CN103811000A - Voice recognition system and voice recognition method - Google Patents

Abstract

The present invention relates to a speech recognition system and method. The speech recognition method includes the following steps: S1: collecting audio information to be recognized and region information corresponding to the audio information; S2: calling the region information according to the region information Corresponding voice database and grammar database, call the corresponding grammar file in the grammar database; S3: Identify the audio information according to the grammar file and the voice database. The invention has the beneficial effect of improving the recognition rate of multi-timbre and large vocabulary.

Description

Speech Recognition System and Method

technical field

The invention relates to the field of speech recognition, in particular to a speech recognition system and method.

Background technique

The existing speech recognition methods mainly include dynamic time correction technology DTW, vector quantization technology VQ, hidden Markov model HMM and artificial neural network ANN.

Dynamic Time Rectification Technology (DTW) is an earlier pattern matching and model training technology. It applies dynamic programming method to successfully solve the problem of unequal duration when comparing speech signal feature parameter sequences, and obtains good performance in isolated word speech recognition. .

Vector quantization technology VQ extracts the feature vector from the training speech, obtains the feature vector set, generates the codebook through the LBG algorithm, extracts the feature vector sequence from the test speech during recognition, matches them with each codebook, and calculates the average quantization error of each, The codebook with the smallest average quantization error is selected as the recognized speech.

The Hidden Markov Model (HMM) is a parametric representation of the time-varying features of the speech signal. It describes the statistical characteristics of the signal by two interrelated stochastic processes, one of which is a hidden (unobservable) Markov chain with finite states. , and the other is the stochastic process (observable) of the observation vector associated with each state of the Markov chain. The characteristics of the hidden Markov chain are revealed by the observable signal characteristics. In this way, the characteristics of a segment of the time-varying speech signal are described by the stochastic process of the corresponding state observation symbols, and the time-varying signal is described by the transition probability of the hidden Markov chain. Model parameters include HMM topology, state transition probabilities, and a set of random functions describing the statistical properties of observed symbols. According to the characteristics of random functions, HMM models can be divided into discrete hidden Markov models (and continuous hidden Markov models and semi-continuous hidden Markov models.

The application of artificial neural network in speech recognition is another research hotspot now. ANN is essentially an adaptive nonlinear dynamic system, which simulates the principle of human neuron activity, and has the ability of self-learning, association, comparison, reasoning and generalization.

There are some deficiencies in the above-mentioned mainstream speech recognition methods. The main disadvantage is that if the same vocabulary is spoken with accents from different regions, the timbre will change to a certain extent, which will greatly reduce the speech recognition rate.

Contents of the invention

A speech recognition system and method are provided aiming at the defect in the existing speech recognition technology that the speech recognition rate will be reduced when the timbre is changed.

The technical solution adopted in the present invention to solve the technical problem is: provide a kind of speech recognition method, comprise the following steps:

S1: Collect audio information to be identified and region information corresponding to the audio information;

S2: call the voice database and the grammar database corresponding to the region information according to the region information, and call the corresponding grammar file in the grammar database;

S3: Identify the audio information according to the grammar file and the voice database.

In the speech recognition method provided by the present invention, the step S1 further includes: collecting service category information corresponding to the audio information;

Calling the grammar file in the grammar database in the step S2 further includes: calling a grammar file corresponding to the business type information in the grammar database according to the business type information.

In the voice recognition method provided by the present invention, step S0 is also included before the step S1: multiple voice databases and multiple grammar databases are established and stored according to different regions, and grammar files are generated corresponding to business types in each grammar database .

In the speech recognition method provided by the present invention, in the step S1, the collecting the area information corresponding to the audio information includes: querying the number of the server that sends the audio information, querying and extracting the Region information corresponding to the audio information.

In the voice recognition method provided by the present invention, the step S3 further includes: setting keywords or words, starting timing when the audio information is recognized, stopping timing when the keywords or words are recognized, and outputting the recognition time.

The present invention also provides a speech recognition system, comprising:

A collection module, the collection module comprising a first collection unit for collecting audio information to be identified and a second collection unit for collecting region information corresponding to the audio information;

A scheduling module, the scheduling module is used to select and call the voice database and the grammar database corresponding to the region according to the region information and call the corresponding grammar file in the grammar database;

A recognition module, configured to recognize the audio information according to the grammar file and the speech database.

