Voice interaction method, vehicle, server, and storage medium

Abstract

The present invention discloses a voice interaction method, a vehicle, a server, and a storage medium. The voice interaction method includes: performing voice recognition on a voice request to obtain text information; querying whether text information exists in a high-frequency set, the high-frequency set including a correspondence between preset text information and a generalized address search intent; when text information exists in the high-frequency set, determining whether the text information is a generalized address search intent based on the high-frequency set; when text information does not exist in the high-frequency set, using an intent prediction model to identify the text information to determine whether the text information is a generalized address search intent. In the above-mentioned voice interaction method, vehicle, server, and storage medium, when text information exists in the high-frequency set, simply determining whether the text information is a generalized address search intent based on the high-frequency set; when text information does not exist in the high-frequency set, identifying the text information based on the intent prediction model to determine whether the text information is a generalized address search intent.

Description

Technical Field

[0001] The present invention relates to voice technology, and in particular to a voice interaction method, a vehicle, a server and a computer-readable storage medium. Background Art

[0002] In related technologies, the voice system's judgment of user intentions is inaccurate, and it is difficult for the voice system to accurately identify the user's intentions, resulting in the user's voice needs not being met and the user experience being poor. Summary of the Invention

[0003] The present invention provides a voice interaction method, a vehicle, a server and a computer-readable storage medium.

[0004] The voice interaction method of the present invention includes: performing voice recognition on a voice request to obtain text information; querying whether the text information exists in a pre-stored high-frequency set, the high-frequency set including a correspondence between preset text information and a generalized address search intention; if the text information exists in the high-frequency set, determining whether the text information is a generalized address search intention based on the high-frequency set; if the text information does not exist in the high-frequency set, using a pre-stored intention prediction model to identify the text information to determine whether the text information is a generalized address search intention.

[0005] In the above-mentioned voice interaction method, by querying whether there is text information in the pre-stored high-frequency set, it is possible to simply determine whether the text information is a generalized address search intention based on the high-frequency set when there is text information in the high-frequency set; when there is no text information in the high-frequency set, the text information is identified based on the pre-stored intention prediction model to determine whether the text information is a generalized address search intention.

[0006] The voice interaction method includes: collecting historical voice requests; performing voice recognition on the historical voice requests to obtain historical text information, the historical text information including valid historical text information and invalid historical text information, the valid historical text information being marked with a generalized address search intention or a non-generalized address search intention; removing the invalid historical text information; and screening from the valid historical text information to obtain valid historical text information having an appearance frequency greater than a preset frequency as the high-frequency set.

[0007] In this way, a high-frequency set can be obtained based on the user's historical voice requests.

[0008] The volume of the intention prediction model is smaller than a preset volume, and the voice interaction method includes: obtaining a model to be processed; and compressing the model to be processed to obtain the intention prediction model.

[0009] In this way, an intention prediction model with a volume smaller than a preset volume can be obtained through compression processing.

[0010] The compressing the model to be processed to obtain the intention prediction model includes: performing distillation processing on the model to be processed to obtain a distillation model; and performing model quantization on the distillation model to obtain the intention prediction model.

[0011] In this way, the intention prediction model can be obtained by performing distillation processing and model quantization processing on the model to be processed.

[0012] The method of obtaining the model to be processed includes: collecting historical voice requests; performing voice recognition on the historical voice requests to obtain historical text information, wherein the historical text information includes valid historical text information and invalid historical text information; removing the invalid historical text information; using the valid historical text information to train a pre-trained model to obtain an updated model; and obtaining the model to be processed based on the updated model.

[0013] In this way, the user's historical voice requests and pre-trained model can be combined to obtain an updated model, which not only utilizes powerful pre-training technology but also fits the user's usage habits, making the updated model more effective.

[0014] The valid historical text information is annotated with the generalized address search intent or the non-generalized address search intent, and the valid historical text information includes a training set, a validation set and a test set; obtaining the model to be processed based on the updated model includes: training a classification model using the updated model, the training set and the validation set to obtain the model to be processed; and testing the model to be processed using the test set.

