TW201935460A - Speech recognition device and speech recognition method - Google Patents

Abstract

A speech recognition device including a speech recognition module and a probability comparison module. The speech recognition module receives a speech feature. The speech recognition module includes an acoustic model, language model, and a plurality of acoustic dictionaries. The speech recognition module analyzes the speech feature by at least one of the plurality of acoustic dictionaries, the acoustic model and the language model to generate at least one string probability and at least one string data. The plurality of acoustic dictionaries correspond to a plurality of different accent type. The probability comparison module judges a highest probability among the at least one string probability to output one of the at least one string data corresponding to the highest probability. In addition, a speech recognition method is also provided.

Description

Voice recognition device and voice recognition method

The present invention relates to a recognition technology, and more particularly, to a voice recognition device and a voice recognition method.

With the development of speech recognition technology, more and more electronic devices have the function of speech recognition. Speech recognition usually takes the parameters of speech features from the input speech signal and compares them with samples in the database to find samples with a low degree of dissimilarity from the input. However, if the voice provided by the user has a special accent, a problem that a voice signal cannot be effectively recognized may occur. Therefore, the general speech recognition method is to establish multiple sets of acoustic models, multiple sets of language models, and multiple acoustic dictionaries for different types of accents to generate multiple string probabilities and multiple string data, respectively.

However, because different acoustic models have different phonemes and probability bases, and different language models also have different probability bases, the general way of speech recognition is that the probabilities of multiple strings produced are not actually comparable, and Low accuracy. In addition, the general speech recognition method requires a large amount of analysis and calculation, and has the problem that the efficiency of speech recognition is too low. In this regard, how to propose a speech recognition technology that can effectively adapt to users of different accent types, to effectively recognize the speech signals of different accent types provided by different users, and to provide efficient speech recognition results is currently important in the field One of the topics.

The invention provides a speech recognition device and a speech recognition method, which can effectively recognize the speech signals of different accents, so as to effectively decode and analyze the speech characteristics of the speech signals according to different types of accents.

The speech recognition device of the present invention includes a speech recognition module and a probability comparison module. The voice recognition module is used to receive voice features. The speech recognition module includes an acoustic model, a language model, and a plurality of acoustic dictionaries. The speech recognition module analyzes the speech features by using at least one of the plurality of acoustic dictionaries, the acoustic model, and the language model to generate at least one string probability and at least one string data. The plurality of acoustic dictionaries correspond to a plurality of different accent types. The probability comparison module is coupled to the speech recognition module. The probability comparison module determines a highest probability among the at least one string probability to output one of the at least one string data corresponding to the highest probability.

In an embodiment of the present invention, the speech recognition module generates multiple string probabilities and multiple string data corresponding to different accent types by using the multiple acoustic dictionaries. The speech feature is of the same accent type as one of the plurality of acoustic dictionaries, so that the string probability generated via one of the plurality of acoustic dictionaries is the highest probability.

In an embodiment of the present invention, the voice recognition device further includes a feature extraction module. The feature extraction module is coupled to the voice recognition module. The feature extraction module is used to receive a voice signal. The feature extraction module analyzes the voice signal to provide the voice feature to a voice recognition module.

In an embodiment of the present invention, the voice recognition device further includes an accent recognition module. The accent recognition module is coupled to the feature extraction module and the voice recognition module. The accent recognition module is configured to analyze the voice signal to determine the accent type of the voice signal, and select one of the plurality of acoustic dictionaries to analyze the voice characteristics.

In an embodiment of the present invention, the above-mentioned accent recognition module selectively outputs the speech feature to the plurality of acoustic dictionaries corresponding to the accent type according to the accent type of the voice signal. One of the plurality of acoustic dictionaries corresponding to the accent type, the acoustic model, and the language model to analyze the speech feature, and output one string of data .

The speech recognition method of the present invention includes the following steps: receiving a speech feature, and analyzing the speech feature to generate at least one string probability and at least one string by using at least one of a plurality of acoustic dictionaries, an acoustic model, and a language model. Data, wherein the plurality of acoustic dictionaries correspond to a plurality of different accent types; and determining the highest probability among the at least one string probability to output one of the at least one string information corresponding to the highest probability One.

In an embodiment of the present invention, the above-mentioned multiple acoustic dictionaries generate multiple string probabilities and multiple string data corresponding to different accent types, and one of the voice characteristics and the multiple acoustic dictionaries is Are the same accent type such that the string probability generated via one of the plurality of acoustic dictionaries is the highest probability.

In an embodiment of the present invention, the above-mentioned voice recognition method further includes the following steps: receiving a voice signal, and analyzing the voice signal to obtain the voice feature.

