being sent into universal identification engine, obtains speech recognition result, sound recognition result is expressed as W1,
W2,...,Wn, wherein n is natural number, while obtaining sound recognition result WiCorresponding position and confidence level in audio, wherein 1
≤ i≤n, i are natural number;
Step 2: audio user being sent into hot word detecting and alarm and carries out hot word retrieval, hot word W, the hot word for obtaining highest scoring are corresponding
Audio position P and score S, be expressed as (W, P, S);
Step 3: judging the score S of the hot word (W, P, S) of highest scoring, if score S is greater than given threshold value, replaced with hot word W
The word of audio position is corresponded in speech recognition result, executes step 4;Otherwise, terminate;
Step 4: if the word in hot word appearance position and current recognition result has overlapping, the word before and after hot word being modified.
2. a kind of hot word recognition methods according to claim 1, it is characterised in that: between step 1 and step 2, also wrap
Include step 1.5: speech recognition result W if it existsi~Wj, i < j, i, j is natural number, Wi~WjAverage confidence lower than given
Threshold value then extracts Wi~WjCorresponding audio fragment executes step 2.
3. a kind of hot word recognition methods according to claim 1, it is characterised in that: step 1 is synchronous with step 2 to be executed.
4. a kind of hot word recognition methods according to claim 1, it is characterised in that: step 2 specifically comprises the following steps:
Step 2-1: according to hot word vocabulary, a filler word is added, filler word is configured to connect all Acoustic Modeling lists
Member, the parallel grammer of construction one identify network;
Step 2-2: the input sound bite of extraction is decoded using the viterbi algorithm of beam-search beam-search
Search;
Step 2-3: backtracking obtains the hot word and the corresponding audio position of hot word of highest scoring;
Step 2-4: it by calculating average posterior probability to the corresponding speech frame of hot word, is exported as the hot word score.
5. a kind of hot word recognition methods according to claim 4, it is characterised in that: in step 2, grammer identifies in network
The posterior probability score exported using universal identification acoustic model.
6. a kind of hot word recognition methods according to claim 1, it is characterised in that: in step 4, hot word appearance position and
Word in current recognition result has overlapping overlapping and end position overlapping including initial position;
When the word in hot word appearance position and current recognition result have initial position it is overlapping when, step 4 specifically includes following step
It is rapid:
Step 4-1: determining the word in recognition result where hot word original position, calculates the starting of the word initial position and hot word
The alternate position spike of position;
Step 4-2: if alternate position spike is greater than the duration of a word, from selecting the word and hot word in vocabulary not in overlapping part, word
Language pronounces similar word as candidate word;
Step 4-3: using preparatory trained language model, word after word and current word is given before beginning of the sentence word to current word
Under the conditions of, predict the probability that each candidate word of current word occurs, and the score as the candidate word;
Step 4-4: if the score of the candidate word of highest scoring is greater than given threshold value, current word is replaced with the candidate word;It is no
Then keep current word constant;
When the word in hot word appearance position and current recognition result have end position it is overlapping when, step 4 specifically includes following step
It is rapid:
Step 4.1: determine hot word end position place in recognition result word, calculate the end position of the word and the knot of hot word
The alternate position spike of beam position;
Step 4.2: if alternate position spike is greater than the duration of a word, from selecting the word and hot word in vocabulary not in overlapping part, word
Language pronounces similar word as candidate word;
Step 4.3: using preparatory trained language model, giving after sentence tail word to current word word before word and current word
Under the conditions of, predict the probability that each candidate word of current word occurs, and the score as the candidate word;
Step 4.4: if the score of the candidate word of highest scoring is greater than given threshold value, replacing current word with the candidate word;It is no
Then keep current word constant.
7. a kind of hot word recognition methods according to claim 6, it is characterised in that: between step 4-3 and step 4-5,
It is further comprising the steps of respectively: to increase the acoustic confidence information of each word, then and between step 4.3 and step 4.5
Predict the probability that each candidate word of current word occurs, and the score as the candidate word.
8. a kind of hot word recognition methods according to claim 1, it is characterised in that: hot word detecting and alarm is configured to and uses
Family ID is corresponding, uploads hot word and User ID when adding hot word simultaneously, inquires pronunciation dictionary, obtains the pronunciation of hot word and its to sound sound
Prime sequences;Then grammer network is added in hot word;It generates hot word and detects resource, hot word correspondence is added to heat corresponding with User ID
Word detecting and alarm.
