A voice recognition method, device, equipment and medium for specifying a scene

Abstract

The application provides a voice recognition method, device and equipment for a specified scene and a medium. The method comprises the following steps: obtaining a plurality of word segmentation feature sequences by performing word segmentation processing on the voice features of the audio to be processed, and performing feature recognition on each word segmentation feature sequence according to a trained voice recognition model to obtain a plurality of feature recognition results of each word segmentation feature sequence. Further, the optimal word segmentation result is selected from the plurality of feature recognition results corresponding to each word segmentation feature sequence based on a preset specified scene-specific vocabulary, and the backup word segmentation result is determined according to the confidence of each feature recognition result. The above process adjusts the optimal word segmentation result of each word segmentation feature sequence through the specified scene-specific vocabulary, and determines the voice recognition result of the audio to be processed according to the optimal word segmentation result and the backup word segmentation result, so that the model has the ability to recognize the specified scene-specific vocabulary.

Description

Technical Field

[0001] This application relates to the field of speech recognition technology, specifically to a speech recognition method, apparatus, device, and medium for a specified scenario. Background Technology

[0002] Speech recognition is a technology encompassing acoustics, linguistics, mathematics, and statistics, and is a crucial component of human-computer natural interaction technology. In practical applications, it requires collecting a large amount of speech data and annotating it with corresponding text to construct training samples for the model. These training samples are then used to train the speech recognition model, enabling it to convert speech data into corresponding text.

[0003] However, corpora from specific scenarios such as healthcare and finance are often highly private and difficult to obtain. This results in a scarcity of training samples for these scenarios. Related technologies often use only a small number of samples for training, or employ traditional methods such as cropping, rotation, pitch adjustment, and noise addition to augment the samples before training. Speech recognition models trained using these methods often suffer from underfitting and have poor recognition capabilities for scenario-specific vocabulary. Summary of the Invention

[0004] This application provides a speech recognition method, apparatus, device, and medium for a specified scenario, which is used to enable the speech recognition model to recognize the specific vocabulary of the specified scenario based on the output of a preset specified scenario-specific vocabulary restriction model.

[0005] To achieve the above objectives, the technical solution of this application embodiment is implemented as follows:

[0006] In a first aspect, embodiments of this application provide a speech recognition method for a specified scenario, the method comprising:

[0007] Acquire the audio to be processed and determine the speech features of the audio to be processed;

[0008] The speech features are segmented into words to obtain a segmentation feature sequence; and feature recognition is performed on each segmentation feature sequence to obtain multiple feature recognition results for each segmentation feature sequence and the confidence level of each feature recognition result.

[0009] For each segmentation feature sequence, the optimal segmentation result of the segmentation feature sequence is determined based on the preset specific vocabulary for the specified scenario and the recognition results of each feature; and a first number of candidate segmentation results are determined from each feature recognition result according to the confidence level.

[0010] The speech recognition result of the audio to be processed is determined based on the optimal segmentation result and the alternative segmentation result of each segmentation feature sequence.

[0011] In some possible embodiments, obtaining the speech features of the audio to be processed includes:

[0012] The audio to be processed is digitized to obtain a processing array representing the audio features;

[0013] The likelihood probability of the audio to be processed is determined based on the array to be processed, and the speech features in the audio to be processed are determined based on the likelihood probability.

[0014] In some possible embodiments, determining the likelihood probability of the audio to be processed based on the array to be processed includes:

[0015] The first expectation representing the speech features and the second expectation representing the noise features of the array to be processed are determined by the boundary probability algorithm.

[0016] The likelihood probability is determined based on the first expectation and the second expectation.

[0017] In some possible embodiments, determining the optimal word segmentation result of the word segmentation feature sequence based on preset, scenario-specific vocabulary and the recognition results of each feature includes:

[0018] Detect whether any of the feature recognition results contain a recognition result that is identical to any of the specified scene-specific terms;

[0019] If it exists, the identification result with the highest confidence level is taken as the optimal word segmentation result; otherwise, the feature identification result with the highest confidence level is taken as the optimal word segmentation result.

[0020] In some possible embodiments, before selecting the identification result with the highest confidence level as the optimal word segmentation result, the method further includes:

[0021] The confidence level of the identification result with the highest confidence level is determined to be greater than the confidence threshold.

[0022] The method further includes:

[0023] If the confidence level of the highest confidence-level identification result is not greater than the confidence level threshold, then the feature identification result with the highest confidence level is taken as the optimal word segmentation result.

[0024] In some possible embodiments, determining the speech recognition result of the speech features based on the optimal and alternative segmentation results of each of the segmentation feature sequences includes:

[0025] Based on the order of the segmentation feature sequences in the speech features, the optimal segmentation results are concatenated to obtain a target concatenated feature containing all optimal segmentation results; and,

[0026] The candidate word segmentation results are concatenated according to the sorting to obtain the candidate concatenation features; wherein the candidate concatenation features include the candidate word segmentation results corresponding to each word segmentation feature sequence.

[0027] The speech recognition result is determined based on the target splicing features and the candidate splicing features.

