Embedded speech recognition method and device

Abstract

The invention relates to an embedded voice recognition method and device, and the method comprises the steps: obtaining original voice data through an embedded system, and carrying out the local feature recognition of the original voice data, and obtaining local feature recognition data; judging whether the local feature data corresponding to the voice recognition model on the embedded system is matched with the local feature recognition data or not; if yes, performing voice recognition on the original voice data by using a voice recognition model on the embedded system; and if not, sending the local feature recognition data to a remote server, performing feature matching in the remote server to obtain feature matching data, and updating a voice recognition model on the embedded system according to the feature matching data. The remote server is designed on the basis of the embedded system, the model parameters of the speech recognition model carried on the embedded system can be updated according to the local features of the local languages, and then precise speech recognition can be carried out on various local languages.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition, and in particular to an embedded speech recognition method and device. Background Art

[0002] Due to the special nature of embedded systems and the lack of their own development capabilities, the current speech recognition solutions based on embedded development have only designed a single speech recognition model. This single speech recognition model can usually only perform good speech recognition on Mandarin; while the recognition results for local languages ​​are unsatisfactory and accurate speech recognition results cannot be obtained. Summary of the invention

[0003] The present invention provides an embedded speech recognition method and device to solve at least one of the above technical problems.

[0004] The technical solution of the present invention to solve the above technical problems is as follows: an embedded speech recognition method, which uses a remote server and a speech recognition model mounted on an embedded system to perform speech recognition, comprising:

[0005] S1, using the embedded system to obtain original voice data, and performing local feature recognition on the original voice data to obtain local feature recognition data;

[0006] S2, determining whether the local feature data corresponding to the speech recognition model currently mounted on the embedded system matches the local feature recognition data; if so, performing speech recognition on the original speech data using the speech recognition model currently mounted on the embedded system; if not, executing S3 and S4;

[0007] S3, sending the local feature recognition data to the remote server through the network, performing feature matching in the remote server to obtain feature matching data, and updating the speech recognition model on the embedded system according to the feature matching data;

[0008] S4, performing speech recognition on the original speech data using the updated speech recognition model on the embedded system.

[0009] Based on the above technical solution, the present invention can also be improved as follows.

[0010] Furthermore, the remote service pre-stores a plurality of speech recognition models, each of which has different local feature data; in S3, feature matching is performed in the remote server specifically as follows:

[0011] The local feature recognition data is compared with the local feature data of multiple speech recognition models stored in the remote service to find out the speech recognition model corresponding to the local feature data consistent with the local feature recognition data; wherein the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is the feature matching data.

[0012] Further, in S3, updating the speech recognition model on the embedded system according to the feature matching data is specifically as follows: extracting model parameters from the speech recognition model corresponding to the local feature data consistent with the local feature recognition data, and sending the extracted model parameters to the embedded system through the network to update the model parameters of the speech recognition model on the embedded system.

[0013] Further, in S3, if the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is not found in the remote server, then:

[0014] Using the embedded system to collect a plurality of original voice data with the local feature recognition data, and sending the collected plurality of original voice data to the remote server through a network;

[0015] In the remote server, model training is performed using the local feature recognition data and a plurality of original speech data to obtain and store a speech recognition model corresponding to the local feature recognition data;

[0016] Model parameters are extracted from the speech recognition model corresponding to the local feature recognition data, and the extracted model parameters are sent to the embedded system through a network to update the model parameters of the speech recognition model on the embedded system.

[0017] Furthermore, it also includes:

[0018] The speech recognition models stored in the remote server and having a matching rate lower than a preset threshold are regularly cleared.

[0019] Based on the above-mentioned embedded speech recognition method, the present invention also provides an embedded speech recognition device.