9. a kind of hot word identifying system characterized by comprising
Universal phonetic identifies engine, is configured to export the time location of speech recognition result and each word in audio and sets
Reliability;
Hot word detecting and alarm is configured to detect whether to export ID, audio position and its score of hot word there are hot word;
Hot word modified result module is configured to the speech recognition knot using hot word replacement universal phonetic identification engine output
The word of corresponding position in fruit;
Language model modified result module, when being configured to have overlapping as the word in hot word appearance position and current recognition result
Word before and after hot word is modified.
10. a kind of hot word identifying system according to claim 9, it is characterised in that: further include hot word adding module, matched
It is set to and hot word is added or updated to the hot word detecting and alarm.
11. a kind of hot word identification device, characterized in that it comprises: including processor, memory and program;
Described program stores in the memory, and the processor calls the program of memory storage, requires 1 with perform claim
The hot word recognition methods.
12. a kind of computer readable storage medium, which is characterized in that the computer readable storage medium is configured to store journey
Sequence, described program are configured to hot word recognition methods described in perform claim requirement 1.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910706314.6A CN110415705B (en) | 2019-08-01 | 2019-08-01 | Hot word recognition method, system, device and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910706314.6A CN110415705B (en) | 2019-08-01 | 2019-08-01 | Hot word recognition method, system, device and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110415705A true CN110415705A (en) | 2019-11-05 |
CN110415705B CN110415705B (en) | 2022-03-01 |
Family
ID=68365126
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910706314.6A Active CN110415705B (en) | 2019-08-01 | 2019-08-01 | Hot word recognition method, system, device and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110415705B (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110689881A (en) * | 2018-06-20 | 2020-01-14 | 深圳市北科瑞声科技股份有限公司 | Speech recognition method, speech recognition device, computer equipment and storage medium |
CN110879839A (en) * | 2019-11-27 | 2020-03-13 | 北京声智科技有限公司 | Hot word recognition method, device and system |
CN111028830A (en) * | 2019-12-26 | 2020-04-17 | 大众问问(北京)信息科技有限公司 | Local hot word bank updating method, device and equipment |
CN111090720A (en) * | 2019-11-22 | 2020-05-01 | 北京捷通华声科技股份有限公司 | Hot word adding method and device |
CN111161739A (en) * | 2019-12-28 | 2020-05-15 | 科大讯飞股份有限公司 | Speech recognition method and related product |
CN112349278A (en) * | 2020-11-12 | 2021-02-09 | 苏州思必驰信息科技有限公司 | Local hot word training and recognition method and device |
CN112489651A (en) * | 2020-11-30 | 2021-03-12 | 科大讯飞股份有限公司 | Voice recognition method, electronic device and storage device |
CN112599114A (en) * | 2020-11-11 | 2021-04-02 | 联想(北京)有限公司 | Voice recognition method and device |
CN112735428A (en) * | 2020-12-27 | 2021-04-30 | 科大讯飞(上海)科技有限公司 | Hot word acquisition method, voice recognition method and related equipment |
CN113178194A (en) * | 2020-01-08 | 2021-07-27 | 上海依图信息技术有限公司 | Voice recognition method and system for interactive hot word updating |
WO2021232746A1 (en) * | 2020-05-18 | 2021-11-25 | 科大讯飞股份有限公司 | Speech recognition method, apparatus and device, and storage medium |
CN113836270A (en) * | 2021-09-28 | 2021-12-24 | 深圳格隆汇信息科技有限公司 | Big data processing method and related product |
CN114185511A (en) * | 2021-11-29 | 2022-03-15 | 北京百度网讯科技有限公司 | Audio data processing method and device and electronic equipment |
Citations (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5559925A (en) * | 1994-06-24 | 1996-09-24 | Apple Computer, Inc. | Determining the useability of input signals in a data recognition system |
