CN111583939A - Method and device for specific target wake-up by voice recognition - Google Patents

Abstract

The invention discloses a method and a device for waking up a specific target by voice recognition, wherein the method comprises the following steps: receiving a voice message of a specific target and extracting voice characteristics in the voice message; the voice characteristics of the specific target are used as input data of an HVS model which is trained in an identification mode, training is carried out, a specific target acoustic model is obtained, and the specific target acoustic model is stored; receiving a voice message of a target to be detected, and extracting voice characteristics in the voice message; taking the voice characteristics of the target to be tested as input data of a hidden vector state model trained in an identification mode, and training to obtain an acoustic model of the target to be tested; and comparing the acoustic model of the target to be detected with the acoustic model of the specific target, if the acoustic model of the target to be detected and the acoustic model of the specific target are related, performing language decoding on the voice characteristics of the target to be detected by using the language model, and judging whether to awaken or not according to a language decoding result. According to the invention, the HVS model of discriminant training is used as the acoustic model, so that the target can be accurately and quickly judged, and further the awakening function is achieved.

Description

Method and device for specific target wake-up by speech recognition

technical field

The present invention relates to the field of speech recognition, and in particular, to a method and device for speech recognition.

Background technique

In recent years, smart speakers have gradually changed the way people live. As a voice assistant, smart speakers can assist users in performing tasks in life, such as helping with car calls, shopping, reminders, recording information, etc. Although smart speakers bring more It is convenient, but there are still many security risks in smart speakers. Sometimes smart speakers cannot effectively determine whether the user is the initial user and the possibility of ordering goods with a credit card. Therefore, in order to prevent people from using it, many smart speakers are currently on the market. The speaker will use voice recognition as a protective measure.

General smart speakers usually use voice wake-up to wake up the smart speakers to perform subsequent tasks. The so-called voice wake-up method usually automatically captures some pre-registered voice commands (wake words) from a continuous voice. Traditionally, the Hidden Markov Model (HMM) technique was used to compare the feature vectors of individual phonemes and syllables to find the word with the highest probability (most likely), and later, combined with Gaussian mixture The model (Gaussian Mixture Model, GMM) forms the classic GMM-HMM model. The existing GMM-HMM model often adopts the maximum similarity training method (Maximum Likelihood). However, this method tends to make the competitor's answer probability greater than the correct answer probability under certain factors, resulting in a decline in the correct rate, so there is still progress. Room for improvement.

SUMMARY OF THE INVENTION

The purpose of the present invention is to aim at the defects and deficiencies of the above-mentioned prior art, and propose a method for realizing the wake-up of a specific target by speech recognition, which utilizes the wake-up word of the specific target in combination with the Hidden Vector State Model (Hidden Vector State Model) of discriminative training. Model, referred to as HVS Model), to realize the identification and monitoring of specific targets, so as to achieve the purpose of voice wake-up of specific targets.

In order to achieve the above object, an embodiment of the present invention provides a method for waking up a specific target by voice recognition, including the following steps:

S1: Receive a voice message of a specific target and preprocess the voice message of the specific target to extract a voice feature of the specific target;

S2: take the speech feature of the specific target as the input data of the latent vector state model (HVS Model) of the discriminative training and train to obtain a specific target acoustic model, and store the specific target acoustic model;

S3: Receive a voice message of a target to be tested and preprocess the voice message of the target to be tested to extract a voice feature of the target to be tested;

S4: take the voice feature of the target to be measured as the input data of the latent vector state model trained by the discriminant type and train to obtain an acoustic model of the target to be measured;

S5: Compare the correlation between the acoustic model of the target to be measured and the acoustic model of the specific target, and if the two are correlated, use at least one language model to perform language decoding on the speech feature of the target to be measured, And judge whether to wake up according to the language decoding result.

Specifically, the voice message of the specific target and the voice message of the target to be tested include at least one wake-up word.

Specifically, the preprocessing includes: performing noise suppression processing and echo cancellation processing on the voice message.

Specifically, the speech features are obtained by means of Mel cepstral coefficients (MFCC).

