CN107481718B - Voice recognition method, voice recognition device, storage medium and electronic equipment - Google Patents

Abstract

The embodiment of the present application discloses a speech recognition method, device, storage medium and electronic device. The method comprises obtaining first speech data; inputting the first speech data into a pre-constructed screening model for screening, obtaining a speech segment output by the screening model with set speech features filtered out; and recognizing the speech segment to obtain corresponding text. The above technical solution can effectively reduce the amount of calculation in the speech recognition process and improve the recognition speed.

Description

Speech recognition method, device, storage medium and electronic device

technical field

The embodiments of the present application relate to speech recognition technology, and in particular, to a speech recognition method, apparatus, storage medium, and electronic device.

Background technique

With the rapid development of science and technology applied to electronic equipment, electronic equipment has possessed powerful processing capabilities and has gradually become an indispensable and important tool for people's life, entertainment and work.

Taking a smartphone as an example, most of the current smartphones are equipped with a voice assistant function in order to allow the user to conveniently operate the smartphone even when driving a vehicle, carrying items or other scenarios where it is inconvenient to operate the smartphone through a touch screen. The voice data input by the user can be converted into text through the voice assistant. However, the current speech recognition solution has the defects of a large amount of calculation and a slow recognition speed when performing speech recognition.

SUMMARY OF THE INVENTION

Embodiments of the present application provide a speech recognition method, apparatus, storage medium and electronic device, which can reduce the amount of calculation in the speech recognition process and improve the recognition speed.

In a first aspect, an embodiment of the present application provides a speech recognition method, including:

obtain the first voice data;

The first voice data is input into a pre-built screening model for screening, and a voice segment output by the screening model that filters out the set voice features is obtained, wherein the screening model is composed of voice data with voice features that have no actual meaning added. sample training;

Recognize the speech segment to obtain the corresponding text.

In a second aspect, an embodiment of the present application further provides a speech recognition device, the device comprising:

a voice acquisition module for acquiring the first voice data;

A voice screening module, configured to input the first voice data into a pre-built screening model for screening, and obtain a voice segment output by the screening model to filter out the set voice features, wherein the screening model is not meaningful by adding The speech data samples of the speech features are obtained by training;

The speech recognition module is used for recognizing the speech segment to obtain the corresponding text.

In a third aspect, the embodiments of the present application further provide a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the speech recognition method described in the embodiments of the present application.

In a fourth aspect, the embodiments of the present application also provide an electronic device, including a voice collector for collecting first voice data, a memory, a processor, and a computer program stored in the memory and running on the processor, so When the processor executes the computer program, the speech recognition method according to the embodiment of the present application is implemented.

The present application provides a voice recognition solution, by acquiring first voice data; inputting the first voice data into a pre-built screening model for screening, and obtaining voice fragments output by the screening model that filter out the set voice features; and recognizing voice fragments get the corresponding text. In the above technical solution, before speech recognition, the acquired first speech data is input into the screening model. Since the training samples of the screening model are speech data samples with speech features with or without actual meaning added, the first speech data is input into the screening model for calculation, and the phonemes without actual meaning contained in the first speech data can be filtered out, and the result is obtained without any actual meaning. The actual meaning of the phoneme's speech fragment. Therefore, the data volume of the speech segment output by the screening model is smaller than the data volume of the first speech data. Recognizing the speech segment after the data amount is reduced can effectively reduce the calculation amount in the speech recognition process and improve the recognition speed.

Description of drawings

1 is a flowchart of a speech recognition method provided by an embodiment of the present application;

2 is a schematic diagram of the basic structure of a single neuron provided by an embodiment of the present application;

3 is a flowchart of another speech recognition method provided by an embodiment of the present application;

4 is a flowchart of another speech recognition method provided by an embodiment of the present application;

5 is a structural block diagram of a speech recognition device provided by an embodiment of the present application;

FIG. 6 is a structural block diagram of an electronic device provided by an embodiment of the present application.

