Speech recognition methods and devices, electronic devices, chips

Abstract

This invention provides a speech recognition method, apparatus, electronic device, and chip. The speech recognition method includes: extracting speech features from the speech to be recognized; sending the speech features to an acoustic model to output a first-level phoneme sequence, wherein the acoustic model is trained using a complete pronunciation dictionary module, which records each character and its corresponding phoneme sequence; performing context transcription on the phonemes in the first-level phoneme sequence to obtain a second-level phoneme sequence; using a decoding mini-dictionary sub-module to model the pronunciation dictionary of the second-level phoneme sequence to obtain decoded characters, wherein the decoding mini-dictionary sub-module is obtained by replacing non-keyword characters in the complete pronunciation dictionary with specified characters; and sending the decoded characters to a language model to process the decoded characters and output the recognition result. This invention can reduce the size of the language model while maintaining the accuracy of the acoustic model.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a speech recognition method and apparatus, electronic device, and chip. Background Technology

[0002] In phoneme-level keyword speech recognition, training the acoustic model with a complete pronunciation dictionary improves its modeling accuracy, but it also leads to an excessively large language model size during decoding. For example, in speech recognition using GMM and HMM-DNN systems (Gaussian Mixture Model and Hidden Markov-Deep Neural Network system), the acoustic model built with a DNN models a complete phoneme set. The phoneme set in the complete pronunciation dictionary module consists of phonemes from keyword characters and numerous non-keyword character phoneme sequences. The language model is designed around the keywords and other characters in the complete pronunciation dictionary module. While this approach allows the acoustic model to clearly model the audio features corresponding to each phoneme, and after context transcription processing, the language model outputs the recognition result, the presence of a large number of non-keyword characters makes the language model excessively large.

[0003] The approach of mapping all non-keyword characters to "unknown" is employed. For example, in speech recognition using GMM and HMM-DNN systems, the acoustic model only models the phonemes corresponding to keywords and the "unknown" phoneme, and then the language model is designed around the keyword characters and "unknown". While this method can reduce the size of the language model, it results in the pronunciation features of many non-keyword characters being mapped to "unknown" phonemes. This not only increases the difficulty of modeling the pronunciation dictionary but also reduces the modeling accuracy of the acoustic model because the pronunciation differences between different characters are significant, requiring the fitting of numerous phonemes corresponding to non-keyword characters to a single "unknown" phoneme.

[0004] Therefore, it is urgent to solve the contradiction between the modeling accuracy of acoustic models and the size of language models. Summary of the Invention

[0005] To address the aforementioned issues, the speech recognition method, apparatus, electronic device, and chip provided by this invention train the acoustic model through a complete pronunciation dictionary module and unify the non-keyword characters in the decoding sub-dictionary module, thereby reducing the design of decoding paths for non-keyword characters in the language model and reducing the size of the language model while ensuring the accuracy of the acoustic model.

[0006] In a first aspect, the present invention provides a speech recognition method, comprising:

[0007] Extract speech features from the speech to be recognized;

[0008] The speech features are sent to the acoustic model so that the acoustic module outputs a first-level phoneme sequence through the acoustic model; the acoustic model is trained by a complete pronunciation dictionary module, which records each character and the phoneme sequence corresponding to each character;

[0009] Context transcription is performed on the phonemes in the primary phoneme sequence to obtain the secondary phoneme sequence;

[0010] The method further includes: using a decoding mini-dictionary submodule to perform pronunciation dictionary modeling on the secondary phoneme sequence to obtain decoded characters, wherein the decoding mini-dictionary submodule is obtained by the complete pronunciation dictionary module by replacing non-keyword characters in the complete pronunciation dictionary with specified characters, wherein the specified characters are a fixed character different from the keyword characters, and the decoded characters are the specified characters and / or keyword characters;

[0011] The decoded characters are sent to the language model so that the language model processes the decoded characters and outputs the recognition result.

[0012] Optionally, the step of using the decoding mini-dictionary submodule to perform pronunciation dictionary modeling on the secondary phoneme sequence includes:

[0013] Based on the secondary phoneme sequence, query the phoneme sequence recorded in the decoding dictionary submodule that matches the secondary phoneme sequence, and determine the character corresponding to the phoneme sequence that matches the secondary phoneme sequence in the decoding dictionary submodule as the decoded character.