[0028] In some possible embodiments, determining the speech recognition result based on the target splicing feature and the candidate splicing features includes:

[0029] Based on the confidence level of each candidate word segmentation result contained in each candidate splicing feature, a second number of target splicing features are determined from each candidate splicing feature;

[0030] If the feature similarity between the target splicing feature and any other target splicing feature is greater than a similarity threshold, then the speech recognition result is determined based on the target splicing feature.

[0031] Otherwise, the comprehensive confidence of each target splicing feature is determined based on the confidence of each candidate word segmentation result, and the speech recognition result is determined based on the target splicing feature with the highest comprehensive confidence.

[0032] Secondly, embodiments of this application provide a speech recognition device for a specified scenario, the device comprising:

[0033] The feature acquisition module is configured to acquire the audio to be processed and determine the speech features of the audio to be processed.

[0034] The feature recognition module is configured to perform word segmentation processing on the speech features to obtain word segmentation feature sequences of the speech features; and to perform feature recognition on each word segmentation feature sequence to obtain multiple feature recognition results for each word segmentation feature sequence and the confidence level of each feature recognition result;

[0035] The word segmentation processing module is configured to perform the following for each word segmentation feature sequence: determine the optimal word segmentation result of the word segmentation feature sequence based on a preset specified scenario-specific vocabulary and the feature recognition results; and determine a first number of candidate word segmentation results from each feature recognition result according to the confidence level.

[0036] The speech recognition module is configured to determine the speech recognition result of the audio to be processed based on the optimal segmentation result and alternative segmentation results of each of the segmentation feature sequences.

[0037] In some possible embodiments, the acquisition of speech features of the audio to be processed is performed, and the feature acquisition module is configured to:

[0038] The audio to be processed is digitized to obtain a processing array representing the audio features;

[0039] The likelihood probability of the audio to be processed is determined based on the array to be processed, and the speech features in the audio to be processed are determined based on the likelihood probability.

[0040] In some possible embodiments, the feature acquisition module is configured to perform the determination of the likelihood probability of the audio to be processed based on the array to be processed, wherein:

[0041] The first expectation representing the speech features and the second expectation representing the noise features of the array to be processed are determined by the boundary probability algorithm.

[0042] The likelihood probability is determined based on the first expectation and the second expectation.

[0043] In some possible embodiments, the optimal word segmentation result of the word segmentation feature sequence is determined based on a preset, specified scenario-specific vocabulary and the recognition results of each feature. The word segmentation processing module is configured as follows:

[0044] Detect whether any of the feature recognition results contain a recognition result that is identical to any of the specified scene-specific terms;

[0045] If it exists, the identification result with the highest confidence level is taken as the optimal word segmentation result; otherwise, the feature identification result with the highest confidence level is taken as the optimal word segmentation result.

[0046] In some possible embodiments, before performing the step of taking the identification result with the highest confidence as the optimal word segmentation result, the word segmentation processing module is further configured to:

[0047] The confidence level of the identification result with the highest confidence level is determined to be greater than the confidence threshold.

[0048] The word segmentation processing module is also configured to:

[0049] If the confidence level of the highest confidence-level identification result is not greater than the confidence level threshold, then the feature identification result with the highest confidence level is taken as the optimal word segmentation result.

[0050] In some possible embodiments, the speech recognition module is configured to perform the speech recognition result determined based on the optimal and alternative segmentation results of each of the segmentation feature sequences, wherein the speech recognition module is configured to:

[0051] Based on the order of the segmentation feature sequences in the speech features, the optimal segmentation results are concatenated to obtain a target concatenated feature containing all optimal segmentation results; and,

[0052] The candidate word segmentation results are concatenated according to the sorting to obtain the candidate concatenation features; wherein the candidate concatenation features include the candidate word segmentation results corresponding to each word segmentation feature sequence.

[0053] The speech recognition result is determined based on the target splicing features and the candidate splicing features.

[0054] In some possible embodiments, the speech recognition module is configured to determine the speech recognition result based on the target splicing feature and the candidate splicing features, wherein the speech recognition module is configured to:

[0055] Based on the confidence level of each candidate word segmentation result contained in each candidate splicing feature, a second number of target splicing features are determined from each candidate splicing feature;

[0056] If the feature similarity between the target splicing feature and any other target splicing feature is greater than a similarity threshold, then the speech recognition result is determined based on the target splicing feature.

[0057] Otherwise, the comprehensive confidence of each target splicing feature is determined based on the confidence of each candidate word segmentation result, and the speech recognition result is determined based on the target splicing feature with the highest comprehensive confidence.

[0058] Thirdly, embodiments of this application provide an electronic device, including:

[0059] Memory, used to store program instructions;

[0060] A processor is configured to invoke program instructions stored in the memory and execute the steps of the method described in any one of the first aspects according to the obtained program instructions.

[0061] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the method described in any one of the first aspects.

[0062] Fifthly, embodiments of this application provide a computer program product comprising: computer program code, which, when executed on a computer, causes the computer to perform the method described in any of the first aspects.