Detailed ways

The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, the drawings only show some but not all the structures related to the present application.

Before discussing the exemplary embodiments in greater detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowchart depicts the steps as a sequential process, many of the steps may be performed in parallel, concurrently, or concurrently. Furthermore, the order of the steps can be rearranged. The process may be terminated when its operation is complete, but may also have additional steps not included in the figures. The processes may correspond to methods, functions, procedures, subroutines, subroutines, and the like.

In the related art, speech recognition methods generally include endpoint detection, feature extraction and matching operations. Among them, in order to accurately find the moment when speech starts and ends, a double-threshold detection algorithm is usually used. At the same time, the short-term zero-crossing rate and the short-term average energy are used to detect the voice data respectively, and the above-mentioned methods are comprehensively used to determine the endpoints (start time and end time) of the voice signal. The essence of feature extraction of voice data is to convert voice data from analog signal to digital signal, and represent voice data with a series of characteristic parameters reflecting the characteristics of voice data. Since the Mel Frequency Cepstral Coefficents (MFCC for short) are proposed based on the auditory model of the human ear, it is close to the auditory features of the human and can improve the recognition performance. Therefore, the feature extraction process is described by taking the extraction method of MFCC parameters as an example. The extraction method of MFCC parameters includes the following steps: using a preset window function, the audio signal is divided into frames according to a fixed frame length and frame shift, for example, the frame length can be 25ms, and the frame shift can be 10ms; The Liye transform (fastFourier transform, referred to as FFT) converts the time domain signal into the power spectrum of the signal; then a set of Mel filters are used to process the above frequency spectrum to obtain the Mel spectrum; the cepstrum analysis is performed on the Mel spectrum (including taking logarithm and discrete cosine transform) to obtain MFCC parameters. The MFCC parameters of each sound frame are taken as the speech feature vector sequence of the sound frame. Input the speech feature vector sequence of each sound frame into the hidden Markov model, and obtain the state output by the hidden Markov model that matches at least one sound frame (that is, compare the probability that the sound frame matches the state, and set the maximum The state corresponding to the probability as the state matching the sound frame). The three states are sequentially acquired to form phonemes, and the pronunciation of the word is determined according to the phonemes, thereby realizing speech recognition. However, the above-mentioned speech recognition solution cannot distinguish between phonemes with actual meaning and phonemes without actual meaning (such as "this", "that", "how do you say" and "that is to say", etc. in the user's expression habits), resulting in In the process of speech recognition, the amount of calculation is relatively large, and the speed of speech recognition is slow.

FIG. 1 is a flowchart of a speech recognition method provided by an embodiment of the present application. The method may be executed by a speech recognition apparatus, wherein the apparatus may be implemented by software and/or hardware, and may generally be integrated in an electronic device. As shown in Figure 1, the method includes:

Step 110: Acquire first voice data.

Wherein, the first voice data includes a voice signal input by a user. For example, the voice signal input by the user when using the voice input function in the SMS application. Another example is the voice signal input by the user when using the voice input function in the memo application. Another example is the voice signal input by the user when using the voice input function in the mail application. Another example is the voice signal input by the user when using the voice input function of the instant messaging application.

A voice collector is integrated on the electronic device, and the first voice data can be acquired through the voice collector. Among them, the voice collector includes a microphone, and wireless earphones such as Bluetooth earphones and infrared earphones. Exemplarily, taking a smartphone as an example, when the user enables the voice input function of the short message application, the short message input method can use voice input instead of manual input. The voice signal corresponding to the voice instruction is converted into text and displayed on the short message application interface. The first voice data can be obtained by preprocessing the voice signal corresponding to the voice instruction input by the user. The above preprocessing includes filtering, analog-to-digital conversion, and the like. It should be noted that because users often unconsciously bring oral expressions when speaking, it may lead to the fact that the first voice data includes "this", "that", "how do you say" and "that is to say", etc. vocabulary of meaning.