[0020] An embedded speech recognition device, applied to the above-mentioned embedded speech recognition method, comprises an embedded system, wherein the embedded system comprises a sound pickup module, a microprocessor and a communication module;

[0021] The sound pickup module is used to obtain original voice data;

[0022] The microprocessor is equipped with a speech recognition model and is connected to the sound pickup module, and is used to perform local feature recognition on the original speech data to obtain local feature recognition data, and to determine whether the local feature data corresponding to the currently equipped speech recognition model matches the local feature recognition data; and when the local feature data corresponding to the currently equipped speech recognition model matches the local feature recognition data, the currently equipped speech recognition model is used to perform speech recognition on the original speech data;

[0023] The communication module is connected to the microprocessor and is connected to a remote server via a network, and is used to send the local feature recognition data to the remote server when the local feature data corresponding to the currently installed speech recognition model does not match the local feature recognition data;

[0024] The remote server is used to perform feature matching on the local feature recognition data to obtain feature matching data, and update the speech recognition model on the microprocessor according to the feature matching data;

[0025] The microprocessor is also used to perform speech recognition on the original speech data using an updated speech recognition model.

[0026] Based on the above technical solution, the present invention can also be improved as follows.

[0027] Furthermore, the remote service pre-stores a plurality of speech recognition models, each of which has different local feature data; the remote server is specifically used to: compare the local feature recognition data with the local feature data of the stored plurality of speech recognition models, and find out the speech recognition model corresponding to the local feature data consistent with the local feature recognition data; wherein the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is the feature matching data.

[0028] Furthermore, the remote server is also specifically used to: extract model parameters from the speech recognition model corresponding to the local feature data consistent with the local feature recognition data, and send the extracted model parameters to the microprocessor through the network to update the model parameters of the speech recognition model on the microprocessor.

[0029] Further, if the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is not found in the remote server, then:

[0030] The sound pickup module is also used to collect a plurality of original voice data with the local feature recognition data, and transmit the collected plurality of original voice data to the microprocessor;

[0031] The microprocessor is also used to send the collected multiple original voice data to the remote server through the network via the communication module;

[0032] The remote server is also used to perform model training using the local feature recognition data and multiple original speech data to obtain and store a speech recognition model corresponding to the local feature recognition data, and to extract model parameters from the speech recognition model corresponding to the local feature recognition data, and to send the extracted model parameters to the microprocessor via a network to update the model parameters of the speech recognition model on the microprocessor.

[0033] Furthermore, the remote server is also used to regularly clean up stored speech recognition models whose matching rates are lower than a preset threshold.

[0034] The beneficial effects of the present invention are as follows: an embedded speech recognition method and device of the present invention designs a remote server based on an embedded system, and can update the model parameters of the speech recognition model carried by the embedded system according to the local characteristics of the local language, thereby being able to perform accurate speech recognition on various local languages. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flowchart of an embedded speech recognition method of the present invention;

[0036] Figure 2 The structure block diagram of an embedded speech recognition device of the present invention. DETAILED DESCRIPTION

[0037] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0038] like Figure 1 As shown, an embedded speech recognition method uses a remote server and a speech recognition model mounted on an embedded system to perform speech recognition, including:

[0039] S1, using the embedded system to obtain original voice data, and performing local feature recognition on the original voice data to obtain local feature recognition data;

[0040] S2, determining whether the local feature data corresponding to the speech recognition model currently mounted on the embedded system matches the local feature recognition data; if so, performing speech recognition on the original speech data using the speech recognition model currently mounted on the embedded system; if not, executing S3 and S4;

[0041] S3, sending the local feature recognition data to the remote server through the network, performing feature matching in the remote server to obtain feature matching data, and updating the speech recognition model on the embedded system according to the feature matching data;

[0042] S4, performing speech recognition on the original speech data using the updated speech recognition model on the embedded system.

[0043] The embedded speech recognition method of the present invention designs a remote server based on an embedded system, and can update the model parameters of the speech recognition model carried on the embedded system according to the local characteristics of the local language, thereby being able to perform accurate speech recognition on various local languages.