CN101510222A (en) * | 2009-02-20 | 2009-08-19 | 北京大学 | Multilayer index voice document searching method and system thereof |
WO2013000136A1 (en) * | 2011-06-29 | 2013-01-03 | 宇龙计算机通信科技(深圳)有限公司 | Mobile terminal and method, system for inputting network hot words into mobile terminal |
US20160027439A1 (en) * | 2014-07-25 | 2016-01-28 | Google Inc. | Providing pre-computed hotword models |
US20160104480A1 (en) * | 2014-10-09 | 2016-04-14 | Google Inc. | Hotword detection on multiple devices |
CN106326484A (en) * | 2016-08-31 | 2017-01-11 | 北京奇艺世纪科技有限公司 | Error correction method and device for search terms |
CN106782607A (en) * | 2012-07-03 | 2017-05-31 | 谷歌公司 | Determine hot word grade of fit |
US20180158454A1 (en) * | 2016-12-07 | 2018-06-07 | Google Inc. | Preventing of audio attacks |
US20180182390A1 (en) * | 2016-12-27 | 2018-06-28 | Google Inc. | Contextual hotwords |
US20180330717A1 (en) * | 2017-05-11 | 2018-11-15 | International Business Machines Corporation | Speech recognition by selecting and refining hot words |
CN108984529A (en) * | 2018-07-16 | 2018-12-11 | 北京华宇信息技术有限公司 | Real-time court's trial speech recognition automatic error correction method, storage medium and computing device |
CN109271495A (en) * | 2018-08-14 | 2019-01-25 | 阿里巴巴集团控股有限公司 | Question and answer recognition effect detection method, device, equipment and readable storage medium storing program for executing |
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 |
CN110689881A (en) * | 2018-06-20 | 2020-01-14 | 深圳市北科瑞声科技股份有限公司 | Speech recognition method, speech recognition device, computer equipment and storage medium |
-
2019
- 2019-08-01 CN CN201910706314.6A patent/CN110415705B/en active Active
Patent Citations (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5559925A (en) * | 1994-06-24 | 1996-09-24 | Apple Computer, Inc. | Determining the useability of input signals in a data recognition system |
CN101510222A (en) * | 2009-02-20 | 2009-08-19 | 北京大学 | Multilayer index voice document searching method and system thereof |
WO2013000136A1 (en) * | 2011-06-29 | 2013-01-03 | 宇龙计算机通信科技(深圳)有限公司 | Mobile terminal and method, system for inputting network hot words into mobile terminal |
CN106782607A (en) * | 2012-07-03 | 2017-05-31 | 谷歌公司 | Determine hot word grade of fit |
US20160027439A1 (en) * | 2014-07-25 | 2016-01-28 | Google Inc. | Providing pre-computed hotword models |
US20160104480A1 (en) * | 2014-10-09 | 2016-04-14 | Google Inc. | Hotword detection on multiple devices |
CN106326484A (en) * | 2016-08-31 | 2017-01-11 | 北京奇艺世纪科技有限公司 | Error correction method and device for search terms |
US20180158454A1 (en) * | 2016-12-07 | 2018-06-07 | Google Inc. | Preventing of audio attacks |
US20180182390A1 (en) * | 2016-12-27 | 2018-06-28 | Google Inc. | Contextual hotwords |
US20180330717A1 (en) * | 2017-05-11 | 2018-11-15 | International Business Machines Corporation | Speech recognition by selecting and refining hot words |
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 |
CN110689881A (en) * | 2018-06-20 | 2020-01-14 | 深圳市北科瑞声科技股份有限公司 | Speech recognition method, speech recognition device, computer equipment and storage medium |
CN108984529A (en) * | 2018-07-16 | 2018-12-11 | 北京华宇信息技术有限公司 | Real-time court's trial speech recognition automatic error correction method, storage medium and computing device |
CN109271495A (en) * | 2018-08-14 | 2019-01-25 | 阿里巴巴集团控股有限公司 | Question and answer recognition effect detection method, device, equipment and readable storage medium storing program for executing |
Non-Patent Citations (1)
Title |
---|
韩静 等: "基于词语热度的启发式中文句子压缩算法", 《计算机工程与应用》 * |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110689881A (en) * | 2018-06-20 | 2020-01-14 | 深圳市北科瑞声科技股份有限公司 | Speech recognition method, speech recognition device, computer equipment and storage medium |
CN111090720A (en) * | 2019-11-22 | 2020-05-01 | 北京捷通华声科技股份有限公司 | Hot word adding method and device |
CN111090720B (en) * | 2019-11-22 | 2023-09-12 | 北京捷通华声科技股份有限公司 | Hot word adding method and device |
CN110879839A (en) * | 2019-11-27 | 2020-03-13 | 北京声智科技有限公司 | Hot word recognition method, device and system |
CN111028830A (en) * | 2019-12-26 | 2020-04-17 | 大众问问(北京)信息科技有限公司 | Local hot word bank updating method, device and equipment |