[0063] In this embodiment, multiple segmentation feature sequences are obtained by segmenting the speech features of the audio to be processed. Then, feature recognition is performed on each segmentation feature sequence using a trained speech recognition model, resulting in multiple feature recognition results for each segmentation feature sequence. Further, based on a preset vocabulary specific to a designated scenario, the optimal segmentation result is selected from the multiple feature recognition results corresponding to each segmentation feature sequence, and candidate segmentation results are determined based on the confidence level of each feature recognition result. This process adjusts the optimal segmentation result of each segmentation feature sequence using a vocabulary specific to a designated scenario, and determines the speech recognition result of the audio to be processed based on the optimal segmentation result and the candidate segmentation results, enabling the model to recognize vocabulary specific to a designated scenario.

[0064] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing this disclosure. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0065] Figure 1 A schematic diagram of a voice wake-up device provided in an embodiment of this application;

[0066] Figure 2 A flowchart illustrating the overall process of a speech recognition method for a specified scenario provided in this application embodiment;

[0067] Figure 3 This is a schematic diagram of speech feature extraction provided in an embodiment of this application;

[0068] Figure 4 This is a schematic diagram illustrating the expansion of the likelihood probability formula provided in an embodiment of this application;

[0069] Figure 5 This is a schematic diagram of feature recognition provided for an embodiment of this application;

[0070] Figure 6 A schematic diagram of the clustering algorithm provided in the embodiments of this application;

[0071] Figure 7 A flowchart for obtaining the optimal word segmentation result provided in this application embodiment;

[0072] Figure 8 A schematic diagram illustrating the process of obtaining the optimal word segmentation result as provided in an embodiment of this application;

[0073] Figure 9 Another flowchart for obtaining the optimal word segmentation result provided in this application embodiment;

[0074] Figure 10 This application provides a schematic diagram illustrating the process of obtaining alternative word segmentation results in its embodiments;

[0075] Figure 11 This is a schematic diagram of the target splicing features and candidate splicing features provided in the embodiments of this application;

[0076] Figure 12 A schematic diagram illustrating the final speech recognition result output by the model provided in this embodiment of the application;

[0077] Figure 13 A structural diagram of a speech recognition device 1300 for a specified scenario provided in an embodiment of this application;

[0078] Figure 14 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0080] The terms "first" and "second" in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the term "comprising" and any variations thereof are intended to cover non-exclusive protection. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses. The term "multiple" in this application can mean at least two, for example, two, three, or more, and is not limited by the embodiments of this application.

[0081] In practical applications, such as device wake-up and intelligent voice interaction, speech recognition technology is involved. Specifically, it identifies whether the user's voice contains specified words to achieve human-computer voice interaction tasks. For example... Figure 1In the illustrated application scenario, users can wake up the smart device by speaking to it and interact with it via voice. In this process, the smart device receives the user's voice message and provides corresponding feedback. The core of this functionality lies in the smart device's ability to recognize the user's voice content.

[0082] Currently, many speech recognition models face the challenge of obtaining specific contextual corpora during training. For example, medical terms like rasagiline mesylate, metformin, and monosialotetrahexosylganglioside sodium are rarely used in other fields, making it difficult for traditional dictionary-based training methods to recognize such terms. Medical-related corpora are primarily sourced from Electronic Health Records (EHRs) and Electronic Medical Records (EMRs). These data are not open-source, are typically highly confidential, and difficult to access. Current techniques often train models using only a small number of samples or expand the samples using traditional methods such as cropping, rotation, pitch adjustment, and noise addition. However, speech recognition models trained in these ways often suffer from underfitting and exhibit poor vocabulary recognition capabilities for specific contexts.

[0083] To address the aforementioned issues, the inventive concept of this application is as follows: Multiple segmentation feature sequences are obtained by segmenting the speech features of the audio to be processed. Then, feature recognition is performed on each segmentation feature sequence using a trained speech recognition model, resulting in multiple feature recognition results for each segmentation feature sequence. Further, based on a preset vocabulary specific to a designated scenario, the optimal segmentation result is selected from the multiple feature recognition results corresponding to each segmentation feature sequence. Candidate segmentation results are determined based on the confidence level of each feature recognition result. This process adjusts the optimal segmentation results of each segmentation feature sequence using a preset vocabulary specific to a designated scenario, and determines the speech recognition result of the audio to be processed based on the optimal segmentation result and the candidate segmentation results, enabling the model to recognize vocabulary specific to a designated scenario.

[0084] To facilitate understanding of the technical solutions provided in the embodiments of this application, the following description uses a medical scenario as an example to illustrate a speech recognition method for a specific scenario provided in this application. Figure 2 As shown, it includes:

[0085] Step 201: Obtain the audio to be processed and determine the speech features of the audio to be processed;

[0086] In this embodiment, after receiving the audio to be processed, the audio is digitized to obtain a processing array for characterizing audio features. Then, a boundary probability algorithm is used to determine a first expectation representing speech features and a second expectation representing noise features in the processing array. Finally, the likelihood probability of the audio to be processed is determined based on the first and second expectations, and the speech features in the audio to be processed are determined according to the likelihood probability.