[0014] In a second aspect, the present invention provides a voice recognition device, comprising:

[0015] The extraction module is used to extract speech features from the speech to be recognized;

[0016] An acoustic module is used to output a first-level phoneme sequence based on the speech features; the acoustic module includes an acoustic model, which is trained by a complete pronunciation dictionary module. The acoustic model is used to model the speech features, and the complete pronunciation dictionary module is used to record the phoneme sequence corresponding to each character.

[0017] The phoneme processing module is used to perform contextual transcription on the phonemes in the primary phoneme sequence to obtain the secondary phoneme sequence;

[0018] The speech recognition device further includes: a pronunciation dictionary conversion module, used to perform pronunciation dictionary modeling on the secondary phoneme sequence using a decoding mini-dictionary sub-module to obtain decoded characters, wherein the decoding mini-dictionary sub-module is obtained by the complete pronunciation dictionary module by replacing non-keyword characters in the complete pronunciation dictionary with specified characters, wherein the specified characters are a fixed character different from the keyword characters, and the decoded characters are the specified characters and / or keyword characters;

[0019] A language model is used to process the decoded characters and output the recognition results;

[0020] The output of the extraction module is connected to the input of the acoustic module, the output of the acoustic module is connected to the input of the phoneme processing module, the output of the phoneme processing module is connected to the input of the pronunciation dictionary conversion module, and the output of the pronunciation dictionary conversion module is connected to the input of the language model.

[0021] Optionally, the pronunciation dictionary conversion module includes: a decoding submodule and a decoding mini-dictionary submodule;

[0022] The decoding submodule is used to query the phoneme sequence recorded by the decoding dictionary submodule that matches the secondary phoneme sequence, and to determine the character in the decoding dictionary submodule that corresponds to the secondary phoneme sequence as the decoding character;

[0023] The decoding dictionary submodule is connected to the decoding submodule. The input of the decoding submodule is connected to the output of the phoneme processing module, and the output of the decoding submodule is connected to the input of the language model.

[0024] Thirdly, the present invention provides an electronic device, the electronic device comprising:

[0025] At least one processor; and

[0026] A memory that is communicatively connected to at least one processor; wherein,

[0027] The memory stores instructions that can be executed by at least one processor, such that the at least one processor is able to perform any of the methods described above.

[0028] Fourthly, the present invention provides a chip, the chip comprising:

[0029] At least one processor; and

[0030] A memory that is communicatively connected to at least one processor; wherein,

[0031] The memory stores instructions that can be executed by at least one processor, such that the at least one processor is able to perform any of the methods described above.

[0032] Fifthly, the present invention provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method described above.

[0033] The speech recognition method, device, electronic device, and chip provided in this invention train the acoustic model through a complete pronunciation dictionary module and unify the non-keyword characters in the decoding small dictionary sub-module, reducing the design of decoding paths for non-keyword characters in the language model, thereby reducing the size of the language model while ensuring the modeling accuracy of the acoustic model. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic flowchart illustrating a speech recognition method according to an embodiment of this application;

[0036] Figure 2 This is a schematic structural diagram of a speech recognition system according to an embodiment of this application;

[0037] Figure 3 This is a schematic flowchart illustrating a speech recognition method according to an embodiment of this application.

[0038] Figure Labels

[0039] 200. Speech recognition device; 210. Extraction module; 220. Acoustic model; 230. Phoneme processing module; 231. Hidden Markov submodule; 232. Context transcription submodule; 240. Pronunciation dictionary conversion module; 241. Decoding submodule; 242. Decoding mini-dictionary submodule; 250. Language model. Detailed Implementation

[0040] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0041] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0042] Spatial relation terms such as “below,” “under,” “below,” “under,” “above,” “above,” etc., are used herein to describe the relationship between one element or feature shown in the figure and other elements or features. It should be understood that, in addition to the orientation shown in the figure, spatial relation terms also include different orientations of the device in use and operation. For example, if the device in the figure is flipped, the element or feature described as “below,” “under,” or “below” will be oriented “above” the other element or feature. Therefore, the exemplary terms “below” and “under” can include both above and below orientations. Furthermore, the device may also include other orientations (e.g., rotated 90 degrees or other orientations), and the spatial descriptive terms used herein will be interpreted accordingly.

[0043] It should be noted that when an element is referred to as "fixedly connected" to another element, it can be directly on the other element or there may be an intervening element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intervening element. Conversely, when an element is referred to as being "directly on" another element, there is no intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0044] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising,” “including,” or “having,” etc., specify the presence of the stated feature, whole, step, operation, component, part, or combination thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof.