[0015] In this way, the model to be processed can be obtained by training with effective historical text information.

[0016] The volume of the intention prediction model is smaller than a preset volume, and the voice interaction method is applied to a vehicle without a network.

[0017] In this way, in the absence of a network, even if the vehicle is limited by computing power, it is still possible to use an intention prediction model with a volume smaller than the preset volume to realize the recognition of generalized address search intentions.

[0018] The vehicle of the present invention includes one or more processors and memories, wherein the memories store a computer program, and when the computer program is executed by the processor, the steps of the voice interaction method of any one of the above-mentioned embodiments are implemented.

[0019] In the above-mentioned vehicle, by querying whether there is text information in the pre-stored high-frequency set, it is possible to simply determine whether the text information is a generalized address search intention based on the high-frequency set when there is text information in the high-frequency set; when there is no text information in the high-frequency set, the text information is identified based on the pre-stored intention prediction model to determine whether the text information is a generalized address search intention.

[0020] The server of the present invention includes one or more processors and memories, wherein the memories store computer programs, and when the computer programs are executed by the processors, the steps of the above-mentioned voice interaction method are implemented.

[0021] In the above-mentioned server, by querying whether there is text information in the pre-stored high-frequency set, it is possible to simply determine whether the text information is a generalized address search intention based on the high-frequency set when there is text information in the high-frequency set; when there is no text information in the high-frequency set, the text information is identified based on the pre-stored intention prediction model to determine whether the text information is a generalized address search intention.

[0022] The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the voice interaction method of any one of the above-mentioned embodiments.

[0023] In the above-mentioned computer-readable storage medium, by querying whether there is text information in the pre-stored high-frequency set, it is possible to simply determine whether the text information is a generalized address search intention based on the high-frequency set when there is text information in the high-frequency set; when there is no text information in the high-frequency set, the text information is identified according to the pre-stored intention prediction model to determine whether the text information is a generalized address search intention.

[0024] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0026] Figure 1 and Figure 2 It is a flowchart of the voice interaction method of the present invention;

[0027] Figure 3 is a schematic diagram of a vehicle of the present invention;

[0028] Figure 4 is a schematic diagram of the speech system of the present invention;

[0029] Figures 5 to 9It is a flowchart of the voice interaction method of the present invention;

[0030] Figure 10 It is a schematic diagram of the connection between a vehicle and a computer-readable storage medium of the present invention.

[0031] Description of main component symbols:

[0032] Speech system 10 , vehicle 12 , server 14 , processor 101 , memory 102 , and computer-readable storage medium 50 . DETAILED DESCRIPTION

[0033] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0034] In related technologies, the voice system's judgment of user intentions is inaccurate, and it is difficult for the voice system to accurately identify the user's intentions, resulting in the user's voice needs not being met and the user experience being poor.

[0035] See also Figure 1 and Figure 2 , the voice interaction method of the present invention includes:

[0036] 012: Perform voice recognition on voice requests to obtain text information;

[0037] 014: Query whether there is text information in the pre-stored high-frequency set, where the high-frequency set includes the correspondence between the preset text information and the generalized address search intent;

[0038] 016: When there is text information in the high-frequency set, determine whether the text information is a generalized address search intent based on the high-frequency set;

[0039] 018: When there is no text information in the high-frequency set, the pre-stored intent prediction model is used to identify the text information to determine whether the text information is a generalized address search intent.