In the speech recognition system provided by the present invention, the collection module further includes a third collection unit for collecting business type information corresponding to the audio information, and the scheduling module is also used for Call the grammar file corresponding to the business type information in the grammar database.

In the speech recognition system provided by the present invention, the speech recognition system further includes a storage module, the storage module is used to store the plurality of speech databases and the plurality of grammar databases established according to different regions, each of the Generate grammar files corresponding to business types in the grammar database.

In the speech recognition system provided by the present invention, the second collection unit includes a first query subunit for querying the number of the server that sends out the audio information, and a first query subunit for querying and extracting the corresponding audio information according to the number. The first extraction subunit of the region information.

In the speech recognition system provided by the present invention, it also includes a timing module and a setting module, the setting module is used to set keywords or words, and the timing module is used to time when the recognition module starts to recognize the audio information, When the recognition module recognizes the keywords or words, it stops counting and outputs the recognition time.

Compared with the prior art, the speech recognition system and method provided by the present invention have the following beneficial effects: when the speech recognition system and method provided by the present invention recognize the audio information to be recognized, they call the region information corresponding to the audio information. Speech database and grammatical database, so that the same vocabulary can be avoided due to different accents in different regions, resulting in a decrease in the speech recognition rate, and because the types of business in different regions are different, it is possible to call the corresponding grammatical database for the types of business in different regions Improve the efficiency of speech recognition; therefore, the present invention has the beneficial effect of improving the multi-timbre recognition rate.

Description of drawings

The present invention will be further described below in conjunction with accompanying drawing and embodiment, in the accompanying drawing:

Fig. 1 is the functional block diagram of the speech recognition system in the first embodiment of the present invention;

Fig. 2 is the functional block diagram of the second acquisition unit in the embodiment shown in Fig. 1 of the present invention;

Fig. 3 is the functional block diagram of the speech recognition system in the second embodiment of the present invention;

Fig. 4 is the functional block diagram of the speech recognition system in the third embodiment of the present invention;

Fig. 5 is the functional block diagram of the speech recognition system in the fifth embodiment of the present invention;

Fig. 6 is a block flow diagram of the speech recognition method in the first embodiment of the present invention;

Fig. 7 is a block flow diagram of the speech recognition method in the second embodiment of the present invention;

Fig. 8 is a flowchart of the voice recognition method in the third embodiment of the present invention.

Detailed ways

In order to solve the defect in the prior art that the speech recognition rate is greatly reduced due to the change of timbre, the innovation of the present invention is to provide corresponding speech databases and grammar databases for different timbres in different regions, so as to improve the speech recognition rate.

In order to make the purpose, technical solution and description of the advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Fig. 1 shows the speech recognition system in the first embodiment of the present invention, and it is mainly used in business monitoring, and the monitoring server automatically imitates the real user to issue instructions to the operator's server and recognizes the audio information that the operator's server replies, To judge whether the business service level meets expectations. The speech recognition system includes: a collection module 1 , a scheduling module 2 and a recognition module 3 . The collection module 1 includes a first collection unit 11 for collecting audio information to be identified and a second collection unit 12 for collecting region information corresponding to the audio information. The scheduling module 2 is used to select and call the speech database and the grammar database corresponding to the region according to the region information, and call the corresponding grammar file in the grammar database. The recognition module 3 is used to recognize the audio information according to the called grammar file and the voice database and output the recognition result. The acquisition module 1, the scheduling module 2 and the identification module 3 are connected in sequence.

The voice recognition system is realized by installing a program on the monitoring server. That is, the functions of the acquisition module 1 , the scheduling module 2 and the identification module 3 are realized by the CPU of the monitoring server working according to the software program for realizing these functions.

The monitoring server simulates a real user to issue instructions to the operator's server, and the operator's server automatically replies with audio information, and the first acquisition unit 11 of the acquisition module 1 installed on the monitoring server collects the audio information to be identified through a communication link. The second collection unit 12 of the collection module 1 is configured to obtain region information corresponding to the audio information to be recognized and send the collected region information to the scheduling module 2 . According to the region information, the scheduling module 2 invokes the voice database and the grammar database corresponding to the region information, and invokes the corresponding grammar files in the grammar database. The speech recognition module 3 recognizes the audio information to be recognized according to the called speech database and grammar file, and outputs the recognition result.