[0087] Specifically, such as Figure 3 As shown, assuming the received user speech is X, the user speech is first digitized, transforming it into an array XD representing audio using conventional techniques. Then, the likelihood probability of the array XD is calculated to determine the speech features and noise features in the user speech. Noise features are then removed to obtain the speech features in the user speech. The formula for calculating the likelihood probability is shown in formula (1) below:

[0088]

[0089]

[0090] Where L is the likelihood probability of the array XD to be processed, Y1 is the noise feature, and Y2 is the speech feature. The expansion calculation process of the above formula (1) is as follows: Figure 4 As shown, the final obtained D KL The (q∥p) term represents the divergence between the true distribution of the hidden variables and the predicted distribution of the hidden variables, while L b The term represents the joint probability of Y1 and Y2. L b The calculation of this term is currently only at the academic level; in practical applications, its value is usually defaulted to zero. The downside of this is that the speech feature space cannot fully represent real speech, ultimately leading to a decrease in the accuracy of the speech recognition model.

[0091] To address the aforementioned problems, embodiments of this application provide... Figure 4 Extending the formula shown, we transform the joint probability into the expectation problem of the boundary probabilities, making L b The ability to calculate this feature can fundamentally and significantly improve the quality of speech data, thereby enhancing the model's recognition accuracy. In practice, the first expectation representing the speech features and the second expectation representing the noise features of the array to be processed are determined using a boundary probability algorithm, as shown in the following formula (2):

[0092]

[0093] Where S is In expression S The two terms represent the distributions of Y1 and Y2 in the theoretical space, respectively, and can be ignored in the subsequent solution process. Furthermore, since this paper assumes Y1 is in a noisy space, the equation... This term can be understood as the probability of noise occurring under XD conditions, so it can be ignored. Therefore, the above formula (2) can be simplified to

[0094] Therefore, the joint probability is transformed into the expectation problem of the boundary probability, thus enabling the determination of L in practical applications. b The value of is then used to obtain the likelihood probability of the audio to be processed according to the above formula (1). Finally, the speech features in the audio to be processed, namely Y2, can be determined by existing algorithms based on the likelihood probability. This speech feature is the array data obtained by calculating the array to be processed using the above formula (2).

[0095] Step 202: Perform word segmentation processing on the speech features to obtain the word segmentation feature sequence of the speech features; and perform feature recognition on each word segmentation feature sequence to obtain multiple feature recognition results for each word segmentation feature sequence and the confidence level of each feature recognition result;

[0096] In this application, the speech recognition model is pre-trained using a traditional dictionary training set. Specifically, a conventional model training method such as CTC+Attention can be used, and this application does not limit this approach. The trained speech recognition model can accurately identify commonly used words, and then outputs the corresponding speech recognition result based on the identified speech content and the contextual semantics.

[0097] When performing step 202 above, if Figure 5 As shown, the speech features of the audio to be processed are input into a trained speech recognition model, enabling the model to segment the speech features into multiple segmented feature sequences. Each segmented feature sequence is equivalent to a subset of the array representing speech features. The model processing flow in this step can be simply understood as follows: first, the speech features are recognized character by character; after recognizing the complete speech content, the recognized speech content is segmented into words. Next, the model performs feature recognition on each segmented feature sequence, obtaining multiple sets of feature recognition results for each segmented feature sequence and the model confidence score corresponding to each set of feature recognition results.

[0098] Step 203: For each segmentation feature sequence, determine the optimal segmentation result of the segmentation feature sequence based on the preset specific scene-specific vocabulary and the feature recognition results of each; and determine a first number of candidate segmentation results from each feature recognition result according to the confidence level;

[0099] In conventional speech recognition models, the network often uses beam search for output. To facilitate understanding of the technical solution in this application's embodiments, a brief explanation of the beam search algorithm is provided below. The beam search algorithm is a heuristic graph search algorithm. When the graph space is large, to reduce the space and time occupied by the search, some low-quality nodes are pruned and some high-quality nodes are retained during each depth expansion step. The beam search algorithm requires a pre-defined beam size parameter to characterize the candidate results retained at each time step.

[0100] Assume that each time step processes a segmentation feature sequence, and each segmentation feature sequence has 5 possible outputs, A to E. Assume the beamwidth parameter is 2, meaning that each time step retains the two candidate segmentation results with the best conditional probability up to the current step. Specifically, as follows... Figure 6 As shown, in the first time step, A and C are the two optimal choices, resulting in two outcomes [A] and [C], while the other three are discarded. The second step continues generation based on these two outcomes. Branch A yields five possibilities: [AA], [AB], [AC], [AD], and [AE]. Similarly, branch C yields five possibilities. At this point, the confidence levels of these ten possibilities are used to filter them, re-selecting the two possibilities with the highest confidence. Figure 5 The model identifies [AB] and [CE] as shown in the diagram. Then, in the next time step, it determines the possible values ​​corresponding to [AB] and [CE], and selects the two possibilities with the highest confidence levels, [ABD] and [CED], based on the confidence level. Finally, it outputs [ABD] with the highest confidence level in text format as the final recognition result of the model.