Specifically, the discriminative training adopts the maximum mutual information method (MMI) for training.

Specifically, the language model includes a vocabulary model or a grammar model or a combination thereof.

Specifically, judging whether the wake-up of speech recognition is achieved according to the language decoding result, the steps include: performing language decoding on the speech feature of the target to be tested; judging whether the voice message of the target to be tested contains the wake-up word; If the wake-up word is included, the voice recognition wake-up is activated, and if the wake-up word is not included, the voice recognition wake-up is not activated.

Another aspect of the embodiments of the present invention provides a voice recognition device for waking up a specific target, including:

an acquisition module including a plurality of microphone arrays for receiving voice messages of a specific target and a target to be tested, wherein the voice messages include a wake-up word;

an extraction module, connected to the acquisition module, for extracting the MFCC voice features in the voice messages of the specific target and the target to be tested;

a training module, connected to the extraction module, for using the MFCC voice features in the voice messages of the specific target and the target to be tested as the input data of the latent vector state model trained by the maximum mutual information method, and Obtain the acoustic model of the specific target after training and the acoustic model of the target to be tested;

a storage module, connected to the training module, for saving the acoustic model of the specific target that has been trained;

a decoding module, connected to the extraction module, for performing language decoding on the voice message of the target to be tested; and

a processor module, connected to the training module, the storage module and the decoding module, for comparing the acoustic model of the specific target in the storage module with the acoustic model of the target to be tested, and According to the comparison result, it is judged whether to activate the decoding module to perform language decoding of the voice message of the target to be tested, and to confirm whether the voice message of the target to be tested after the language decoding contains a wake-up word to wake up the device.

Specifically, the device further includes a registration module, the registration module is connected to the acquisition module and the storage module, and the registration module is used to start saving the acoustic model of a specific target to the storage module Group.

Specifically, the device further includes a wireless communication module, wherein the wireless communication module is used for external communication connection.

Compared with the prior art, the voice recognition method and device for specific target awakening of the present invention adopts the latent vector state model of discriminative training as the acoustic model. The probability of a competitor's appearance is reduced, the ability to distinguish between the correct answer and the competitor is increased, and it can quickly and accurately determine whether the target to be tested is a specific target, thereby achieving the function of awakening.

Description of drawings

FIG. 1 is a schematic flowchart of a method for waking up a specific target by voice recognition according to an embodiment of the present invention.

FIG. 2 is a schematic diagram of an apparatus for waking up a specific target by voice recognition according to an embodiment of the present invention.

The reference numerals in the figure are explained as follows:

100 Speech Recognition Device 11 Acquisition Module

12 Extraction module 13 Training module

14 Storage Module 15 Decoding Module

16 Processor Module 17 Register Module

18 Wireless communication module

S101～S105 Process steps.

Detailed ways

In order to describe the technical content, structural features, achieved goals and effects of the present invention in detail, the following examples are given and described in detail with the drawings.

Please refer to FIG. 1. FIG. 1 is a schematic flowchart of a method for waking up a specific target by speech recognition disclosed in an embodiment of the present invention, including the following steps:

Step S101: Receive a voice message of a specific target and preprocess the voice message of the specific target to extract a voice feature of the specific target;

Specifically, the specific target in this step refers to the registered user who has reached the wake-up condition in the speech recognition, and the voice message is a text prepared in advance. The text content will contain a preset wake-up word, and the specific target will read the text aloud first. content and collect voice information of a specific target through a collection module 11 of a voice recognition device 100 according to an embodiment of the present invention.

Specifically, the voice information collected in this step is an analog voice signal, and the analog voice signal needs to be converted into a digital voice signal before subsequent voice recognition processing can be performed. In addition, other environmental noises may be included in the voice information, so it is also necessary to preprocess the voice information to filter out the useless environmental noise and obtain an effective voice signal. The preprocessing includes noise suppression processing on the digital voice signal. and echo cancellation processing, the above-mentioned preprocessing can refer to the existing noise reduction processing technology.