As shown in Figure 2, the second collection unit 12 may include a first query subunit for querying the number of the operator's server and a first extraction subunit for querying and extracting the area information of the operator's server according to the number . The scheduling module 2 schedules the voice database and grammar database corresponding to the area information.

It can be understood that the second acquisition unit 12 can also acquire the area information corresponding to the audio information to be identified in other ways.

In this embodiment, the recognition module 3 implements speech recognition by using the template matching HMM method. The recognition module 3 includes a feature extraction unit, an acoustic recognition unit, and a language recognition unit. Wherein, the feature extraction unit is used to perform feature extraction on the speech waveform corresponding to the collected audio information, so as to obtain speech acoustic features. Traditional speech feature extraction algorithms can be used to extract features from speech waveforms, such as extracting MFCC (Mel Frequency Cepstrum System), LPC (Linear Predictive Coding Coefficient), and speech energy. The acoustic recognition unit compares the feature quantities of each acoustic model in the speech database called by the scheduling module 2 with the feature quantities extracted by the feature extraction unit in order to obtain the phoneme string corresponding to the audio information.

Among them, the acoustic model is data obtained by modeling what kind of feature quantity the speech will be. Since the pronunciation of the same word is different in each region, its acoustic model is also different. Therefore, multiple speech databases should be established according to the differences in each region. Each speech database is equipped with an acoustic model established according to the pronunciation habits of the region. , which can improve speech recognition accuracy and recognition speed. The language recognition unit recognizes the phoneme string according to the grammar file invoked by the scheduling module 2, obtains and outputs the word string. Since the business conditions of each region are different, establishing corresponding grammar databases according to the business conditions of different regions can improve the speech recognition rate.

For the same business in the same region, its phonetic vocabulary is basically fixed. Therefore, it is possible to write a small-scale grammar file for each business. The grammar file combines some vocabulary phrases through specific rules to define the data model of the output result of the speech recognition unit. Therefore, as shown in FIG. 3, in the second embodiment, on the basis of the first embodiment, the acquisition module 1 can also include a third acquisition unit 13, which is used to acquire the audio information corresponding to the audio information. business type information. The dispatching module 2 first invokes the grammar database corresponding to the region information according to the region information, and then invokes the grammar file corresponding to the business category in the grammar database according to the service category information. The grammar file produced according to the type of business mainly includes key vocabulary phrases related to the business, so the recognition efficiency can be improved.

For example, region B has four services b1, b2, b3, and b4. Corresponding to area B, a grammar database is established according to these four kinds of services, and the grammar database includes four grammar files correspondingly. Each grammatical file only has various common grammatical words related to the business and phoneme strings corresponding to the common grammatical words. When the recognition unit recognizes the phoneme string, it outputs the word string corresponding to the phoneme string as a recognition result.

In this embodiment, the scheduling module 2 schedules the speech database and the grammar database in a specific process: acquiring audio information and corresponding province information. Obtain the information of the current voice database: the region information corresponding to the current voice database, the name of the current voice database, and the path of the current voice database. Obtain the information of the speech database to be recognized: the province information to be recognized, the name of the speech database to be recognized, and the path of the speech database to be recognized. The current voice database information is compared with the voice database information to be recognized. If they are not consistent, switch between the backup voice database of the corresponding province and the current voice database, and return the grammar file of the province corresponding to the current voice database. If it is consistent, it will directly return the grammar file of the province corresponding to the current speech database.

As shown in FIG. 4 , in the third embodiment, on the basis of the second embodiment, the speech recognition system provided by the present invention further includes a storage module 4 . The storage module 4 is used for storing multiple speech databases and multiple grammar databases generated according to different regions. The scheduling module 2 calls the speech database and the grammar database from the storage module 4 .

In the fourth embodiment, on the basis of the third embodiment, the speech recognition system provided by the present invention further includes a repair module. When the scheduling module 2 fails to schedule the voice database, it automatically repairs the voice database. In the actual work process, there are many reasons why the scheduling module 2 fails to schedule the voice database. Voice bank switching can be done in the form of renaming. If the renaming is wrong, the automatic repair process is to modify the incorrectly renamed voice library, or manually delete the voice library, then the automatic repair program will find it in the backup voice library path and copy it to the existing voice library path . It is also possible that the voice database is in the occupied state. At this time, it is necessary to wait. After the occupied voice database is used up, it can be called again by the scheduling module 2. It may be that there is no voice database corresponding to the region in the storage module 4, so at this time, a prompt message that the voice database lacks needs to be issued. The repair module has the function of automatically repairing the voice database. Each voice database has the same standby voice database, and when calling the voice database in a certain area fails, it can switch to call the corresponding standby voice database.