[0101] In performing step 203 above, this embodiment introduces preset, scenario-specific vocabulary. In a medical scenario, this scenario-specific vocabulary can be medical terminology. Speech recognition models trained with conventional dictionaries cannot recognize such vocabulary. Therefore, this application constrains the output rules of conventional speech recognition models based on the scenario-specific vocabulary, enabling the model to recognize and output the specialized vocabulary. In implementation, this embodiment selects the beamforming algorithm output for each segmentation feature sequence based on the multiple feature recognition results corresponding to that segmentation feature sequence. The specific process is as follows... Figure 7 As shown, for each segmentation feature sequence, the optimal segmentation result in that segmentation feature sequence is obtained through the following steps 701 to 703.

[0102] Step 701: Detect whether there are any unprocessed recognition results in each feature recognition result;

[0103] In this embodiment, after determining multiple feature recognition results corresponding to the word segmentation feature sequence, each feature recognition result is compared with a preset set of specific scenario-specific vocabulary. The feature recognition result that is identical to any preset vocabulary for a specified scenario is taken as the recognition result to be processed. Simply put, in a medical scenario, if any feature recognition result contains medical-related specific terms, then that feature recognition result is determined as the recognition result to be processed.

[0104] As mentioned earlier, word segmentation feature sequences are arrays that can represent some speech features. Therefore, the feature recognition results mentioned above are the text recognition results obtained by the speech recognition model through feature recognition of the array-formatted word segmentation feature sequences. For example, feature recognition is performed on the word segmentation feature sequence array representing "ether" to obtain the text "ether".

[0105] Step 702: If it exists, the identification result with the highest confidence is taken as the optimal word segmentation result;

[0106] Step 703: If it does not exist, the feature recognition result with the highest confidence is taken as the optimal word segmentation result.

[0107] Assuming the model identifies the word segmentation feature sequence and obtains the following... Figure 8 The five feature recognition results shown (A through E) correspond to confidence levels of 75%, 54%, 27%, 13%, and 2%, respectively. These five feature recognition results are compared with pre-stored medical terminology, and the recognition results belonging to the medical terminology category are selected for further processing (i.e.,...). Figure 8 The five feature recognition results are shown as [B] and [E], and then the highest confidence recognition result [B] among [B] and [E] is taken as the optimal word segmentation result of the word segmentation feature sequence. If none of these five feature recognition results belong to medical professional terms, then the feature recognition result [A] with the highest confidence among these five feature recognition results is taken as the optimal word segmentation result of the word segmentation feature sequence.

[0108] To improve model accuracy and avoid misidentifying speech that does not contain medical terminology as such, before performing step 703, it is necessary to ensure that the confidence level of the highest-confidence recognition result is greater than a preset confidence threshold. The confidence threshold in this embodiment can be set according to actual needs, for example, 5%. Specifically, the above... Figure 7 The process shown is optimized as follows: Figure 9 As shown, it includes:

[0109] Step 901: Detect whether there are any unprocessed recognition results in each feature recognition result;

[0110] Step 902: If there are identification results to be processed, check whether the confidence of the identification result with the highest confidence is greater than the confidence threshold.

[0111] Step 903: If the confidence level is greater than the confidence threshold, the recognition result with the highest confidence level is taken as the optimal word segmentation result.

[0112] Step 904: If the confidence threshold is not greater than the threshold or there is no recognition result to be processed, then the feature recognition result with the highest confidence is taken as the optimal word segmentation result.

[0113] Based on the above Figure 8 For example, suppose that among the five feature recognition results A to E, only [E] belongs to medical professional terms and has a confidence level of 2%, which is less than the confidence threshold of 5%. In this case, [A] with the highest confidence level is taken as the optimal word segmentation result of the word segmentation feature sequence.

[0114] Next, the candidate segmentation results of the segmentation feature sequence will be explained. The candidate segmentation results in the segmentation feature sequence are determined according to the bundle width parameter in the bundle algorithm mentioned above. In this embodiment, the bundle width parameter is 2, that is, the first quantity in step 203 above is 2. Figure 8 For example, from the five feature recognition results, the two with the highest confidence [A] and [B] are selected as candidate word segmentation results for this word segmentation feature sequence. Furthermore, to further improve the model's ability to recognize the aforementioned specialized vocabulary, the feature recognition results can be reordered based on the recognition results to be processed, specifically as follows: Figure 10 As shown, assuming A to E are originally sorted by confidence level, after identifying the medical-related professional terms [B] and [E] in A to E, these results are sorted before the non-processable results [A], [C], and [D], and then sorted by confidence level from highest to lowest for each result. This process yields the following results: Figure 10 The order of [B], [E], [A], [C], and [D] is shown in the figure, and then the top two [B] and [E] are selected as candidate word segmentation results.

[0115] It should be understood that the above word segmentation results and the recognition results to be processed are all selected from the feature recognition results of each feature in the word segmentation feature sequence, and their data format is the same as that of the feature recognition results, which is text format.

[0116] Therefore, the above process can yield an optimal segmentation result and two alternative segmentation results for each segmentation feature sequence.