[0040] See also Figure 3 and Figure 4The voice interaction method of the present invention can be applied to the vehicle 12 or server 14 of the present invention. When the voice interaction method is applied to the vehicle 12, the microphone of the vehicle 12 can receive the user's voice request, which can then be forwarded to the processor 101 of the vehicle 12 for processing. When the voice interaction method is applied to the server 14, the microphone of the vehicle 12 can receive the user's voice request, which can then be forwarded to the processor 101 of the server 14 for processing. The server 14 can then forward the processed result of the voice request to the vehicle 12. The vehicle 12 and server 14 form the voice system 10. Of course, the vehicle 12 and server 14 of the present invention can also cooperate to implement the voice interaction method of the present invention. That is, some steps of the voice interaction method of the present invention can be implemented by the vehicle 12, and other steps can be implemented by the server 14. This is not specifically limited here. For example, the vehicle 12 of the present invention can be used to implement the online process of the voice interaction method, and the server 14 can be used to implement the offline process of the voice interaction method. The online process includes the application of the high-frequency set and the intent prediction model, and the offline process includes the process of establishing the high-frequency set and training the intent prediction model. Step 012, step 014, step 016 and step 018 can be regarded as an online process of the voice interaction method.

[0041] The voice request may be an original audio component, and text information may be obtained by performing speech recognition (Automatic Speech Recognition, ASR) on the voice request.

[0042] A query is performed to determine whether the current text information exists in a pre-stored high-frequency set, which includes a correspondence between the preset text information and a generalized address search intent. A generalized address search intent refers to a user's destination during the driving of the vehicle 12 being a generalized address, such as "Please help me find a nearby fast food restaurant that is still open, preferably cheaper," or "Please help me navigate to a free parking lot next to a certain park," rather than a specific location or other non-address search intent (such as music playback, weather query, etc.). Hereinafter, specific location search intent and other non-address search intents are collectively referred to as non-generalized address search intent.

[0043] If the current text information exists in the high-frequency set, it can be quickly and easily determined whether the current text information represents a generalized address search intent. If the current text information does not exist in the high-frequency set, a pre-stored intent prediction model can be used to identify the current text information and determine whether the current text information represents a generalized address search intent.

[0044] In the above-mentioned voice interaction method, by querying whether there is text information in the pre-stored high-frequency set, it is possible to simply determine whether the text information is a generalized address search intention based on the high-frequency set when there is text information in the high-frequency set; when there is no text information in the high-frequency set, the text information is identified based on the pre-stored intention prediction model to determine whether the text information is a generalized address search intention.

[0045] In the case where the text information is a generalized address search intent, the segmented intent prediction model can be used to further identify the search intent category of the text information. The search intent categories include the first category of search intent, the second category of search intent, and the third category of search intent. Among them, the first category of search intent can be a point search intent, such as "navigate to a certain university", where "a certain university" is a "point". The second category of search intent can be a line search intent, such as "find a free charging station on a certain road", where "a certain road" is a "line". The third category of search intent can be a surface search intent, such as "help me find a hospital in a certain district", where "a certain district" is a "surface". Of course, the first category of search intent can also be a line or surface search intent, the second category of search intent can also be a point or surface search intent, and the third category of search intent can also be a point or line search intent, which will not be repeated here. The present invention is explained by taking the first category of search intent as a point search intent, the second category of search intent as a line search intent, and the third category of search intent as a surface search intent as an example.

[0046] The search intent category can be a single category among the first category, second category and third category search intent, or it can be the intersection of at least two of the first category, second category and third category search intent. For example, the search intent category is the intersection of the first category and the third category search intent. No specific limitation is made here.

[0047] After further identifying the search intent category of text information using the segmented intent prediction model, the specific search intent category can be sent to a downstream service, such as a map server, which can perform searches based on the specific search intent category.

[0048] See also Figure 2 and Figure 5 , the voice interaction methods include:

[0049] 022: Collect historical voice requests;

[0050] 024: Perform voice recognition on historical voice requests to obtain historical text information. The historical text information includes valid historical text information and invalid historical text information. The valid historical text information is marked with generalized address search intent or non-generalized address search intent.

[0051] 026: Remove invalid historical text information;

[0052] 028: Filter out valid historical text information with an appearance frequency greater than a preset frequency from the valid historical text information to obtain the valid historical text information as a high-frequency set.

[0053] In this way, a high-frequency set can be obtained based on the user's historical voice requests.