[0045] Firstly, this embodiment provides a speech recognition method, see [link to relevant documentation]. Figure 1 The method includes steps S101 to S105:

[0046] Step S101: Extract speech features from the speech to be recognized.

[0047] Specifically, when extracting speech features, the speech signal formed by the speech to be recognized is first processed by DC removal, noise reduction and other processes; then, MFCC (Mel frequency cepstral coefficients), Fbank (Mel band energy) and LPC (linear prediction coefficients) are extracted as speech features as input to the acoustic model.

[0048] Step S102: Send speech features to the acoustic model so that the acoustic module can obtain the first-level phoneme sequence through the acoustic model.

[0049] The acoustic model is trained using a complete pronunciation dictionary module, a large number of speech tags, and a large amount of corpus. The complete pronunciation dictionary module is a standard file that records each character and the phoneme sequence corresponding to each character. Speech tags refer to the character sequence corresponding to a speech segment. The character sequence is composed of multiple characters, including keyword characters and non-keyword characters.

[0050] In this embodiment, the characters recorded by the complete pronunciation dictionary module are the commonly used characters corresponding to the respective language systems. ， Rarely occurring characters may be excluded from this comprehensive pronunciation dictionary module, but are not limited to this; the acoustic model is a neural network model. Furthermore, the speech features sent to the acoustic model are arranged temporally, and the acoustic model outputs multiple results for each set of input speech features depending on the classification. Because there are numerous phonemes, each phoneme also carries a different probability.

[0051] It should be noted that step S102 further includes: after the speech features are processed by the acoustic model and then modeled by the Hidden Markov Model (HMM) submodule, a first-level phoneme sequence is obtained. Specifically, the acoustic model provides multiple observation probabilities with HMM after processing the speech features; the HMM submodule performs HMM modeling on the speech features based on the HMM observation probabilities to output the first-level phoneme sequence.

[0052] Step S103: Context transcription is performed on the phonemes in the primary phoneme sequence to obtain the secondary phoneme sequence.

[0053] Step S104: Use the decoding small dictionary sub-module to perform pronunciation dictionary modeling on the secondary phoneme sequence to obtain decoded characters. Among them, the decoding small dictionary sub-module is obtained by converting the complete pronunciation dictionary module by replacing non-keyword characters in the complete pronunciation dictionary with specified characters, and the specified character is a fixed character different from the keyword character. It should be noted that both the keyword character and the non-keyword character are set manually according to the application scenario, and no specific limitation is made in this embodiment; the characters in the complete pronunciation dictionary module either belong to the keyword character or belong to the non-keyword character. And for the characters in the decoding small dictionary sub-module, the characters that are not replaced with the specified character belong to the keyword character, and the remaining non-keyword characters are all replaced with the same specified character.

[0054] The decoded character is a specified character and / or a keyword character, that is, in the decoding small dictionary sub-module, a phoneme sequence may correspond to at least one keyword character, may also correspond to at least one keyword character and at least one specified character, or may also correspond to at least one specified character.

[0055] Specifically, in the complete pronunciation dictionary module, a phoneme sequence does not only correspond to one character. For example, the characters corresponding to "da" are "打", "大", "搭", etc. Similarly, in the decoding small dictionary sub-module, the characters corresponding to "d a" are "打" and "非", where it is limited that only "打" among the characters corresponding to "d a" is the keyword character. In this way, the decoded characters corresponding to "d a" are "打" and "非". If some phoneme sequences only correspond to keyword characters in the decoding small dictionary, the obtained decoded characters only contain keyword characters. If some phoneme sequences only correspond to specified characters in the decoding small dictionary, the obtained decoded characters only contain specified characters. In addition, there are multiple characters corresponding to a segment of speech features, so in the process of decoding through the decoding small dictionary sub-module, a decoded character may contain multiple characters at the same time, such as "打非", "打非开", or "打开", etc.

[0056] In this embodiment, it is assumed that a phoneme sequence corresponds to one character for illustration. Further, the step of using the decoding small dictionary sub-module to perform pronunciation dictionary modeling on the secondary phoneme sequence includes: according to the secondary phoneme sequence, query the phoneme sequence recorded in the decoding small dictionary sub-module that matches the secondary phoneme sequence, and determine the character corresponding to the phoneme sequence that matches the secondary phoneme sequence in the decoding small dictionary sub-module as the decoded character.