Specifically, the voice signal after the preprocessing needs to extract the voice features of the specific target. In the embodiment of the present invention, the method of Mel-frequency Cepstral Coefficients (MFCC) is used to extract the voice features of the specific target. The pre-processed speech signal is cut into multiple frames (Frame blocking), pre-emphasis (Pre-emphasis) and windowing (Window) are performed for the part that needs to be emphasized, so as to obtain a clearer and clearer piece of speech. feature.

Step S102: Use the speech feature of the specific target as the input data of the discriminative training Hidden Vector State Model (HVS Model for short) and train it to obtain a specific target acoustic model, and store the specific target acoustic model. target acoustic model;

Specifically, in this step, the voice feature of the specific target is used as the input data to train the acoustic model. In the embodiment of the present invention, the latent vector state model is used and the training is performed by using the discriminative training method. The discriminative training does not maximize the training. The similarity of the acoustic corpus is the goal, and the classification (or identification) error is minimized to improve the identification rate.

Among them, the discriminative training is based on the maximum mutual information method (Maximum Mutual Information, referred to as MMI) for training, which can maximize the probability of the correct answer, and effectively reduce the probability of competitors, and increase the probability of correct answers and Competitor discrimination.

Specifically, storing the specific target acoustic model in this step refers to storing in a storage module 14 of the speech recognition apparatus 100 according to the embodiment of the present invention.

Step S103: Receive a voice message of a target to be tested and preprocess the voice message of the target to be tested to extract a voice feature of the target to be tested;

Specifically, in this step, the target to be measured refers to a user who wants to perform voice recognition and comparison. The target to be measured outputs a piece of voice information, and the target to be measured is collected by a collection module 11 of the speech recognition device 100 according to the embodiment of the present invention. 's voice message.

Specifically, in this step, the voice information of the target to be tested is preprocessed, and the voice features of the target to be tested are extracted. The flow of speech features.

Step S104: take the speech feature of the target to be measured as the input data of the latent vector state model trained by the discriminant type and train it to obtain an acoustic model of the target to be measured;

Specifically, in this step, the speech features of the target to be tested are used as input data to train the acoustic model. In the embodiment of the present invention, the latent vector state model is used and the training is performed by means of discriminative training. The discriminative training is based on the maximum mutual information. The maximum Mutual Information (MMI) is used as the criterion for training.

Step S105: Compare the correlation between the acoustic model of the target to be tested and the acoustic model of the specific target, and if there is a correlation between the two, use at least one language model to decode the speech feature of the target to be tested. , and judge whether to wake up according to the language decoding result.

Specifically, in this step, when the acoustic model of the target to be tested conforms to the acoustic model of the specific target, language decoding is performed, and if the acoustic model of the target to be tested does not conform to the acoustic model of the specific target, no action is performed. The speech feature of the test target is used as input data to train the language model. In the embodiment of the present invention, the language model includes a vocabulary model and a grammar model.

When the acoustic model of the target to be tested is determined to be the acoustic model of a specific target, it means that the target to be tested is a specific target at this time, so language decoding is performed to confirm whether the voice message of the target to be tested contains wake words. The speech features of the target to be tested are trained on thesaurus model and grammar model, and the content of the voice message of the target to be measured is obtained by parsing, and then it is judged whether the voice message content of the target to be tested contains a wake-up word. If it contains a wake-up word, the voice recognition wakes up. Start, if the wake-up word is not included, the voice recognition wake-up is not started.

Please refer to FIG. 2 , an embodiment of the present invention is an apparatus for waking up a specific target by voice recognition. A speech recognition device 100 includes an acquisition module 11, an extraction module 12, a training module 13, a storage module 14, a decoding module 15, a processor module 16, a registration module 17 and a Wireless communication module 18.

The acquisition module 11 is connected with the extraction module 12 and the registration module 17, wherein the acquisition module 11 is provided with a plurality of microphone arrays for receiving the voice information of the specific target and the target to be measured, and the collected voice information is required for analog voice signals. Convert the digital voice signal into a digital voice signal, and at the same time perform noise suppression processing and echo cancellation processing on the digital voice signal, and then transmit the processed digital voice information to the extraction module 12.