[0054] Specifically, to protect user privacy, historical voice requests can be collected with user permission. These historical voice requests can be collected from different vehicles. To ensure more accurate high-frequency collection, the number of collected historical voice requests can exceed 100,000. After performing voice recognition on these historical voice requests, historical text information is obtained. This historical text information includes valid and invalid historical text information. Therefore, this historical text information can be filtered (command filtering). Specifically, invalid historical text information, such as historical text information with obvious ambiguous meanings and short historical text information containing only modal particles such as "ah" and "oh," is filtered out. Only historical text information with clear meaning and specific purposes is retained, such as "Navigate to X University," "Turn on the air conditioner for me," "Search for nearby hospitals," "Play Zhou X's songs," and "What's the weather like today?" can be manually annotated (data labeling). The labeling target is <0, 1>, where 0 represents a non-generalized address search intent and 1 represents a generalized address search intent. Valid historical text information with a frequency greater than a preset frequency can be screened from the valid historical text information to be used as a high-frequency set. Then, part of the valid historical text information can be randomly selected as a coverage test set and the coverage of the high-frequency set can be tested. The coverage is the proportion of valid historical text information that can be hit in the entire coverage test set. The valid historical text information that can be hit refers to the valid historical text information that can be queried in the high-frequency set in the coverage test set. When the coverage is less than the preset coverage, the preset frequency can be reduced to obtain more historical text information as a high-frequency set; when the coverage is greater than the preset coverage, the preset frequency can be increased to obtain less historical text information as a high-frequency set, wherein the preset coverage is, for example, 50%. A cache engine can be constructed using a high-frequency set, that is, the cache engine can be loaded with valid historical text information with a frequency greater than a preset frequency and annotations indicating whether it is a generalized address search intention.

[0055] In related technologies, voice systems need to respond to users' voice requests when there is a network. However, in actual applications, users will always encounter situations where the network signal is poor or there is no network. In this case, related technologies usually can only not respond to users' voice requests, answer "I don't understand" or simply not respond directly, so that users' voice needs are not met and the user experience is poor.

[0056] The volume of the intention prediction model is smaller than a preset volume, and the voice interaction method is applied to a vehicle 12 without a network.

[0057] In this way, in the absence of a network, even if the vehicle 12 is limited by computing power, it is still possible to use an intention prediction model with a volume smaller than a preset volume to realize the recognition of generalized address search intentions.

[0058] Specifically, after performing voice recognition on a voice request to obtain text information, the present invention can first use a high-frequency set pre-stored in the vehicle 10 to determine whether the text information is a generalized address search intention. If the text information does not exist in the pre-stored high-frequency set, an intention prediction model pre-stored in the vehicle 10 with a volume smaller than a preset volume can be used to realize the recognition of the generalized address search intention, thereby reducing the utilization rate of the processor 101 of the vehicle 12 and improving the overall intention recognition effect.

[0059] See also Figure 2 and Figure 6 ,The volume of the intent prediction model is smaller than the preset volume.,The voice interaction method includes:

[0060] 032: Get the model to be processed;

[0061] 034: Compress the model to be processed to obtain an intent prediction model.

[0062] In this way, an intention prediction model with a volume smaller than a preset volume can be obtained through compression processing.

[0063] Specifically, the model to be processed can be obtained first. The model to be processed is usually a large model obtained through training with a large amount of data. Therefore, in order to reduce the volume of the model to be processed, the model to be processed can be compressed to obtain an intent prediction model.

[0064] See also Figure 2 and Figure 7 , compressing the model to be processed to obtain an intention prediction model (step 034), including:

[0065] 0342: Perform distillation on the model to be processed to obtain a distilled model;

[0066] 0344: Perform model quantization on the distillation model to obtain an intent prediction model.

[0067] In this way, the intention prediction model can be obtained by performing distillation processing and model quantization processing on the model to be processed.