Step 120: Input the first voice data into a pre-built screening model for screening, and obtain a voice segment output by the screening model for filtering out the set voice feature.

Wherein, the screening model is obtained by training speech data samples with speech features without actual meaning added. Exemplarily, taking the screening model as a neural network model as an example, the training process of the screening model includes:

Model initialization, including setting the number of hidden layers, the number of nodes in the input layer, the hidden layer and the output layer, the connection weight between the layers, and the thresholds for initializing the hidden layer and the output layer, etc., to initially obtain the framework of the neural network model .

Speech recognition, calculate the output parameters of the hidden layer and the output layer of the output layer according to the formula included in the neural network model, and calculate the neural network according to the calculation result of the previous layer, the connection weight between the two layers and the external bias value of the own node. the output of the model.

Error calculation, using supervised learning to adjust the parameters in the neural network model. Obtain the voice data and the corresponding text entered by voice input in the user's historical short messages sent. Since the user confirms that the short message sent is data that does not have meaningless words after adjustment and conforms to the user's expression habits, it can be as a standard speech data sample. Correspondingly, the expected output corresponding to the speech data sample is the speech (or pronunciation) of the text corresponding to the above speech data. A training sample is obtained by adding voice features without actual meaning to the voice data sample. Among them, the way to obtain the speech features without actual meaning may be by counting the expression habits of a set number of sample groups, and analyzing and obtaining words with a high occurrence probability that have no actual meaning as the speech features. It may also be that the user selects the commonly used meaningless words by himself, or automatically counts the meaningless words commonly used by the user as speech features, and so on.

Calculate the actual output and the expected output of the neural network model, and obtain the error signal between the actual output and the expected output. Then, the connection weight and external bias value of each neuron in the neural network model are updated according to the error signal. Fig. 2 shows a schematic diagram of the basic structure of a single neuron provided by an embodiment of the present application. In Fig. 2, ω _i1 is the connection weight between neuron i and the neuron in the upper layer of the layer where it is located, which can also be understood as the input x ₁ The weight of ; θ _i is the external bias of this neuron. According to the prediction error of the network, the error back propagation in the neural network modifies the connection weight and external bias value of each neuron. Determine whether the algorithm iteration is over, and if so, complete the construction of the screening model.

Input the first voice data into the constructed screening model. For the path corresponding to the pronunciation without actual meaning in the first voice data, the connection weight is small, and the input parameters are between the hidden layers of the neural network model, or between the hidden layer and the output. In the process of layer transfer, since the input parameters are gradually reduced by multiplying the connection weight, the voice features (such as phonemes) that have no actual meaning in the first voice data are filtered out after multiple calculations. The output result of the screening model is to filter out speech segments that have no actual meaning.

Step 130: Recognize the voice segment to obtain the corresponding text.

Calculate the speech fragment and carry out the distance comparison with the preset reference template, take the pronunciation with the shortest distance in each sound frame in the speech fragment and the reference template as the pronunciation of this sound frame, and the combination of the pronunciation of each sound frame is the pronunciation of this speech fragment . After learning the voice of the voice segment, a preset dictionary can be queried to determine the text corresponding to the voice.

In the technical solution of this embodiment, the acquired first speech data is input into the screening model before speech recognition. Since the training samples of the screening model are speech data samples with speech features with or without actual meaning added, the first speech data is input into the screening model for calculation, and the phonemes without actual meaning contained in the first speech data can be filtered out, and the result is obtained without any actual meaning. The actual meaning of the phoneme's speech fragment. Therefore, the data volume of the speech segment output by the screening model is smaller than the data volume of the first speech data. Recognizing the speech segment after the data amount is reduced can effectively reduce the calculation amount in the speech recognition process and improve the recognition speed.