The definition of the specific target is the object that is used for the wake-up of the specific target according to the speech recognition of the present invention, and the definition of the target to be tested is the object that the speech recognition apparatus 100 performs speech recognition.

The voice message of the specific target includes a preset wake-up word.

The extraction module 12 is connected with the acquisition module 11, the training module 13 and the decoding module 15, and the extraction module 12 is used to receive the voice message processed by the acquisition module 11, and to extract the specific target and the target to be tested. The speech features are then sent to the training module 13 for acoustic model training or sent to the decoding module 15 for decoding.

The extraction of the voice features of the specific target and the target to be measured is to extract the voice features of the voice messages by using Mel-frequency Cepstral Coefficients (MFCC for short).

The training module 13 is connected to the extraction module 12 , the storage module 14 and the processor module 16 . The training module 13 is used to receive the voice features of the specific target and the target to be tested extracted by the extraction module 12, and use the voice features of the specific target and the target to be tested as the latent vector state model trained by the maximum mutual information method. Input data, and finally obtain the trained acoustic model, and perform different steps according to the specific target and the target to be tested. If it is a specific target, the acoustic model of the specific target is transmitted to the storage module 14 , and if it is a target to be measured, the acoustic model of the target to be measured is transmitted to the processor module 16 .

The storage module 14 is connected with the training module 13 , the processor module 16 and the registration module 17 . The storage module 14 is used to save the acoustic model of the specific target trained by the training module 13 . In the embodiment of the present invention, when the specific target performs the operation of the registration module 17, the acoustic model of the specific target trained by the training module 13 will be transmitted to the storage module 14 for saving. In addition, when the processor module 16 compares the acoustic model of the target to be tested with the specific target, the storage module 14 transmits the saved acoustic model of the specific target to the processor module 16 .

The decoding module 15 is connected to the extraction module 12 and the processor module 16 . The decoding module 15 is used to perform language decoding on the voice information of the target to be tested. More specifically, the extraction module 12 uses the voice feature of the target to be tested as the input data of the thesaurus model and the grammar model for training, and The results are passed to the processor module 16 .

The processor module 16 is connected with the training module 13 , the storage module 14 , the decoding module 15 and the wireless communication module 18 . The processor module 16 is used to compare the acoustic model of the specific target with the acoustic model of the target to be tested, and judge whether to activate the decoding module 15 to perform language decoding according to the comparison result of the two acoustic models, and more specifically. It means that when the training module 13 transmits the acoustic model of the target to be tested, the processor module 16 simultaneously obtains the acoustic model of the specific target from the storage module 14, and compares the two acoustic models in the processor module 16. right.

When it is confirmed that the acoustic model of the specific target is related to the acoustic model of the target to be tested, it means that the target to be tested is a specific target. Therefore, the speech message language decoding of the target to be tested is performed to determine whether it contains a wake-up word. Therefore, the processor module 16 will The decoding module 15 is activated, and the decoding module 15 performs language decoding.

The decoding module 15 obtains the speech features of the target to be tested from the extraction module 12, and returns the operation result of language decoding to the processor module 16, and the processor module 16 will determine the target according to the acoustic model of the target to be tested and After language decoding, it is determined whether the voice message of the target to be tested contains wake words.

When the processor module 16 obtains that the voice message of the target to be tested contains a wake-up word, the wake-up of the voice recognition device 100 is executed, otherwise, it is not executed.

The registration module 17 is connected with the acquisition module 11 and the storage module 14 . The registration module 17 is used to provide a specific target for registration of the speech recognition device 100, wherein the registration module 17 includes an activation element and a display element. When the specific target touches the activation element, the storage module 14 is activated at the same time, indicating that the acquisition is performed. The acoustic model of the voice message collected by the module 11 this time after being trained by the training module 13 needs to be saved to the storage module 14. In addition, when a specific target touches the activation element, the display element is activated to provide a specific target to confirm whether it is currently in the registration stage. .