[0068] Specifically, after obtaining the model to be processed, the model to be processed can be distilled to obtain a distilled model. For example, if the model to be processed is a 110M model, a 10M distilled model can be obtained after distillation. The distillation process can be implemented using soft labels, feature distillation and other technologies. The distillation process can reduce some parameters in the model to be processed. After obtaining the distillation model, the effect of the distillation model can be verified. If the verification fails, more valid historical text information can be added to train the classification model to obtain the model to be processed (see below for details). If the verification passes, the distillation model can be quantized. Model quantization can specifically be an adjustment of accuracy, for example, quantization from float32 to int8, thereby further compressing the distillation model. Model quantization methods include BinnaryConnect, DeepCompression, etc., which are not specifically limited here.

[0069] See also Figure 2 and Figure 8 , obtaining the model to be processed (step 032), including:

[0070] 0321: Collect historical voice requests;

[0071] 0323: Perform voice recognition on historical voice requests to obtain historical text information, where the historical text information includes valid historical text information and invalid historical text information;

[0072] 0325: Remove invalid historical text information;

[0073] 0327: Use valid historical text information to train the pre-trained model to obtain an updated model;

[0074] 0329: Get the model to be processed based on the updated model.

[0075] In this way, the user's historical voice requests and pre-trained model can be combined to obtain an updated model, which not only utilizes powerful pre-training technology but also fits the user's usage habits, making the updated model more effective.

[0076] Specifically, to protect user privacy, historical voice requests can be collected with user permission, and these historical voice requests can be collected from different vehicles. To improve the training effectiveness of intent prediction models and segmented intent prediction models, the number of historical voice requests collected can exceed 100,000. It is understood that the more historical voice requests collected, the better the training results; the fewer historical voice requests collected, the faster the training speed. After performing voice recognition on the historical voice requests, historical text information is obtained. This historical text information includes valid historical text information and invalid historical text information. Therefore, the historical text information can be filtered (command filtering). Specifically, invalid historical text information, such as historical text information with obvious unclear meanings and short historical text information containing only modal particles such as "ah" and "oh," is filtered out, while historical text information with clear meanings and specific purposes is retained, such as "Navigate to X University," "Turn on the air conditioner for me," "Search for nearby hospitals," "Play Zhou X's songs," and "What's the weather like today?" The pre-trained model can be trained using the retained valid historical text information. The pre-trained model can be an open-source pre-trained model, such as the BERT model, the SBERT model, the Nezha model, etc. After training with the valid historical text information, the weights of the trained pre-trained model can be better adapted to scenarios where users speak heavily in a vehicle. The newly weighted pre-trained model is then output as an updated model. The updated model can be used to assist in obtaining the model to be processed. Steps 0321 and 022 can be the same step, steps 0323 and 024 can be the same step, and steps 0325 and 026 can be the same step.

[0077] See also Figure 2 and Figure 9 , valid historical text information is annotated with generalized address search intent or non-generalized address search intent, and the valid historical text information includes a training set, a validation set, and a test set; obtaining a model to be processed according to the updated model (step 0329), including:

[0078] 03292: Train the classification model using the updated model, training set, and validation set to obtain the model to be processed;

[0079] 03294: Use the test set to test the model to be processed.

[0080] In this way, the model to be processed can be obtained by training with effective historical text information.

[0081] Specifically, the valid historical text information can be manually annotated (data annotation), and the annotation target is <0, 1>, where 0 represents a non-generalized address search intent and 1 represents a generalized address search intent. In order to improve the training effect, the effective annotation amount of the valid historical text information can be greater than 80,000, and 60,000 valid historical text information can be divided into a training set, 10,000 valid historical text information can be divided into a validation set, and 10,000 valid historical text information can be divided into a test set. The training set and the validation set are preprocessed using the updated model, and the preprocessed training set and validation set can be used to train the classification model, wherein the classification model can be a deep learning model, such as a convolutional neural network (CNN). The training process can be accelerated in the GPU of server 14, and the powerful processing power of the GPU of server 14 can be used to increase the training speed. After the training is completed, the classification model can be output as a model to be processed. The test set is used to test the model to verify the effectiveness of the model to be processed. If the model to be processed passes the test, the model to be processed can be used to recognize text information; if the model to be processed fails the test, the amount of valid historical text information data can be increased and training can be performed again until it passes the test.