It can be understood that, on the basis of the above-mentioned embodiments, in the fifth embodiment, as shown in FIG. Sensitivity, recognition language information, etc. This setting module 5 can also set keywords or words, and timing module 6 counts when recognition module 3 starts to identify, stops timing when recognition module 3 recognizes the keywords or words that setting module 6 presets, and timing module 6 outputs recognition module 3 The time it takes to recognize a key word or phrase. This time can be used to judge whether the recognition efficiency of the recognition module 3 meets expectations.

As shown in Fig. 6, the present invention also provides a kind of speech recognition method, and this speech recognition method comprises the following steps in the first embodiment:

S1: Collect audio information to be identified and region information corresponding to the audio information. In this step, the monitoring server simulates a real person to input an instruction and sends it to the operator's server, and the operator's server automatically replies to the audio information according to the instruction, and the first acquisition unit 11 of the acquisition module 1 installed in the monitoring server passes the communication link Collect the audio information. The second collection unit 12 of the collection module 1 can query the number of the operator's server that sends out the audio information, and then query and extract the area information corresponding to the operator's server according to the number, that is, the area information corresponding to the audio information.

S2: Select the speech database and the grammar database corresponding to the region according to the region information; call the grammar files in the grammar database. In this step, the dispatching module 2 invokes the speech database and the grammar database corresponding to the district information in the storage module 4 according to the region information collected by the second collection unit 12, and further dispatches the grammar files in the grammar database.

S3: The recognition module 3 recognizes the audio information according to the scheduled grammar file and the voice database, and outputs the recognition result.

On the basis of the first embodiment of the speech recognition method, as shown in FIG. 7 , in the second embodiment, step S1 may further include the following steps: the third acquisition unit 13 of the acquisition module 1 acquires the corresponding audio information business type information. Correspondingly, in step S2, invoking the grammar file in the grammar database specifically includes: the scheduling module 2 calls the grammar file corresponding to the business type information in the grammar database according to the business type information. In this step, the identification sensitivity can be set in advance for the identification module according to the type of identified business. Also keyword or vocabulary can be set, timing module 5 begins timing when identification module 3 begins to identify, stops timing when identification module 3 recognizes the keyword or word that setting module 6 presets, and timing module 5 outputs identification module 3 and recognizes The time used for keywords or phrases. This time can be used to judge whether the recognition efficiency of the recognition module 3 meets expectations. As shown in Figure 8, on the basis of the second embodiment, the voice recognition method may also include step S0 in the third embodiment: set up multiple voice databases and multiple grammar databases according to different regions and store them in the storage module 4 In each grammar database, multiple grammar files are generated according to different types of business.

In the fourth embodiment, the region information is the province information, and the voice recognition method includes the following steps:

S1: Acquisition module 1 collects the path and name of the audio information to be identified, the province information corresponding to the audio information to be identified, identifies the business information corresponding to the audio information to be identified, and keywords related to the business, and collects these The information is transmitted to the scheduling module.

S2: The scheduling module 2 reads the province, name, and path to which the current voice database belongs, and compares the region information to which the current voice database belongs with the province information corresponding to the audio information to be recognized.

If the scheduling module 2 judges that the province information corresponding to the current voice database is the same as the province information corresponding to the audio information with recognition, it does not need to switch the voice database, and directly returns the grammar corresponding to the province information of the audio information to be recognized The path and name of the database.

If scheduling module 2 judges that the province information corresponding to the current voice database is different from the incoming province, then scheduling module 2 performs voice database switching, that is, renaming the current voice database database to the backup voice database of the corresponding province, and then combining with the audio to be identified The voice database corresponding to the information is renamed to the current voice database, and finally the grammar database of the province corresponding to the current voice database is returned. And call the grammar file in the grammar database according to the business category information.

S3: If the scheduling module 2 fails to schedule the speech database and grammar database, it will directly exit the recognition process, and perform automatic retrieval and automatic repair of the speech database. If the scheduling is successful, the recognition module 3 is set according to the parameters required for recognition, such as sensitivity, accuracy, and recognition language.

A recognition context is created according to the recognition process of the recognition module 3, and is used to record information generated during the recognition process, including recognition success information, error information, abnormal information, and the like.