FIG. 3 is a flowchart of another speech recognition method provided by an embodiment of the present application. As shown in Figure 3, the method includes:

Step 301: Acquire first voice data.

Step 302: Determine whether the user corresponding to the first voice data is a registered user, if yes, go to Step 303, otherwise go to Step 306.

When the first voice data is detected, the camera of the electronic device is controlled to be turned on, and at least one frame of the user image is captured. By performing image processing and image recognition on the user image, it is determined whether the user who inputs the first voice data is a registered user. Wherein, whether the user who inputs the first voice data is a registered user may be determined by image matching. Exemplarily, when a user registers, a user image is obtained as a matching template. When the first voice data is detected, a user image is acquired, and the user image is matched with the matching template, so that it can be determined whether the user corresponding to the first voice data is a registered user.

Step 303: Acquire historical voice data of at least one registered user, and determine the speech rate and pause interval of each registered user according to the historical voice data.

When the user corresponding to the first voice data is a registered user, the historical voice data of the registered user is acquired. The historical voice data includes the user's historical call data, historical voice control data, and historical voice messages. The average speech rate and average pause interval of each registered user can be determined by analyzing historical speech data. Among them, the average speech rate and the average pause interval are calculated by weighting. The speech rate and pause interval of each registered user in different scenarios can also be further determined.

Step 304: Query a preset framing strategy set according to the speaking rate and pause habit, and determine a framing strategy corresponding to the registered user.

The framing strategy includes the selection of a window function, the value of frame length and the value of frame shift, and the framing strategy is associated with the language habits of different users. The framing strategy set is a set of framing strategies, in which the correspondence between the speech rate interval and the pause interval interval and the window function, frame length and frame shift is stored.

According to the speech rate and pause interval determined in the above steps, query the speech rate interval and pause interval interval stored in the framing strategy set, locate the interval corresponding to the speech rate and pause interval, and use the window function, frame length and frame shift corresponding to the interval as Register the framing policy of the current voice data input by the user.

Step 305 : Framing the first voice data according to the framing policy corresponding to the registered user to obtain at least two second voice data, and then go to step 307 .

Since the voice data only exhibits stationarity in a short period of time, it is necessary to divide the voice data into one short period of time, that is, sound frames.

Exemplarily, using the window function included in the framing strategy determined in the above steps, and processing the first voice data according to the frame shift included in the framing strategy to obtain at least two second voice data. The window length of the window function is equal to the frame length of the framing strategy. After obtaining at least two second voice data, go to step 307 . The division of the first voice data is related to the speech rate and pause interval of the registered user. Therefore, the frame length of the second voice data obtained after framing varies with the speech rate and pause interval of the registered user, and the frame length is not fixed. It can reduce the division of speech with practical significance and speech without practical significance into the same sound frame, which is beneficial to improve the efficiency of speech recognition.

Step 306: Framing the first voice data according to a default framing strategy to obtain at least two second voice data.

When the user corresponding to the first voice data is not a registered user, a default window function is used, and at least two second voice data are obtained by processing the first voice data according to the default frame shift. The window length of the window function is the default frame length. The frame length of the second voice data obtained after frame division is fixed, and it is often the case that voices with practical significance and voices with no practical significance are divided into one voice frame.

Step 307: Extract the first speech feature vector sequence corresponding to the second speech data.

Wherein, the first speech feature vector sequence includes MFCC features. The method of extracting MFCC features from the second speech data includes: filtering the spectrogram of the second speech data through a series of mel filters to obtain a mel spectrum; performing cepstral analysis on the mel spectrum to obtain a mel spectrum The Mel frequency cepstral coefficient is used as the dynamic feature vector of the input screening model, that is, the first speech feature vector sequence.

Step 308: After normalizing the first speech feature vector sequence, input a pre-built recurrent neural network model for screening.