[0057] For example, in the complete pronunciation dictionary module, the character corresponding to the phoneme sequence "q ing" is "请"; the character corresponding to the phoneme sequence "d a" is "打"; the character corresponding to the phoneme sequence "k ai" is "开"; the character corresponding to the phoneme sequence "zh i" is "智"; the character corresponding to the phoneme sequence "n eng" is "能"; the character corresponding to the phoneme sequence "y in" is "音"; the character corresponding to the phoneme sequence "x i ang" is "箱". Among them, the characters "请", "智", and "能" are defined as non-keyword characters in the complete pronunciation dictionary module. In this embodiment, the specified character is "非".

[0058] Then, in the decoding small dictionary sub-module, the character corresponding to the phoneme sequence "q ing" is "非", the character corresponding to the phoneme sequence "d a" is "打"; the character corresponding to the phoneme sequence "k ai" is "开"; the character corresponding to the phoneme sequence "zh i" is "非"; the character corresponding to the phoneme sequence "n eng" is "非"; the character corresponding to the phoneme sequence "y in" is "音"; the character corresponding to the phoneme sequence "x i ang" is "箱". In this way, the phoneme sequence matching the secondary phoneme sequence recorded in the decoding small dictionary sub-module can be queried, and the character corresponding to the matching phoneme sequence can be determined as the decoding character.

[0059] For multiple secondary phoneme sequences - "q ing", "d a", "k ai", "zh i", "n eng", "y in", and "xi ang", after modeling through the pronunciation dictionary, the corresponding decoding characters - "非", "打", "开", "非", "非", "音", and "箱" can be obtained.

[0060] Step S105: Send the decoding character to the language model so that the language model processes the decoding character and outputs an identification result.

[0061] Specifically, if the decoding characters sent to the speech model are "非", "打", "开", "非", "非", "音", and "箱" respectively, then after relevant steps such as collocation and probability statistics, the speech model outputs "非打开非非音箱" as the identification result.

[0062] In an optional embodiment, during the decoding process, the operation steps before executing step S104 can be implemented through speech recognition of the GMM and HMM-DNN systems, and this embodiment will not elaborate on this too much.

[0063] When training the acoustic model for this speech recognition method, corresponding speech labels for the training corpus are used. When constructing the pronunciation dictionary, the complete pronunciation dictionary module is made to perfectly contain each character in the speech labels, which is the large dictionary. This allows each piece of speech to have its own relatively independent phoneme sequence and label sequence. In this way, when training, the acoustic model can better fit the features of each piece of speech and can also more clearly model each phoneme, ensuring the modeling accuracy of the acoustic model. At the same time, in the decoding stage, this speech recognition method does not directly use the complete pronunciation dictionary module constructed during training. Instead, all non-keyword characters in the complete pronunciation dictionary module are replaced with the same specified character, such as "non", and the complete pronunciation dictionary module is modified to obtain the decoding small dictionary sub-module. At this time, the character types in the decoding small dictionary sub-module become very few, which is the small dictionary. After the decoding process is modeled with the small dictionary, there are only keyword characters and the specified character "non", but the numerous "non"s in the small dictionary still each correspond to an independent phoneme sequence. From the perspective of the entire speech recognition method, it improves the modeling accuracy of the acoustic model to a certain extent. In addition, when designing the language model in the decoding stage, only the decoding paths for keywords and the specified character "non" are designed. In this way, the language model originally needed to design decoding paths for numerous characters, but now only needs to design the decoding paths for very few characters, that is, keyword characters and the specified character, thus greatly reducing the volume of the language model.

[0064] The following illustrates the process of speech feature processing through acoustic modeling, Hidden Markov Modeling, contextual transcription, pronunciation dictionary modeling, and language modeling: Assume a set of audio features is input to the acoustic module every 10ms. Thus, from the first set of audio features, the acoustic module outputs phonemes: A1, B1, C1, D1, where A1, B1, C1, and D1 are in a parallel relationship, and each phoneme corresponds to a probability value. Similarly, from the second set of audio features, the acoustic module outputs phonemes: A2, B2, C2, D2; from the third set of audio features, the acoustic module outputs phonemes: A3, B3, C3, D3. After 30ms, 64 first-level phoneme sequences are obtained, such as A1, A2, A3; A1, A2, B3; A1, B2, C3, and B1, A2, B3, etc. Subsequently, during context transcription, the probabilities of the 64 primary phoneme sequences are superimposed. For example, if A2 has a higher probability of appearing after A1, the weight of A2 in A1, A2, A3, and A1, A2, B3 is further increased; while the probability of B3 appearing after A2 is low, the weight of B3 in the A1, A2, B3 sequence is decreased. Based on the weights of each sequence, the sequence with the highest total weight is selected as the secondary phoneme sequence, such as A1, A2, A3. Then, a pronunciation dictionary model is performed on A1, A2, A3. If A1, A2, A3 are used to traverse the decoding dictionary submodule and a corresponding phoneme sequence is found, the character corresponding to that phoneme sequence is output as the decoded character. After receiving the decoded characters output from the decoding dictionary submodule, the language model combines them with the decoded characters output from previous rounds of time-series processing and selects the combination with the highest probability as the speech recognition result, which is then output. The probabilities referenced by the speech model include: the observation probabilities output by the acoustic model, the transition probabilities given in Hidden Markov Modeling, the context probabilities given during context transcription, the pronunciation dictionary probabilities given during pronunciation dictionary modeling, and the language model probabilities given during language model processing. The context probabilities, pronunciation dictionary probabilities, and language model probabilities given during language model processing are all prior probabilities and will not be elaborated upon further here.