[0044] In some embodiments, the remote service pre-stores a plurality of speech recognition models, each of which has different local feature data; in S3, feature matching is performed in the remote server specifically as follows:

[0045] The local feature recognition data is compared with the local feature data of multiple speech recognition models stored in the remote service to find out the speech recognition model corresponding to the local feature data consistent with the local feature recognition data; wherein the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is the feature matching data.

[0046] Specifically, each local language has its corresponding local features, which refer to language features, and multiple speech recognition models are pre-stored in the remote service, each of which can accurately recognize a local language. Usually, multiple speech recognition models are pre-stored in the remote service, including speech recognition models for recognizing Mandarin and speech recognition models for some popular local languages.

[0047] In some embodiments, in S3, updating the speech recognition model on the embedded system according to the feature matching data is specifically as follows: extracting model parameters from the speech recognition model corresponding to the local feature data consistent with the local feature recognition data, and sending the extracted model parameters to the embedded system through the network to update the model parameters of the speech recognition model on the embedded system.

[0048] Specifically, a plurality of speech recognition models having the same model architecture but different model parameters are stored in the remote service, and different model parameters are targeted at different local languages. Therefore, when a speech recognition model corresponding to the local feature data consistent with the local feature recognition data is found, the model parameters can be extracted, and the extracted model parameters can be sent to the embedded system through the network to update the model parameters of the speech recognition model on the embedded system. In addition, updating the speech recognition model by updating the model parameters is easy to implement, and the amount of data transmission is small, so the model update speed is also fast.

[0049] In some embodiments, in S3, if the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is not found in the remote server, then:

[0050] Using the embedded system to collect a plurality of original voice data with the local feature recognition data, and sending the collected plurality of original voice data to the remote server through a network;

[0051] In the remote server, model training is performed using the local feature recognition data and a plurality of original speech data to obtain and store a speech recognition model corresponding to the local feature recognition data;

[0052] Model parameters are extracted from the speech recognition model corresponding to the local feature recognition data, and the extracted model parameters are sent to the embedded system through a network to update the model parameters of the speech recognition model on the embedded system.

[0053] Specifically, when the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is not found in the remote server, the model can be trained online in the remote server to obtain the corresponding speech recognition model, thereby ensuring the diversity of model types in the remote server, and further enabling speech recognition of more local languages.

[0054] In some embodiments, it also includes:

[0055] The speech recognition models stored in the remote server and having a matching rate lower than a preset threshold are regularly cleared.

[0056] Specifically, since the capacity of the model database storing the speech recognition model in the remote server is limited, in order to ensure the normal operation of the remote server, the stored speech recognition models with a matching rate lower than a preset threshold can be cleaned up regularly; a low matching rate refers to the ratio of the number of matches to the total number of matches when performing feature matching in the remote server, that is, there has been no matching speech recognition model in the remote server for a long time.

[0057] Based on the above-mentioned embedded speech recognition method, the present invention also provides an embedded speech recognition device.

[0058] like Figure 2 As shown, an embedded speech recognition device is applied to the above-mentioned embedded speech recognition method, including an embedded system, wherein the embedded system includes a sound pickup module, a microprocessor and a communication module;

[0059] The sound pickup module is used to obtain original voice data;

[0060] The microprocessor is equipped with a speech recognition model and is connected to the sound pickup module, and is used to perform local feature recognition on the original speech data to obtain local feature recognition data, and to determine whether the local feature data corresponding to the currently equipped speech recognition model matches the local feature recognition data; and when the local feature data corresponding to the currently equipped speech recognition model matches the local feature recognition data, the currently equipped speech recognition model is used to perform speech recognition on the original speech data;

[0061] The communication module is connected to the microprocessor and is connected to a remote server via a network, and is used to send the local feature recognition data to the remote server when the local feature data corresponding to the currently installed speech recognition model does not match the local feature recognition data;

[0062] The remote server is used to perform feature matching on the local feature recognition data to obtain feature matching data, and update the speech recognition model on the microprocessor according to the feature matching data;

[0063] The microprocessor is also used to perform speech recognition on the original speech data using an updated speech recognition model.