[0082] The valid historical text information of the generalized address search intent is annotated with the historical search intent category. The valid historical text information of the generalized address search intent includes a category training set, a category verification set, and a category test set. The voice interaction method includes: using the update model, the category training set, and the category verification set to train the category classification model to obtain a segmented intent prediction model; and using the category test set to test the segmented intent prediction model.

[0083] In this way, the effective historical text information of the generalized address search intent can be used to train a segmented intent prediction model.

[0084] Specifically, the effective historical text information of the generalized address search intent can be manually annotated with subdivided intents, and the annotated target is <0, 1, 2, 3>, where 0 represents the first type of search intent, 1 represents the second type of search intent, 2 represents the third type of search intent, and 3 represents the cross intent. In order to improve the training effect, the effective annotation amount of the effective historical text information of the generalized address search intent can be greater than the preset annotation amount, and the effective historical text information of the generalized address search intent can be divided into a category training set, a category verification set, and a category test set. The category training set and the category verification set are preprocessed using the updated model, and the preprocessed category training set and category verification set can be used to train the category class model, wherein the category classification model can be a deep learning model, such as a convolutional neural network (CNN). The training process can be accelerated in the GPU of the server 14, and the powerful processing power of the GPU of the server 14 can be used to increase the training speed. After training is complete, the category classification model can be output as a segmented intent prediction model. The segmented intent prediction model can be tested using a category test set to verify its effectiveness. If the segmented intent prediction model passes the test, it can be used to identify the search intent category of text information. If the segmented intent prediction model fails the test, the amount of valid historical text information for generalized address search intent can be increased and training can be repeated until it passes the test. In addition, the segmented intent prediction model can be compressed to reduce its size, thereby facilitating its deployment in vehicle 10 to identify the search intent category of text information.

[0085] The text information includes valid text information and invalid text information. The voice interaction method includes: removing invalid text information; correcting valid text information to obtain corrected text information; querying whether text information exists in a pre-stored high-frequency set (step 014), including: querying whether corrected text information exists in a pre-stored high-frequency set.

[0086] In this way, invalid text information can be removed and valid text information can be corrected, thereby facilitating determination of whether the text information is intended for a generalized address search.

[0087] Specifically, text information includes valid and invalid text information. Therefore, the text information can be filtered. Specifically, invalid text information, such as text information with obvious unclear semantics and short text information containing only modal particles such as "ah" and "oh," is filtered out, while retaining text information with clear semantics and specific purposes, such as "Navigate to X University," "Help me turn on the air conditioner," "Search for nearby hospitals," "Play Zhou X's songs," and "What's the weather like today?" In actual application scenarios, due to hardware limitations of the vehicle 12, unstable network conditions, and user colloquialisms or dialects, text information may not be clear and accurate. Therefore, valid text information can be corrected, such as performing some routine text corrections, such as correcting "Navigate to nearby good food" to "Navigate to nearby delicious food." After obtaining the corrected text information, a query can be performed to determine whether the corrected text information exists in a pre-stored high-frequency collection.

[0088] Using a pre-stored intention prediction model to identify text information to determine whether the text information is a generalized address search intention (step 018), including: using the intention prediction model to identify text information to obtain a discrimination probability that the text information is a generalized address search intention; when the discrimination probability is greater than a first preset probability, determining that the text information is a generalized address search intention; when the discrimination probability is less than the first preset probability, determining that the text information is a non-generalized address search intention.

[0089] In this way, whether the text information is a generalized address search intention can be determined by the discrimination probability of the generalized address search intention.

[0090] Specifically, after using the intent prediction model to identify text information, the discrimination probability of the two categories <0, 1> can be given. If the probability of being judged as 1 exceeds the first preset probability (for example, 0.9), then the text information can be considered to be a generalized address search intention, and the segmented intent prediction model can be further used to identify the search intent category of the text information; in other cases, the text information can be considered to be a non-generalized address search intention, that is, it may be a clear location search intention or other non-address search intention (such as music playback, weather query, etc.), so other schemes can be used for further identification, which is not specifically limited here.