In an embodiment of the present invention, the voice recognition method further includes the steps of analyzing the voice signal to determine an accent type of the voice signal, and selecting one of the plurality of acoustic dictionaries to analyze the voice dictionary. The speech features are described.

In an embodiment of the present invention, the above-mentioned voice recognition method further includes the following steps: selectively outputting the voice feature to the plurality of corresponding to the accent type according to the accent type of the voice signal One of the acoustic dictionaries to analyze the speech feature by one of the plurality of acoustic dictionaries corresponding to the accent type, the acoustic model, and the language model; and outputting one of the words String data.

Based on the above, the speech recognition device and speech recognition method of the present invention can analyze speech characteristics through multiple acoustic dictionaries corresponding to multiple different accent types, so that the string probabilities output by the speech recognition module are comparable. In addition, the speech recognition device and speech recognition method of the present invention can also determine the accent type of speech information by using an accent recognition module, and select an acoustic dictionary of the same accent type to generate a speech recognition result.

In order to make the above features and advantages of the present invention more comprehensible, embodiments are hereinafter described in detail with reference to the accompanying drawings.

In order to make the content of the present invention easier to understand, the following specific embodiments are examples based on which the present invention can be implemented. In addition, wherever possible, the same reference numbers are used in the drawings and embodiments to refer to the same or similar components.

FIG. 1 is a schematic diagram of a speech recognition device according to an embodiment of the present invention. Referring to FIG. 1, the speech recognition device 100 includes a processing device 110, an input device 120, a storage device 130, and an output device 140. The processing device 110 is coupled to the input device 120, the storage device 130 and the output device 140. The voice recognition device 100 is, for example, a device having a computing function such as a mobile phone, a smart phone, a Personal Digital Assistant (PDA), a tablet computer, a notebook computer, a desktop computer, or a car computer.

In this embodiment, the processing device 110 is, for example, a Central Processing Unit (CPU), or other programmable general-purpose or special-purpose microprocessor (microprocessor), digital signal processor (Digital Signal Processor) (DSP), programmable controller, Application Specific Integrated Circuits (ASIC), Programmable Logic Device (PLD), other similar processing circuits, or a combination of these devices.

In this embodiment, the input device 120 is used to receive a voice signal. The input device 120 may be, for example, a microphone. The input device 120 is configured to receive an analog voice signal sent by a user, convert the analog voice signal into a digital voice signal, and transmit the analog voice signal to the processing device 110.

In this embodiment, the storage device 130 may be, for example, an electronically erasable rewritable read-only memory (Electrically-Erasable Programmable Read-Only Memory, EEPROM), an embedded Multi Media Card (eMMC), Dynamic Random Access Memory (DRAM), Flash Memory or Non-Volatile Random Access Memory (NVRAM), etc.

In this embodiment, the output device 140 is a display such as a cathode ray tube (CRT) display, a liquid crystal display (LCD), a plasma display, a touch display, and the like. device. The output device 140 can display string data corresponding to a maximum probability among the generated string probabilities. In an embodiment, the output device 140 may also be a speaker, and is used to play string data corresponding to the largest probability among the generated string probabilities. Alternatively, in another embodiment, the output device 140 may also provide string data corresponding to a maximum probability among the generated string probabilities to a specific application program, so that the specific application program may correspondingly execute a specific function or operating.

In this embodiment, the storage device 130 may be used to store a plurality of modules for the processing device 110 to read and execute, so as to implement the voice recognition operation according to the embodiments of the present invention. Specifically, the modules stored in the storage device 130 may include, for example, a feature extraction module, an accent recognition module, a voice recognition module, a probability comparison module, and the like, such as an analysis and calculation module. In this embodiment, the voice recognition device 100 can obtain the voice information through the input device 120 and analyze the voice information through these modules of the storage device 130 to generate corresponding analysis results. That is, the voice recognition device 100 of this embodiment may have a voice recognition function.

FIG. 2 is a flowchart of a speech recognition method according to an embodiment of the present invention. FIG. 3 is a schematic diagram of a speech recognition device according to an embodiment of the invention. Referring to FIG. 1 to FIG. 3 at the same time, in this embodiment, the processing device 110 may execute the voice recognition module 330 and the probability comparison module 340 stored in the storage device 130. The speech recognition module 330 includes an acoustic model 331, a plurality of acoustic dictionaries 332A, 332B, and 332C, a language model 333, and a plurality of decoders 334A, 334B, and 334C.