[0065] In addition, it should be noted that the acoustic model, hidden Markov modeling, context transcription and language model processing involved in this invention are all existing technologies, and the specific processing procedures will not be described in detail in this embodiment.

[0066] Secondly, this embodiment provides a voice recognition device 200, see [link to documentation]. Figure 2 The voice recognition device 200 includes:

[0067] Extraction module 210 is used to extract speech features from the speech to be recognized;

[0068] The acoustic module 220 is used to output multiple sets of first-level phoneme sequences based on speech features. The acoustic module 220 includes an acoustic model, which is trained by a complete pronunciation dictionary module. The acoustic model is used to model speech features to provide HMM observation probabilities. The complete pronunciation dictionary module is used to record the phoneme sequences corresponding to each character.

[0069] The phoneme processing module 230 is used to perform contextual transcription on the phonemes in the primary phoneme sequence to obtain multiple sets of secondary phoneme sequences.

[0070] The pronunciation dictionary conversion module 240 is used to perform pronunciation dictionary modeling on multiple sets of secondary phoneme sequences. In the pronunciation dictionary modeling, non-keyword characters in the primary character sequences corresponding to the secondary phoneme sequences are replaced with unique specified characters that are different from the keyword characters. After the pronunciation dictionary modeling, the secondary character sequences are obtained.

[0071] Language Model 250 is used to process secondary character sequences and output recognition results.

[0072] The output of the extraction module 210 is connected to the input of the acoustic module 220, the output of the acoustic module 220 is connected to the input of the phoneme processing module 230, the output of the phoneme processing module 230 is connected to the input of the pronunciation dictionary conversion module 240, and the output of the pronunciation dictionary conversion module 240 is connected to the input of the language model 250.

[0073] In an optional embodiment, the acoustic module 220 further includes a hidden Markov submodule;

[0074] The Hidden Markov Submodule is used to perform Hidden Markov Modeling on the speech features based on the HMM observation probabilities output by the acoustic model, in order to obtain the first-level phoneme sequence.

[0075] The hidden Markov submodule is connected to the acoustic model.

[0076] In one optional embodiment, the pronunciation dictionary conversion module 240 includes: a decoding submodule 241 and a decoding mini-dictionary submodule 242;

[0077] The decoding submodule 241 is used to query the phoneme sequence recorded by the decoding dictionary submodule 242 that matches the second phoneme sequence, and to determine the character corresponding to the matching phoneme sequence as the decoded character;

[0078] The decoding sub-module 242 is connected to the decoding sub-module 241. The input of the decoding sub-module 241 is connected to the output of the phoneme processing module 230, and the output of the decoding sub-module 241 is connected to the input of the language model 250.

[0079] Thirdly, this embodiment provides a speech recognition method, see [link to relevant documentation]. Figure 3 Combining the speech recognition method of the first aspect and the speech recognition device 200 of the second aspect, the speech recognition method in the third aspect includes steps S301 to S305:

[0080] Step S301: Extract speech features from the speech to be recognized using the extraction module 210.

[0081] Step S302: Send the speech features to the acoustic module 220 to obtain the first-level phoneme sequence through the acoustic module 220.

[0082] Step S303: Send the primary phoneme sequence to the phoneme processing module 230 so that the phoneme processing module 230 outputs the secondary phoneme sequence.