CN111028830B (en) * | 2019-12-26 | 2022-07-15 | 大众问问(北京)信息科技有限公司 | Local hot word bank updating method, device and equipment |
CN111161739A (en) * | 2019-12-28 | 2020-05-15 | 科大讯飞股份有限公司 | Speech recognition method and related product |
CN111161739B (en) * | 2019-12-28 | 2023-01-17 | 科大讯飞股份有限公司 | Speech recognition method and related product |
CN113178194A (en) * | 2020-01-08 | 2021-07-27 | 上海依图信息技术有限公司 | Voice recognition method and system for interactive hot word updating |
CN113178194B (en) * | 2020-01-08 | 2024-03-22 | 上海依图信息技术有限公司 | Voice recognition method and system for interactive hotword updating |
WO2021232746A1 (en) * | 2020-05-18 | 2021-11-25 | 科大讯飞股份有限公司 | Speech recognition method, apparatus and device, and storage medium |
CN112599114A (en) * | 2020-11-11 | 2021-04-02 | 联想(北京)有限公司 | Voice recognition method and device |
CN112599114B (en) * | 2020-11-11 | 2024-06-18 | 联想(北京)有限公司 | Voice recognition method and device |
CN112349278A (en) * | 2020-11-12 | 2021-02-09 | 苏州思必驰信息科技有限公司 | Local hot word training and recognition method and device |
CN112489651A (en) * | 2020-11-30 | 2021-03-12 | 科大讯飞股份有限公司 | Voice recognition method, electronic device and storage device |
CN112735428A (en) * | 2020-12-27 | 2021-04-30 | 科大讯飞(上海)科技有限公司 | Hot word acquisition method, voice recognition method and related equipment |
CN113836270A (en) * | 2021-09-28 | 2021-12-24 | 深圳格隆汇信息科技有限公司 | Big data processing method and related product |
CN114185511A (en) * | 2021-11-29 | 2022-03-15 | 北京百度网讯科技有限公司 | Audio data processing method and device and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
CN110415705B (en) | 2022-03-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110415705A (en) | A kind of hot word recognition methods, system, device and storage medium | |
CN107679033B (en) | Text sentence break position identification method and device | |
CN103559881B (en) | Keyword recognition method that languages are unrelated and system | |
JP5901001B1 (en) | Method and device for acoustic language model training | |
CN108710704B (en) | Method and device for determining conversation state, electronic equipment and storage medium | |
CN111145733B (en) | Speech recognition method, speech recognition device, computer equipment and computer readable storage medium | |
CN112257437B (en) | Speech recognition error correction method, device, electronic equipment and storage medium | |
CN110689881B (en) | Speech recognition method, speech recognition device, computer equipment and storage medium | |
WO2014187096A1 (en) | Method and system for adding punctuation to voice files | |
CN107077843A (en) | Session control and dialog control method | |
CN110930993A (en) | Specific field language model generation method and voice data labeling system | |
US11526512B1 (en) | Rewriting queries | |
US11741948B2 (en) | Dilated convolutions and gating for efficient keyword spotting | |
CN113380238A (en) | Method for processing audio signal, model training method, apparatus, device and medium | |
CN110751234A (en) | OCR recognition error correction method, device and equipment | |
CN110503943B (en) | Voice interaction method and voice interaction system | |
US10468031B2 (en) | Diarization driven by meta-information identified in discussion content | |
JP2014206642A (en) | Voice recognition device and voice recognition program | |
CN113204667B (en) | Method and device for training audio annotation model and audio annotation | |
CN111128181B (en) | Recitation question evaluating method, recitation question evaluating device and recitation question evaluating equipment | |
CN114399992A (en) | Voice instruction response method, device and storage medium | |
Sarikaya et al. | Word level confidence measurement using semantic features | |
CN110781072A (en) | Code auditing method, device and equipment based on machine learning and storage medium | |
KR100449912B1 (en) | Apparatus and method for detecting topic in speech recognition system | |
CN108511001A (en) | Voice monitoring method and device, storage medium, terminal |
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 |