A possible implementation is described with reference to the flowchart of FIG. 2. In step S210, the voice recognition module 330 receives the voice feature VC, and uses at least one of these acoustic dictionaries 332A, 332B, and 332C, an acoustic model 331, and The language model 333 generates at least one string probability and at least one string data. In step S220, the probability comparison module 340 determines the highest probability among the at least one string probability to output the string data SD having the highest probability.

Specifically, in this embodiment, these acoustic dictionaries 332A, 332B, and 332C correspond to a plurality of different accent types, and these accent types are, for example, a Beijing area accent, a Shanghai area accent, a Guangzhou area accent, or a Fujian area accent. In this embodiment, the voice recognition module 330 is configured to receive the voice feature VC and correspondingly generate multiple string probabilities and multiple string data to the probability comparison module 340. The probability comparison module 340 compares the string probabilities to determine the highest probability among the string probabilities, and outputs the string data SD corresponding to the highest probability among the string probabilities to the output device 140.

In this embodiment, these decoders 334A, 334B, and 334C are used to generate string data and string probabilities that are more appropriate or have the highest probability, respectively, based on the acoustic model 331, the acoustic dictionary 332A, 332B, 332C, and the language model 333. It is worth noting that, in order to make the string probabilities generated by the speech recognition module 330 comparable, the speech recognition module 330 of this embodiment analyzes the speech feature VC only through one acoustic model 331 and one language model 333. In this embodiment, the acoustic model 331 is obtained by training through a voice database. For example, the acoustic model 331 is modeled by using a hidden Markov model (HMM). The language model 333 is obtained through training of a text corpus. For example, a probability statistics method is used to reveal the statistical rules inherent in the language unit. In addition, the speech recognition module 330 of this embodiment establishes these acoustic dictionaries 332A, 332B, and 332C for a plurality of different accent types. Among them, these acoustic dictionaries 332A, 332B, and 332C are also obtained by training through a voice database, but corresponding For different accent types. These acoustic dictionaries 332A, 332B, and 332C are used to process vocabulary and pronunciation of different accent types, respectively.

That is, the processing device 110 of this embodiment may analyze a speech feature VC by executing an acoustic model 331 to obtain a corresponding phone or syllable, and then pass these acoustic dictionaries 332A, 332B, and 332C. At least one of them is used to obtain the corresponding word or word, and finally a language model 333 is used to determine the probability of a series of words becoming a sentence. It is worth noting that both the acoustic model 331 and the language model 333 of this embodiment belong to a probability model, and since the speech recognition module 330 of this embodiment only builds a single acoustic model 331 and a single language model 333, it corresponds to different The probabilities of the multiple strings generated by these accent-type acoustic dictionaries 332A, 332B, and 332C are comparable.

For example, the acoustic dictionary 332A is established, for example, for the Guangzhou area, and thus, acoustic data such as “bei jin (Beijing)”, “ci fan (meal)”, and “re qi (热气)” may be recorded. The acoustic dictionary 332B is, for example, created for the Beijing area, and therefore, acoustic data such as "bei jing (Beijing)", "chi fan (eating)", and "re qi (热气)" may be recorded. The acoustic dictionary 332C is established, for example, for the Fujian region, and therefore, acoustic data such as “bei jin (Beijing)”, “ci fan (dining)”, and “le qi (热气)” may be recorded. That is, since these acoustic dictionaries 332A, 332B, and 332C are respectively established for different accent types, as long as the speech feature VC and one of these acoustic dictionaries are the same accent type, the acoustic model 331, the language model 333, and The string probability generated by one of these acoustic dictionaries 332A, 332B, 332C corresponding to the same accent type will be the highest probability.

However, it should be noted that the number and types of the acoustic dictionaries of the present invention are not limited to those shown in FIG. 3, and the schematic diagram of FIG. 3 is only used to illustrate an exemplary embodiment of a feasible speech recognition module.

FIG. 4 is a schematic diagram of another speech recognition device according to an embodiment of the invention. FIG. 5 is a flowchart of another speech recognition method according to an embodiment of the present invention. Referring to FIG. 1, FIG. 4, and FIG. 5 at the same time, in this embodiment, the processing device 110 may execute the feature extraction module 410, the accent recognition module 420, the voice recognition module 430, and the probability comparison stored in the storage device 130. Module 440. Compared with the embodiments of FIG. 2 and FIG. 3, the processing device 110 of this embodiment may further include an execution feature extraction module 410 and an accent recognition module 420. In this embodiment, the feature extraction module 410 is configured to receive a voice signal provided by the input device 120 and analyze the voice signal to obtain a voice feature VC '. The accent recognition module 420 is used to analyze a voice signal to determine what type of accent the voice signal belongs to, and to select the voice feature VC 'according to the judgment result for calculation and processing by a corresponding acoustic dictionary.