[0064] The embedded speech recognition device of the present invention designs a remote server based on an embedded system, and can update the model parameters of the speech recognition model carried on the embedded system according to the local characteristics of the local language, thereby being able to perform accurate speech recognition on various local languages.

[0065] In some embodiments, the remote service pre-stores a plurality of speech recognition models, each of which has different local feature data; the remote server is specifically used to: compare the local feature recognition data with the local feature data of the plurality of stored speech recognition models, and find out the speech recognition model corresponding to the local feature data consistent with the local feature recognition data; wherein the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is the feature matching data.

[0066] Specifically, each local language has its corresponding local features, which refer to language features, and multiple speech recognition models are pre-stored in the remote service, each of which can accurately recognize a local language. Usually, multiple speech recognition models are pre-stored in the remote service, including speech recognition models for recognizing Mandarin and speech recognition models for some popular local languages.

[0067] In some embodiments, the remote server is also specifically used to: extract model parameters from the speech recognition model corresponding to the local feature data consistent with the local feature recognition data, and send the extracted model parameters to the microprocessor through the network to update the model parameters of the speech recognition model on the microprocessor.

[0068] Specifically, a plurality of speech recognition models having the same model architecture but different model parameters are stored in the remote service, and different model parameters are targeted at different local languages. Therefore, when a speech recognition model corresponding to the local feature data consistent with the local feature recognition data is found, the model parameters can be extracted, and the extracted model parameters can be sent to the embedded system through the network to update the model parameters of the speech recognition model on the embedded system. In addition, updating the speech recognition model by updating the model parameters is easy to implement, and the amount of data transmission is small, so the model update speed is also fast.

[0069] In some embodiments, if the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is not found in the remote server, then:

[0070] The sound pickup module is also used to collect a plurality of original voice data with the local feature recognition data, and transmit the collected plurality of original voice data to the microprocessor;

[0071] The microprocessor is also used to send the collected multiple original voice data to the remote server through the network via the communication module;

[0072] The remote server is also used to perform model training using the local feature recognition data and multiple original speech data to obtain and store a speech recognition model corresponding to the local feature recognition data, and to extract model parameters from the speech recognition model corresponding to the local feature recognition data, and to send the extracted model parameters to the microprocessor via a network to update the model parameters of the speech recognition model on the microprocessor.

[0073] Specifically, when the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is not found in the remote server, the model can be trained online in the remote server to obtain the corresponding speech recognition model, thereby ensuring the diversity of model types in the remote server, and further enabling speech recognition of more local languages.

[0074] In some embodiments, the remote server is also used to periodically clean up stored speech recognition models whose matching rates are lower than a preset threshold.

[0075] Specifically, since the capacity of the model database storing the speech recognition model in the remote server is limited, in order to ensure the normal operation of the remote server, the stored speech recognition models with a matching rate lower than a preset threshold can be cleaned up regularly; a low matching rate refers to the ratio of the number of matches to the total number of matches when performing feature matching in the remote server, that is, there has been no matching speech recognition model in the remote server for a long time.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An embedded speech recognition method, characterized in that: Use a remote server and a speech recognition model on an embedded system to perform speech recognition, including: S1, using the embedded system to obtain original voice data, and performing local feature recognition on the original voice data to obtain local feature recognition data; S2, determining whether the local feature data corresponding to the speech recognition model currently mounted on the embedded system matches the local feature recognition data; if so, performing speech recognition on the original speech data using the speech recognition model currently mounted on the embedded system; if not, executing S3 and S4; S3, sending the local feature recognition data to the remote server through the network, performing feature matching in the remote server to obtain feature matching data, and updating the speech recognition model on the embedded system according to the feature matching data; S4, performing speech recognition on the original speech data using the updated speech recognition model on the embedded system.