[0117] Step 204: Determine the speech recognition result of the audio to be processed based on the optimal segmentation result and the alternative segmentation results of each segmentation feature sequence.

[0118] As mentioned earlier, each word segmentation feature sequence is obtained by segmenting speech features, and each word segmentation feature sequence has a fixed positional order within the speech features. During implementation, the optimal word segmentation results are concatenated according to their order within the speech features to obtain the target concatenated feature containing all optimal word segmentation results. Then, the candidate word segmentation results are concatenated according to this order to obtain the candidate concatenated feature.

[0119] Specifically, such as Figure 11 As shown, for example, the speech features with the semantic meaning of "how to treat thyroid nodules" can be segmented into two word segmentation feature sequences A and B with the semantic meanings of "thyroid nodules" and "how to treat". Assume that the optimal word segmentation result for word segmentation feature sequence A is "thyroid nodules", and the alternative word segmentation results are "thyroid sister" and "thyroid nodules". The optimal word segmentation result for word segmentation feature sequence B is "how to treat", and the alternative word segmentation results are "how to treat" and "how to manage".

[0120] Then, according to the order of the word segmentation feature sequences, the optimal word segmentation result corresponding to word segmentation feature sequence A is concatenated with the optimal word segmentation result corresponding to word segmentation feature sequence B to obtain the target concatenated feature "how to treat thyroid nodules". Then, each candidate word segmentation result from word segmentation feature sequence A is combined with each candidate word segmentation result from word segmentation feature sequence B one by one to obtain multiple candidate concatenated features. Each candidate concatenated feature contains the candidate word segmentation result corresponding to each word segmentation feature sequence. Specifically, as follows... Figure 11 As shown, four sets of potential splicing features were obtained: "How to treat thyroid nodules", "How to manage thyroid nodules", "How to treat thyroid nodules" and "How to manage thyroid nodules".

[0121] After determining the target concatenation feature and candidate concatenation features corresponding to the speech feature through the above process, a second number of target concatenation features are determined from the candidate concatenation features based on the confidence levels of the segmentation results of each candidate concatenation feature. If the feature similarity between a target concatenation feature and any other target concatenation feature is greater than a similarity threshold, the speech recognition result is determined based on the target concatenation feature; otherwise, the comprehensive confidence level of each target concatenation feature is determined based on the confidence levels of each candidate segmentation result, and the speech recognition result is determined based on the target concatenation feature with the highest comprehensive confidence level.

[0122] As mentioned above Figure 11 Examples, specifically as follows: Figure 12As shown, if the target concatenation feature "how to treat thyroid nodules" exists among the candidate concatenation features, then the model's final output for that speech feature is determined to be "how to treat thyroid nodules". This allows the medical term "thyroid nodules," which might otherwise not be output by the model, to be successfully used as a speech recognition result. Conversely, if "how to treat thyroid nodules" does not exist among the candidate concatenation features, it means that the semantics of that speech feature does not contain medical terminology. In this case, the confidence level of each candidate concatenation result is determined by the confidence level of each alternative word segmentation result contained in each candidate concatenation feature. For example, the sum of the confidence levels of each alternative word segmentation result is used as the overall confidence level of the candidate concatenation result. Then, the candidate concatenation result with the highest overall confidence level is used as the final speech recognition result output by the model.

[0123] The above process adjusts the optimal segmentation results of each segmentation feature sequence by using pre-defined vocabulary specific to a specified scenario, and determines the speech recognition result of the audio to be processed based on the optimal segmentation result and the alternative segmentation results, so that the model has the ability to recognize vocabulary specific to a specified scenario.

[0124] It should be noted that the technical solution of this application is applicable to any specified scenario where it is difficult to obtain sufficient corpus to train the speech recognition model. The use of the medical scenario as a specified scenario is only for the purpose of introducing the technical solution of this application, and is not intended to limit the scope of application.

[0125] Based on the same inventive concept, this application provides a speech recognition device 1300 for a specific scenario, specifically as follows: Figure 13 As shown, it includes:

[0126] The feature acquisition module 1301 is configured to acquire the audio to be processed and determine the speech features of the audio to be processed;

[0127] The feature recognition module 1302 is configured to perform word segmentation processing on the speech features to obtain a word segmentation feature sequence of the speech features; and to perform feature recognition on each word segmentation feature sequence to obtain multiple feature recognition results for each word segmentation feature sequence and the confidence level of each feature recognition result;

[0128] The word segmentation processing module 1303 is configured to perform the following for each word segmentation feature sequence: determine the optimal word segmentation result of the word segmentation feature sequence based on a preset specified scenario-specific vocabulary and the feature recognition results of each word segmentation feature sequence; and determine a first number of candidate word segmentation results from each feature recognition result according to the confidence level.

[0129] The speech recognition module 1304 is configured to perform speech recognition of the audio to be processed based on the optimal segmentation result and alternative segmentation results of each of the segmentation feature sequences.