In this embodiment, similar to the embodiments of FIG. 2 and FIG. 3, the speech recognition module 430 may include an acoustic model 431, multiple acoustic dictionaries 432A, 432B, 432C, a language model 433, and multiple decoders 434A, 434B. , 434C. In this embodiment, these acoustic dictionaries 432A, 432B, 432C correspond to a plurality of different accent types. In this embodiment, these decoders 434A, 434B, and 434C are used to generate string data and string probabilities that are more appropriate or have the greatest probability based on the acoustic model 431, the acoustic dictionary 432A, 432B, 432C, and the language model 433.

In this embodiment, the accent recognition module 420 is, for example, through a filter bank (Mel-Frequency Cipstal Coefficients, MFCC) to pass a Gaussian Mixture Model (GMM) or Deep Neural Networks (DNN) perform analysis and calculation to identify the type of accent to which the user-supplied speech signal belongs. Therefore, the voice recognition module 430 of this embodiment may select one of these acoustic dictionaries 432A, 432B, 432C for analysis and calculation according to the recognition result of the accent recognition module 420.

That is, in this embodiment, if the accent recognition module 420 can recognize that the accent type of the voice signal can correspond to one of these acoustic dictionaries 432A, 432B, 432C, the voice recognition module 430 receives the voice feature VC 'After that, according to the recognition result of the accent recognition module 420, one of these acoustic dictionaries 432A, 432B, and 432C may be analyzed and calculated to obtain a string data SD through one of the accent types corresponding to the speech signal. The probability comparison module 440 can directly output the string data SD 'to the output device 140.

However, in an embodiment, if the accent type that the accent recognition module 420 cannot recognize the voice signal may correspond to one of these acoustic dictionaries 432A, 432B, 432C, it means that these acoustic dictionaries 432A, 432B, 432C may not have Corresponds to the same accent type. Therefore, in one embodiment, the speech recognition module 430 can perform the speech recognition operation as described in the embodiments of FIG. 2 and FIG. 3 to generate multiple sets of string data and multiple sets of string probabilities. The probability comparison module 440 compares the string probabilities to select the string data with the closest accent type and the highest probability to the output device 140.

For example, if the user inputs the voice signal "bei jing (Beijing)" as a Beijing accent through the input device 120, after the accent recognition module 420, the voice recognition module 430 selects the acoustic dictionary 432B to process the voice feature VC '. In this example, the speech recognition module 430 analyzes the speech features VC 'through the acoustic model 431, the acoustic dictionary 432B, and the language model 433, so that the decoder 334B generates a string of data to the probability comparison module 440. The probability comparison module 440 directly outputs a string of data provided by the speech recognition module 430 without performing a probability comparison. That is, the speech recognition module 430 of this embodiment does not need to perform analysis and calculation through all the acoustic dictionaries 432A, 432B, and 432C, and can efficiently provide speech recognition results.

A possible implementation manner is described with reference to the flowchart of FIG. 5. In step S510, the feature extraction module 410 receives a voice signal and analyzes the voice signal to obtain a voice feature VC ′. In step S520, the accent recognition module 420 analyzes the speech signal to determine the accent type of the speech signal, and selects one of these acoustic dictionaries 432A, 432B, and 432C to analyze the speech feature VC '. In step S530, the voice recognition module 430 selectively outputs a voice feature VC 'to one of the acoustic dictionaries 432A, 432B, and 432C corresponding to the type of the accent according to the type of the accent of the speech signal, so as to correspond to the accent One of these acoustic dictionaries 432A, 432B, 432C, an acoustic model 431, and a language model 433 to analyze speech features. In step S540, the probability comparison module 440 outputs a string data SD 'provided by the analysis result of the speech recognition module 430.

In addition, for detailed implementation details and technical solutions of each model in the speech recognition module 430 of this embodiment, reference may be made to the embodiments of FIG. 2 and FIG. 3 to obtain sufficient teaching, suggestions, and implementation descriptions, and therefore will not be repeated.

In summary, the speech recognition device and speech recognition method of the present invention can be applied to the analysis of speech signals of different types of accents. The speech recognition device can build an acoustic model, a language model, and Multiple acoustic dictionaries corresponding to different accent types are used to analyze speech features, so multiple string probabilities and string data generated by the speech recognition module of the present invention may be comparable. In addition, the speech recognition device and speech recognition method of the present invention can further use an accent recognition module to determine what type of accent the speech signal provided by the user, so as to analyze the speech characteristics directly through the acoustic dictionary corresponding to the type of accent. The speech recognition module can provide speech recognition results more efficiently.