[0091] The search intent category of text information is identified by using a segmented intent prediction model, including: when the text information is a generalized address search intent, using the segmented intent prediction model to identify the text information to obtain the predicted probability that the text information is each search intent category; determining the maximum predicted probability among all predicted probabilities; when the maximum predicted probability is greater than a second preset probability, determining the search intent category of the text information to be the search intent category corresponding to the maximum predicted probability.

[0092] In this way, the search intention category of the text information can be determined by the predicted probability of the search intention category.

[0093] Specifically, when the text information is a generalized address search intention, the segmented intention prediction model is used to identify the segmented intention, and the search intention category corresponding to the maximum predicted probability is selected as the candidate category. It is judged whether the maximum predicted probability exceeds the second preset probability (for example, 0.9). If so, the search intention category of the text information is the search intention category corresponding to the maximum predicted probability (candidate category). If not, it can be considered that the search intention category of the text information has no established segmented intention, and only the text information is returned as a generalized address search intention, without returning the search intention category.

[0094] See also Figure 3 The vehicle 12 of the present invention includes one or more processors 101 and a memory 102. The memory 102 stores a computer program. When the computer program is executed by the processor 101, the steps of the voice interaction method of any one of the above embodiments are implemented.

[0095] For example, when the computer program is executed by the processor 101, the following may be achieved:

[0096] 012: Perform voice recognition on voice requests to obtain text information;

[0097] 014: Query whether there is text information in the pre-stored high-frequency set, where the high-frequency set includes the correspondence between the preset text information and the generalized address search intent;

[0098] 016: When there is text information in the high-frequency set, determine whether the text information is a generalized address search intent based on the high-frequency set;

[0099] 018: When there is no text information in the high-frequency set, the pre-stored intent prediction model is used to identify the text information to determine whether the text information is a generalized address search intent.

[0100] In the above-mentioned vehicle 12, by querying whether there is text information in the pre-stored high-frequency set, it is possible to simply determine whether the text information is a generalized address search intention based on the high-frequency set when there is text information in the high-frequency set; when there is no text information in the high-frequency set, the text information is identified based on the pre-stored intention prediction model to determine whether the text information is a generalized address search intention.

[0101] See also Figure 4 The server 14 of the present invention includes one or more processors 101 and a memory 102. The memory 102 stores a computer program. When the computer program is executed by the processor 101, the steps of the above-mentioned voice interaction method are implemented.

[0102] For example, when the computer program is executed by the processor 101, the following may be achieved:

[0103] 012: Perform voice recognition on voice requests to obtain text information;

[0104] 014: Query whether there is text information in the pre-stored high-frequency set, where the high-frequency set includes the correspondence between the preset text information and the generalized address search intent;

[0105] 016: When there is text information in the high-frequency set, determine whether the text information is a generalized address search intent based on the high-frequency set;

[0106] 018: When there is no text information in the high-frequency set, the pre-stored intent prediction model is used to identify the text information to determine whether the text information is a generalized address search intent.

[0107] In the above-mentioned server 14, by querying whether there is text information in the pre-stored high-frequency set, it is possible to simply determine whether the text information is a generalized address search intention based on the high-frequency set when there is text information in the high-frequency set; when there is no text information in the high-frequency set, the text information is identified based on the pre-stored intention prediction model to determine whether the text information is a generalized address search intention.

[0108] See also Figure 10 The computer-readable storage medium 50 of the present invention stores a computer program thereon, and when the computer program is executed by the processor 101, the voice interaction method of any one of the above-mentioned embodiments is implemented.