In the embodiment of the present invention, the activation element is a button, and the display element is a light-emitting diode.

The wireless communication module 18 is connected to the processor module 16 . The wireless communication module 18 is used for connecting with the external communication after the processor module 16 confirms that the voice recognition device 100 is woken up successfully.

In the embodiment of the present invention, the wireless communication module 18 includes a Wi-Fi module or a Bluetooth module.

As described above, the method and device for voice recognition of the present invention for awakening a specific target adopts the latent vector state model of the discriminative training as the acoustic model, and the discriminative training using the maximum mutual information method not only maximizes the probability of occurrence of the correct answer, but also It will reduce the appearance probability of competitors, increase the ability to discriminate between their correct answers and competitors, and can quickly and accurately determine whether the target to be tested is a specific target, thereby achieving the function of awakening.

Claims (10)

1. a kind of method that speech recognition is used for specific target wake-up, is characterized in that, comprises the steps: S1: Receive a voice message of a specific target and preprocess the voice message of the specific target to extract a voice feature of the specific target; S2: take the speech feature of the specific target as the input data of the latent vector state model (HVS Model) of the discriminative training and train to obtain a specific target acoustic model, and store the specific target acoustic model; S3: Receive a voice message of a target to be tested and preprocess the voice message of the target to be tested to extract a voice feature of the target to be tested; S4: take the speech feature of the target to be measured as the input data of the latent vector state model trained by the discriminant type and train to obtain an acoustic model of the target to be measured; S5: Compare the correlation between the acoustic model of the target to be measured and the acoustic model of the specific target, and if the two are correlated, use at least one language model to perform language decoding on the speech feature of the target to be measured, And judge whether to wake up according to the language decoding result. 2 . The method according to claim 1 , wherein the voice message of the specific target and the voice message of the target to be tested include at least one wake-up word. 3 . 3 . The method according to claim 1 , wherein the preprocessing comprises: performing noise suppression processing and echo cancellation processing on the voice message. 4 . 4 . The method according to claim 1 , wherein the speech features are obtained by means of Mel cepstral coefficients (MFCC). 5 . 5. The method according to claim 1, wherein the discriminative training adopts the maximum mutual information method (MMI) for training. 6. The method according to claim 1, wherein the language model comprises a vocabulary model or a grammar model or a combination thereof. 7. The method according to claim 2, wherein the step of judging whether the wake-up of speech recognition is achieved according to the language decoding result, comprises: performing language decoding on the speech feature of the target to be tested; Determine whether the target voice message to be tested contains the wake-up word; If the wake-up word is included, the voice recognition wake-up is activated, and if the wake-up word is not included, the voice recognition wake-up is not activated. 8. A voice recognition device for waking up a specific target, wherein the device comprises: an acquisition module including a plurality of microphone arrays for receiving voice messages of a specific target and a target to be tested, wherein the voice messages include a wake-up word; an extraction module, connected to the acquisition module, for extracting the MFCC voice features in the voice messages of the specific target and the target to be tested; a training module, connected to the extraction module, for using the MFCC voice features in the voice messages of the specific target and the target to be tested as the input data of the latent vector state model trained by the maximum mutual information method, and Obtain the acoustic model of the specific target after training and the acoustic model of the target to be tested; a storage module, connected to the training module, for saving the acoustic model of the specific target that has been trained; a decoding module, connected to the extraction module, for performing language decoding on the voice message of the target to be tested; and a processor module, connected to the training module, the storage module and the decoding module, for comparing the acoustic model of the specific target in the storage module with the acoustic model of the target to be tested, and According to the comparison result, it is determined whether to activate the decoding module to perform language decoding of the voice message of the target to be tested, and to confirm whether the voice message of the target to be tested after the language decoding contains a wake-up word to wake up the device. 9. The device according to claim 8, further comprising a registration module, the registration module connects the acquisition module and the storage module, the The registration module is used to start saving the acoustic model of a specific target to the storage module. 10 . The apparatus of claim 8 , further comprising a wireless communication module, wherein the wireless communication module is used for external communication connection. 11 .

Images