Create a recognition audio stream in the recognition module 3, set the Format parameters of the audio stream, such as playback frequency, hertz, etc., and bind the audio information that needs to be recognized to the recognition audio stream. Activate the recognition module 3, and the recognition module 3 loads the voice database and grammar files and imports the recognition audio stream to start the recognition process. After the recognition module 3 is activated, it waits for the recognition-specific Windows custom message to be triggered. When the recognition-specific Windows custom message is triggered, if the parameter of the Windows custom message is recognition end, the recognition module exits the recognition process. If the parameter of the Windows custom message is recognition content, the recognition content is extracted and saved in the recognition result. If the parameter of the Windows custom message is an identification error, handle it according to the specific error type.

In this embodiment, it is also possible to create other user-defined prompt messages and ensure that they do not conflict with other prompt messages, and set a message processing function for identifying a dedicated message and a type of identifying message to focus on monitoring.

To sum up, since the speech recognition system and method provided by the present invention are provided with speech databases and grammar databases for different regions, during the recognition process, the speech recognition rate and recognition speed can be provided by correspondingly calling the speech database and grammar database. In addition, a grammar file is specially set up in the grammar database corresponding to each kind of business, and the real business type in the recognition process can be called the corresponding grammar file, which can further improve the speech recognition rate and recognition speed.

It should be understood that the embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementations, and the above-mentioned specific implementations are only illustrative rather than restrictive. Under the enlightenment of the present invention, those skilled in the art can also make many forms without departing from the purpose of the present invention and the scope protected by the claims, and these all belong to the protection of the present invention.

Claims (10)

1. A speech recognition method, characterized in that, comprises the following steps: S1: Collect audio information to be identified and region information corresponding to the audio information; S2: call the voice database and the grammar database corresponding to the region information according to the region information, and call the corresponding grammar file in the grammar database; S3: Identify the audio information according to the grammar file and the voice database. 2. The speech recognition method according to claim 1, wherein the step S1 further comprises: collecting business category information corresponding to the audio information; Calling the grammar file in the grammar database in the step S2 further includes: calling a grammar file corresponding to the business type information in the grammar database according to the business type information. 3. The speech recognition method according to claim 2, characterized in that, before said step S1, step S0 is also included: setting up and saving a plurality of speech databases and a plurality of grammar databases according to different regions, each of said grammar databases Generate grammar files corresponding to business types. 4. The speech recognition method according to any one of claims 1 to 3, characterized in that, in the step S1, the collecting the region information corresponding to the audio information comprises: querying the server that sends the audio information number, query and extract the region information corresponding to the audio information according to the number. 5. The speech recognition method according to any one of claims 1 to 3, characterized in that, in the step S3, it also includes: setting keywords or words, timing when starting to recognize the audio information, recognizing the key Stop timing when the word or phrase is detected, and output the recognition time. 6. The speech recognition system according to claim 1, characterized in that, comprising: A collection module (1), the collection module (1) comprising a first collection unit (11) for collecting audio information to be identified and a second collection unit (12) for collecting area information corresponding to the audio information ; A dispatching module (2), the dispatching module (2) is used to select and call a speech database and a grammar database corresponding to the region according to the region information, and call a corresponding grammar file in the grammar database; A recognition module (3), configured to recognize the audio information according to the grammar file and the voice database. 7. The speech recognition system according to claim 6, characterized in that, the collection module (1) further comprises a third collection unit (13) for collecting business category information corresponding to the audio information, the The scheduling module (2) is further configured to call a grammar file corresponding to the business category information in the grammar database according to the business category information. 8. The voice recognition system according to claim 7, further comprising a storage module (4), the storage module (4) is used to store the multiple voice databases and the multiple voice databases established according to different regions. Grammar databases, and grammar files are generated corresponding to business types in each of the grammar databases. 9. The speech recognition system according to any one of claims 6 to 8, characterized in that, the second collection unit (12) includes a first query subunit for querying the number of the server that sends out the audio information And a first extracting subunit for querying and extracting region information corresponding to the audio information according to the number. 10. The speech recognition system according to any one of claims 6 to 8, characterized in that it also includes a timing module (5) and a setting module (6), and the setting module (5) is used to set keywords or words , the timing module (5) is used to start timing when the recognition module (6) starts to recognize the audio information, and when the recognition module (6) recognizes the keyword or word, stop timing and output the recognition time .

Images