Although the present invention has been disclosed as above with the examples, it is not intended to limit the present invention. Any person with ordinary knowledge in the technical field can make some modifications and retouching without departing from the spirit and scope of the present invention. The protection scope of the present invention shall be determined by the scope of the attached patent application.

100‧‧‧speech recognition device

110‧‧‧Processing device

120‧‧‧ input device

130‧‧‧Storage device

140‧‧‧output device

330, 430‧‧‧speech recognition module

331, 431‧‧‧ acoustic models

332A, 332B, 332C, 432A, 432B, 432C‧‧‧ Acoustic Dictionary

333, 433‧‧‧ language models

334A, 334B, 334C, 434A, 434B, 434C‧‧‧ decoders

340‧‧‧Probability comparison module

410‧‧‧Feature Extraction Module

420‧‧‧Accent recognition module

440‧‧‧ Probability Comparison Module

S210, S220, S510, S520, S530, S540‧‧‧ steps

VC, VC’‧‧‧ Voice Features

SD, SD’‧‧‧ string data

FIG. 1 is a schematic diagram of a speech recognition device according to an embodiment of the present invention. FIG. 2 is a flowchart of a speech recognition method according to an embodiment of the present invention. FIG. 3 is a schematic diagram of a speech recognition device according to an embodiment of the invention. FIG. 4 is a schematic diagram of another speech recognition device according to an embodiment of the invention. FIG. 5 is a flowchart of another speech recognition method according to an embodiment of the present invention.

Claims (10)

A voice recognition device includes: a voice recognition module for receiving voice features, and the voice recognition module includes an acoustic model, a language model, and a plurality of acoustic dictionaries, wherein the voice recognition module uses the plurality of At least one of an acoustic dictionary, the acoustic model, and the language model to analyze the speech features to generate at least one string probability and at least one string data, wherein the plurality of acoustic dictionaries correspond to a plurality of different accents Type; and a probability comparison module coupled to the speech recognition module, and the probability comparison module judges a highest probability among the at least one string probability to output the at least one corresponding to the highest probability One of the string data. The speech recognition device according to item 1 of the scope of patent application, wherein the speech recognition module generates multiple string probabilities and multiple string data corresponding to different accent types by using the multiple acoustic dictionaries, where The phonetic features are of the same accent type as one of the plurality of acoustic dictionaries, so that the string probability generated via one of the plurality of acoustic dictionaries is the highest probability. The speech recognition device according to item 1 of the scope of patent application, further comprising: a feature extraction module coupled to the voice recognition module to receive a voice signal, and the feature extraction module analyzes the voice Signal to provide the voice feature to the voice recognition module. The speech recognition device according to item 3 of the scope of patent application, further comprising: an accent recognition module, coupled to the feature extraction module and the speech recognition module, and the accent recognition module is used to analyze all Speak the speech signal to determine the accent type of the speech signal, and select one of the plurality of acoustic dictionaries to analyze the speech feature. The speech recognition device according to item 4 of the scope of patent application, wherein the accent recognition module selectively outputs the speech features to the corresponding to the accent type according to the accent type of the speech signal. One of the plurality of acoustic dictionaries to analyze the speech feature by one of the plurality of acoustic dictionaries corresponding to the accent type, the acoustic model, and the language model, and outputting one Description of string data. A speech recognition method includes: receiving a speech feature and analyzing the speech feature by at least one of a plurality of acoustic dictionaries, an acoustic model, and a language model to generate at least one string probability and at least one string data, wherein The plurality of acoustic dictionaries correspond to a plurality of different accent types; and determining a highest probability among the at least one string probability to output one of the at least one string data corresponding to the highest probability. The speech recognition method according to item 6 of the scope of patent application, wherein the plurality of acoustic dictionaries generate a plurality of string probabilities and a plurality of string data corresponding to different accent types, and the speech feature and the plurality of One of the acoustic dictionaries is of the same accent type such that the string probability generated via one of the plurality of acoustic dictionaries is the highest probability. The speech recognition method according to item 6 of the patent application scope further comprises: receiving a speech signal, and analyzing the speech signal to obtain the speech characteristics. The speech recognition method according to item 8 of the scope of patent application, further comprising: analyzing the speech signal to determine the accent type of the speech signal, and selecting one of the plurality of acoustic dictionaries to analyze the speech feature. The speech recognition method according to item 9 of the scope of patent application, further comprising: selectively outputting the speech feature to the plurality of acoustic dictionaries corresponding to the accent type according to the accent type of the speech signal. One of the plurality of acoustic dictionaries corresponding to the type of the accent, the acoustic model, and the language model to analyze the speech feature; and outputting one of the string data .