[0109] For example, when the computer program is executed by the processor 101, it can achieve:

[0110] 012: Perform voice recognition on voice requests to obtain text information;

[0111] 014: Query whether there is text information in the pre-stored high-frequency set, where the high-frequency set includes the correspondence between the preset text information and the generalized address search intent;

[0112] 016: When there is text information in the high-frequency set, determine whether the text information is a generalized address search intent based on the high-frequency set;

[0113] 018: When there is no text information in the high-frequency set, the pre-stored intent prediction model is used to identify the text information to determine whether the text information is a generalized address search intent.

[0114] In the above-mentioned computer-readable storage medium 50, by querying whether there is text information in the pre-stored high-frequency set, it is possible to simply determine whether the text information is a generalized address search intention based on the high-frequency set when there is text information in the high-frequency set; when there is no text information in the high-frequency set, the text information is identified based on the pre-stored intention prediction model to determine whether the text information is a generalized address search intention.

[0115] In the present invention, a computer program includes computer program code. The computer program code may be in source code form, object code form, an executable file, or some intermediate form. Memory 102 may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device. Processor 101 may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor, or the processor may also be any conventional processor, etc.

[0116] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0117] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0118] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0119] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A voice interaction method, characterized in that: The voice interaction method includes: Perform speech recognition on voice requests to obtain text information; querying whether the text information exists in a pre-stored high-frequency set, wherein the high-frequency set includes a correspondence between preset text information and a generalized address search intention; In the case where the text information exists in the high-frequency set, determining whether the text information is a generalized address search intention based on the high-frequency set; In the case where the text information does not exist in the high-frequency set, using a pre-stored intention prediction model to identify the text information to determine whether the text information is a generalized address search intention; The voice interaction method includes: Collect historical voice requests; Performing voice recognition on the historical voice request to obtain historical text information, the historical text information including valid historical text information and invalid historical text information, the valid historical text information being marked with a generalized address search intent or a non-generalized address search intent; Removing the invalid historical text information; Filtering the valid historical text information to obtain valid historical text information having an appearance frequency greater than a preset frequency as the high-frequency set; After filtering out valid historical text information having an appearance frequency greater than a preset frequency from the valid historical text information to obtain the valid historical text information as the high-frequency set, the voice interaction method further includes: Randomly select some valid historical text information as a coverage test set, and test the coverage of the high-frequency set; When the coverage rate is less than the preset coverage rate, reducing the preset frequency to obtain more historical text information as a new high-frequency set; When the coverage rate is greater than a preset coverage rate, the preset frequency is increased to obtain less historical text information as a new high-frequency set.

2. The voice interaction method according to claim 1, characterized in that: The volume of the intention prediction model is smaller than a preset volume, and the voice interaction method includes: Get the model to be processed; The model to be processed is compressed to obtain the intention prediction model.

3. The voice interaction method according to claim 2, characterized in that: The compressing the model to be processed to obtain the intention prediction model includes: Performing distillation processing on the model to be processed to obtain a distillation model; Model quantization is performed on the distillation model to obtain the intent prediction model.

4. The voice interaction method according to claim 2, wherein: The step of obtaining the model to be processed includes: Collect historical voice requests; Performing voice recognition on the historical voice request to obtain historical text information, where the historical text information includes valid historical text information and invalid historical text information; Removing the invalid historical text information; Using the valid historical text information to train the pre-trained model to obtain an updated model; The model to be processed is obtained according to the updated model.

5. The voice interaction method according to claim 4, characterized in that: The valid historical text information is annotated with the generalized address search intent or the non-generalized address search intent, and the valid historical text information includes a training set, a validation set, and a test set; and obtaining the to-be-processed model according to the updated model includes: Training a classification model using the updated model, the training set, and the validation set to obtain the model to be processed; The model to be processed is tested using the test set.

6. The voice interaction method according to claim 1, characterized in that: The volume of the intention prediction model is smaller than a preset volume, and the voice interaction method is applied to a vehicle without a network.

7. A vehicle, characterized in that: The vehicle includes one or more processors and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the voice interaction method according to any one of claims 1 to 6 are implemented.

8. A server, characterized in that: The server includes one or more processors and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the voice interaction method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the voice interaction method described in any one of claims 1 to 6 is implemented.

Images