[0130] In some possible embodiments, the acquisition of speech features of the audio to be processed is performed, and the feature acquisition module 1301 is configured to:

[0131] The audio to be processed is digitized to obtain a processing array representing the audio features;

[0132] The likelihood probability of the audio to be processed is determined based on the array to be processed, and the speech features in the audio to be processed are determined based on the likelihood probability.

[0133] In some possible embodiments, the feature acquisition module 1301 is configured to perform the determination of the likelihood probability of the audio to be processed based on the array to be processed, wherein:

[0134] The first expectation representing the speech features and the second expectation representing the noise features of the array to be processed are determined by the boundary probability algorithm.

[0135] The likelihood probability is determined based on the first expectation and the second expectation.

[0136] In some possible embodiments, the optimal word segmentation result of the word segmentation feature sequence is determined based on a preset, specified scenario-specific vocabulary and the recognition results of each feature. The word segmentation processing module 1303 is configured to:

[0137] Detect whether any of the feature recognition results contain a recognition result that is identical to any of the specified scene-specific terms;

[0138] If it exists, the identification result with the highest confidence level is taken as the optimal word segmentation result; otherwise, the feature identification result with the highest confidence level is taken as the optimal word segmentation result.

[0139] In some possible embodiments, before performing the step of taking the identification result with the highest confidence as the optimal word segmentation result, the word segmentation processing module 1303 is further configured to:

[0140] The confidence level of the identification result with the highest confidence level is determined to be greater than the confidence threshold.

[0141] The word segmentation processing module 1303 is also configured to:

[0142] If the confidence level of the highest confidence-level identification result is not greater than the confidence level threshold, then the feature identification result with the highest confidence level is taken as the optimal word segmentation result.

[0143] In some possible embodiments, the speech recognition module 1304 is configured to perform the speech recognition result determined based on the optimal and alternative segmentation results of each of the segmentation feature sequences, wherein the speech recognition module 1304 is configured to:

[0144] Based on the order of the segmentation feature sequences in the speech features, the optimal segmentation results are concatenated to obtain a target concatenated feature containing all optimal segmentation results; and,

[0145] The candidate word segmentation results are concatenated according to the sorting to obtain the candidate concatenation features; wherein the candidate concatenation features include the candidate word segmentation results corresponding to each word segmentation feature sequence.

[0146] The speech recognition result is determined based on the target splicing features and the candidate splicing features.

[0147] In some possible embodiments, the speech recognition module 1304 is configured to perform the step of determining the speech recognition result based on the target splicing feature and the candidate splicing features, wherein the speech recognition module 1304 is configured to:

[0148] Based on the confidence level of each candidate word segmentation result contained in each candidate splicing feature, a second number of target splicing features are determined from each candidate splicing feature;

[0149] If the feature similarity between the target splicing feature and any other target splicing feature is greater than a similarity threshold, then the speech recognition result is determined based on the target splicing feature.

[0150] Otherwise, the comprehensive confidence of each target splicing feature is determined based on the confidence of each candidate word segmentation result, and the speech recognition result is determined based on the target splicing feature with the highest comprehensive confidence.

[0151] The following reference Figure 14 To describe an electronic device 130 according to this embodiment of the present application. Figure 14 The electronic device 130 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0152] like Figure 14 As shown, the electronic device 130 is presented in the form of a general-purpose electronic device. The components of the electronic device 130 may include, but are not limited to: at least one processor 131, at least one memory 132, and a bus 133 connecting different system components (including memory 132 and processor 131).

[0153] Bus 133 represents one or more of several bus structures, including a memory bus or memory controller, peripheral bus, processor, or local bus using any of the various bus structures.

[0154] The memory 132 may include a readable medium in the form of volatile memory, such as random access memory (RAM) 1321 and / or cache memory 1322, and may further include read-only memory (ROM) 1323.

[0155] The memory 132 may also include a program / utility 1325 having a set (at least one) of program modules 1324, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0156] Electronic device 130 can also communicate with one or more external devices 134 (e.g., keyboard, pointing device, etc.), and with one or more devices that enable a user to interact with electronic device 130, and / or with any device that enables electronic device 130 to communicate with one or more other electronic devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 135. Furthermore, electronic device 130 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 136. As shown, network adapter 136 communicates with other modules used in electronic device 130 via bus 133. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 130, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0157] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 132 including instructions, which can be executed by the processor 131 of the aforementioned device to perform the aforementioned method. Optionally, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0158] In an exemplary embodiment, a computer program product is also provided, including a computer program / instructions that, when executed by a processor 131, implement any of the methods in the speech recognition method for a specified scenario provided in this application.

[0159] In an exemplary embodiment, various aspects of the speech recognition method for a specified scenario provided in this application can also be implemented in the form of a program product, which includes program code. When the program product is run on a computer device, the program code is used to cause the computer device to perform the steps in the speech recognition method for a specified scenario according to the various exemplary embodiments of this application described above.

[0160] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0161] The speech recognition program product for a specified scenario according to the embodiments of this application can be a portable compact disc read-only memory (CD-ROM) and include program code, and can run on an electronic device. However, the program product of this application is not limited thereto. In this document, the readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0162] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. This propagated data signal may take many forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0163] The program code contained on the readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wired, fiber optic, RF, etc., or any suitable combination thereof.