[0083] Step S304: Send the secondary phoneme sequence to the pronunciation dictionary conversion module 240, so that the decoding submodule 241 replaces the non-keyword characters in the primary character sequence corresponding to the secondary phoneme sequence with a unique specified character different from the keyword character according to the decoding mini-dictionary submodule 242, and obtains the decoded character.

[0084] Step S305: Send the decoded characters to the language model 250 so that the language model 250 processes the decoded characters and outputs the recognition results.

[0085] Fourthly, this embodiment provides an electronic device, which includes:

[0086] At least one processor; and

[0087] A memory that is communicatively connected to at least one processor; wherein,

[0088] The memory stores instructions that can be executed by at least one processor, such that the at least one processor is able to perform any of the methods described above.

[0089] Fifthly, this embodiment provides a chip, the chip comprising:

[0090] At least one processor; and

[0091] A memory that is communicatively connected to at least one processor; wherein,

[0092] The memory stores instructions that can be executed by at least one processor, such that the at least one processor is able to perform any of the methods described above.

[0093] Sixthly, this embodiment provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement any of the methods described above.

[0094] In the description of this specification, the references to terms such as "some embodiments," "other embodiments," "ideal embodiments," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example that are included in at least one embodiment or example of this application. In this specification, the illustrative descriptions of the above terms do not necessarily refer to the same embodiments or examples.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A speech recognition method, comprising: Extract speech features from the speech to be recognized; The speech features are sent to the acoustic model so that the acoustic module obtains a first-level phoneme sequence through the acoustic model; Context transcription is performed on the phonemes in the primary phoneme sequence to obtain the secondary phoneme sequence; The acoustic model is characterized in that it is trained by a complete pronunciation dictionary module, which records each character and the phoneme sequence corresponding to each character. The method further includes: using a decoding mini-dictionary submodule to perform pronunciation dictionary modeling on the secondary phoneme sequence to obtain decoded characters, wherein the decoding mini-dictionary submodule is obtained by the complete pronunciation dictionary module by replacing non-keyword characters in the complete pronunciation dictionary with specified characters, wherein the specified characters are a fixed character different from the keyword characters, and the decoded characters are the specified characters and / or keyword characters; The step of using the decoding mini-dictionary submodule to perform pronunciation dictionary modeling on the secondary phoneme sequence includes: Based on the secondary phoneme sequence, query the phoneme sequence that matches the secondary phoneme sequence recorded in the decoding dictionary submodule, and determine the character corresponding to the phoneme sequence that matches the secondary phoneme sequence in the decoding dictionary submodule as the decoded character; The decoded characters are sent to the language model so that the language model processes the decoded characters and outputs the recognition result.

2. A voice recognition device, comprising: The extraction module is used to extract speech features from the speech to be recognized; The acoustic module is used to output a first-level phoneme sequence based on the speech features; The phoneme processing module is used to perform contextual transcription on the phonemes in the primary phoneme sequence to obtain the secondary phoneme sequence; The acoustic module is characterized in that it includes an acoustic model, which is trained by a complete pronunciation dictionary module. The acoustic model is used to model speech features, and the complete pronunciation dictionary module is used to record the phoneme sequence corresponding to each character. The speech recognition device further includes: a pronunciation dictionary conversion module, used to perform pronunciation dictionary modeling on the secondary phoneme sequence using a decoding mini-dictionary sub-module to obtain decoded characters, wherein the decoding mini-dictionary sub-module is obtained by the complete pronunciation dictionary module by replacing non-keyword characters in the complete pronunciation dictionary with specified characters, wherein the specified characters are a fixed character different from the keyword characters, and the decoded characters are the specified characters and / or keyword characters; A language model is used to process the decoded characters and output the recognition results; The output of the extraction module is connected to the input of the acoustic module, the output of the acoustic module is connected to the input of the phoneme processing module, the output of the phoneme processing module is connected to the input of the pronunciation dictionary conversion module, and the output of the pronunciation dictionary conversion module is connected to the input of the language model. The pronunciation dictionary module includes: a decoding submodule and a decoding mini-dictionary submodule; The decoding submodule is used to query the phoneme sequence recorded by the decoding dictionary submodule that matches the second phoneme sequence, and to determine the character corresponding to the matching phoneme sequence as the decoding character; The decoding dictionary submodule is connected to the decoding submodule, the input of the decoding submodule is connected to the output of the phoneme processing module, and the output of the decoding submodule is connected to the input of the language model.

3. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of claim 1.

4. A chip, characterized in that, The chip includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of claim 1.

5. A computer-readable storage medium, wherein, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method as described in claim 1.

Images