2. The embedded speech recognition method according to claim 1, characterized in that: The remote service pre-stores a plurality of speech recognition models, each of which has different local feature data; in S3, feature matching is performed in the remote server specifically as follows: The local feature recognition data is compared with the local feature data of multiple speech recognition models stored in the remote service to find out the speech recognition model corresponding to the local feature data consistent with the local feature recognition data; wherein the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is the feature matching data.

3. The embedded speech recognition method according to claim 2, characterized in that: In S3, updating the speech recognition model on the embedded system according to the feature matching data is specifically as follows: extracting model parameters from the speech recognition model corresponding to the local feature data consistent with the local feature recognition data, and sending the extracted model parameters to the embedded system through the network to update the model parameters of the speech recognition model on the embedded system.

4. The embedded speech recognition method according to claim 2, characterized in that: In S3, if the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is not found in the remote server, then: Using the embedded system to collect a plurality of original voice data with the local feature recognition data, and sending the collected plurality of original voice data to the remote server through a network; In the remote server, model training is performed using the local feature recognition data and a plurality of original speech data to obtain and store a speech recognition model corresponding to the local feature recognition data; Model parameters are extracted from the speech recognition model corresponding to the local feature recognition data, and the extracted model parameters are sent to the embedded system through a network to update the model parameters of the speech recognition model on the embedded system.

5. The embedded speech recognition method according to claim 2, characterized in that: Also includes: The speech recognition models stored in the remote server and having a matching rate lower than a preset threshold are regularly cleared.

6. An embedded speech recognition device, characterized in that: The embedded speech recognition method according to any one of claims 1 to 5 comprises an embedded system, wherein the embedded system comprises a sound pickup module, a microprocessor and a communication module; The sound pickup module is used to obtain original voice data; The microprocessor is equipped with a speech recognition model and is connected to the sound pickup module, and is used to perform local feature recognition on the original speech data to obtain local feature recognition data, and to determine whether the local feature data corresponding to the currently equipped speech recognition model matches the local feature recognition data; and when the local feature data corresponding to the currently equipped speech recognition model matches the local feature recognition data, the currently equipped speech recognition model is used to perform speech recognition on the original speech data; The communication module is connected to the microprocessor and is connected to a remote server via a network, and is used to send the local feature recognition data to the remote server when the local feature data corresponding to the currently installed speech recognition model does not match the local feature recognition data; The remote server is used to perform feature matching on the local feature recognition data to obtain feature matching data, and update the speech recognition model on the microprocessor according to the feature matching data; The microprocessor is also used to perform speech recognition on the original speech data using an updated speech recognition model.

7. The embedded speech recognition device according to claim 6, characterized in that: The remote service pre-stores a plurality of speech recognition models, each of which has different local feature data; the remote server is specifically used to: compare the local feature recognition data with the local feature data of the stored plurality of speech recognition models, and find out the speech recognition model corresponding to the local feature data consistent with the local feature recognition data; wherein the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is the feature matching data.

8. The embedded speech recognition device according to claim 7, characterized in that: The remote server is also specifically used to extract model parameters from the speech recognition model corresponding to the local feature data consistent with the local feature recognition data, and send the extracted model parameters to the microprocessor through the network to update the model parameters of the speech recognition model on the microprocessor.

9. The embedded speech recognition device according to claim 7, characterized in that: If the speech recognition model corresponding to the local feature data consistent with the local feature recognition data is not found in the remote server, then: The sound pickup module is also used to collect a plurality of original voice data with the local feature recognition data, and transmit the collected plurality of original voice data to the microprocessor; The microprocessor is also used to send the collected multiple original voice data to the remote server through the network via the communication module; The remote server is also used to perform model training using the local feature recognition data and multiple original speech data to obtain and store a speech recognition model corresponding to the local feature recognition data, and to extract model parameters from the speech recognition model corresponding to the local feature recognition data, and to send the extracted model parameters to the microprocessor via a network to update the model parameters of the speech recognition model on the microprocessor.

10. The embedded speech recognition device according to claim 7, characterized in that: The remote server is also used to periodically clean up stored speech recognition models whose matching rates are lower than a preset threshold.

Images