[0164] Program code for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as Java or similar languages. The program code can execute entirely on the user's electronic device, partially on the user's device, as a standalone software package, partially on the user's electronic device and partially on a remote electronic device, or entirely on a remote electronic device or server. In cases involving remote electronic devices, the remote electronic device can be connected to the user's electronic device via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external electronic device (e.g., via the Internet using an Internet service provider).

[0165] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0166] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0167] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0168] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable image scaling device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable image scaling device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0169] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable image scaling device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions can also be loaded onto a computer or other programmable image scaling device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0171] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0172] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A speech recognition method for a specified scenario, characterized in that, The method includes: Acquire the audio to be processed and determine the speech features of the audio to be processed; The speech features are segmented into words to obtain a segmentation feature sequence; and feature recognition is performed on each segmentation feature sequence to obtain multiple feature recognition results for each segmentation feature sequence and the confidence level of each feature recognition result. For each segmentation feature sequence, the optimal segmentation result of the segmentation feature sequence is determined based on the preset specific vocabulary for the specified scenario and the recognition results of each feature; and a first number of candidate segmentation results are determined from each feature recognition result according to the confidence level. Based on the order of the segmentation feature sequences in the speech features, the optimal segmentation results are concatenated to obtain a target concatenated feature containing all optimal segmentation results; and, The candidate word segmentation results are concatenated according to the sorting to obtain the candidate concatenation features; wherein, the candidate concatenation features include the candidate word segmentation results corresponding to each word segmentation feature sequence; The speech recognition result is determined based on the target splicing features and the candidate splicing features.

2. The method according to claim 1, characterized in that, The acquisition of speech features of the audio to be processed includes: The audio to be processed is digitized to obtain a processing array representing the audio features; The likelihood probability of the audio to be processed is determined based on the array to be processed, and the speech features in the audio to be processed are determined based on the likelihood probability.

3. The method according to claim 2, characterized in that, Determining the likelihood probability of the audio to be processed based on the array to be processed includes: The first expectation representing the speech features and the second expectation representing the noise features of the array to be processed are determined by the boundary probability algorithm. The likelihood probability is determined based on the first expectation and the second expectation.

4. The method according to claim 1, characterized in that, The process of determining the optimal word segmentation result of the word segmentation feature sequence based on preset, scenario-specific vocabulary and the recognition results of each feature includes: Detect whether any of the feature recognition results contain a recognition result that is identical to any of the specified scene-specific terms; If it exists, the identification result with the highest confidence level is taken as the optimal word segmentation result; otherwise, the feature identification result with the highest confidence level is taken as the optimal word segmentation result.

5. The method according to claim 4, characterized in that, Before using the identification result with the highest confidence level as the optimal word segmentation result, the method further includes: The confidence level of the identification result with the highest confidence level is determined to be greater than the confidence threshold. The method further includes: If the confidence level of the highest confidence-level identification result is not greater than the confidence level threshold, then the feature identification result with the highest confidence level is taken as the optimal word segmentation result.

6. The method according to claim 1, characterized in that, Determining the speech recognition result based on the target splicing features and the candidate splicing features includes: Based on the confidence level of each candidate word segmentation result contained in each candidate splicing feature, a second number of target splicing features are determined from each candidate splicing feature; If the feature similarity between the target splicing feature and any other target splicing feature is greater than a similarity threshold, then the speech recognition result is determined based on the target splicing feature. Otherwise, the comprehensive confidence of each target splicing feature is determined based on the confidence of each candidate word segmentation result, and the speech recognition result is determined based on the target splicing feature with the highest comprehensive confidence.

7. A speech recognition device for a specified scenario, characterized in that, The device includes: The feature acquisition module is configured to acquire the audio to be processed and determine the speech features of the audio to be processed. The feature recognition module is configured to perform word segmentation processing on the speech features to obtain word segmentation feature sequences of the speech features; and to perform feature recognition on each word segmentation feature sequence to obtain multiple feature recognition results for each word segmentation feature sequence and the confidence level of each feature recognition result; The word segmentation processing module is configured to perform the following for each word segmentation feature sequence: determine the optimal word segmentation result of the word segmentation feature sequence based on a preset specified scenario-specific vocabulary and the feature recognition results; and determine a first number of candidate word segmentation results from each feature recognition result according to the confidence level. The speech recognition module is configured to concatenate the optimal word segmentation results according to the order of the word segmentation feature sequences in the speech features, to obtain a target concatenated feature containing all optimal word segmentation results; and, The candidate word segmentation results are concatenated according to the sorting to obtain the candidate concatenation features; wherein, the candidate concatenation features include the candidate word segmentation results corresponding to each word segmentation feature sequence; The speech recognition result is determined based on the target splicing features and the candidate splicing features.

8. An electronic device, characterized in that, include: Memory, used to store program instructions; A processor is configured to invoke program instructions stored in the memory and execute the steps of the method according to any one of claims 1-6 in accordance with the obtained program instructions.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a computer, cause the computer to perform the method as described in any one of claims 1-6.

Images