Optionally, before the first speech feature vector sequence is input into the pre-built recurrent neural network model, the first speech feature vector sequence may also be normalized. Required steps. Among them, the normalization process is to map all the first speech feature vector sequences into numbers between [0, 1] or [-1, 1], which can eliminate the influence of different units and ranges of input data on speech recognition. , to reduce speech recognition errors.

After the first speech feature vector sequence is normalized, a pre-built neural network model is input for screening, wherein the neural network model is a recurrent neural network model.

Step 309: Obtain an output result of the cyclic neural network model, wherein the output result is a second speech feature vector sequence for filtering out phonemes that have no actual meaning.

Among them, a phoneme is the smallest unit in speech. According to the analysis of the pronunciation action in the syllable, an action constitutes a phoneme, and the phoneme includes vowels and consonants.

Since the recurrent neural network model is constructed by learning and training training samples with phonemes that have no actual meaning added, the output is a speech segment that filters out the phonemes that have no actual meaning. Therefore, when the first speech feature vector sequence is input After the recurrent neural network model, the output speech segment is the second speech feature vector sequence that filters out the phonemes that have no actual meaning.

Step 310: Determine whether the lengths of the second speech feature vector sequence and the preset reference template are equal, if so, go to Step 313, otherwise go to Step 311.

Obtain the length of the second speech feature vector sequence, and compare it with the length of the preset reference template. If the lengths are different, step 311 is executed. If the lengths are the same, step 313 is executed.

Step 311 , using the dynamic time warping algorithm to calculate the frame matching distance between the second speech feature vector sequence and the reference template.

Among them, dynamic time warping (DTW for short) is a method for measuring the similarity between two time series, and is mainly used in the field of speech recognition to identify whether two pieces of speech represent the same word.

Exemplarily, if the length of the second voice feature vector sequence is different from the preset reference template, the frame matching distance matrix between the second voice feature vector sequence and the reference template can be calculated by the DTW algorithm, and find out in the frame matching distance matrix. An optimal path, which is the path corresponding to the minimum matching distance.

Step 312 , determine the pronunciation corresponding to the minimum frame matching distance, and then perform step 314 .

Determine the speech in the reference template as the endpoint corresponding to the minimum frame matching distance and the second speech feature vector sequence, and use the speech in the reference template as the pronunciation of the second speech feature vector sequence.

Step 313: Directly match the second speech feature vector sequence with the reference template to determine the pronunciation corresponding to the speech segment.

If the length of the second speech feature vector sequence is the same as that of the preset reference template, the second speech feature vector sequence and the reference template are directly matched to determine the pronunciation corresponding to the speech segment.

Step 314: Match the corresponding text according to the pronunciation as a speech recognition result.

In the technical solution of this embodiment, before speech recognition, a framing strategy is determined according to the user's speech rate and pause interval, and a personalized framing strategy is used to perform framing processing on the first voice data, so as to realize personalized framing, It effectively reduces the number of sound frames that divide speech features with actual meaning from those without actual meaning in one frame. Inputting the first speech feature vector sequence corresponding to the framed second speech data into the screening model can further improve speech recognition efficiency.

FIG. 4 is a flowchart of another speech recognition method provided by an embodiment of the present application. As shown in Figure 4, the method includes:

Step 401: Determine whether the model update condition is satisfied, if yes, go to Step 402, otherwise, go to Step 408.

The model update condition may be that the system time reaches a preset time, or a preset update period is met. For example, if the model update condition is set to update the screening model at 12:00 pm every Friday, when it is detected that the system time is 12:00 pm on Friday, it is determined that the model update condition is currently satisfied. For another example, if the model update condition is set to be updated every 7 days, when it is detected that the time from the last model update satisfies the update cycle, it is determined that the model update condition is currently met.

Step 402: Acquire the sent short message input by voice, and/or the stored memo input by voice.

Get sent text messages with voice input, and stored memos. Since the short message sent by the user's confirmation can be regarded as the adjusted data which has no meaningless words and conforms to the user's expression habits, it can be regarded as a standard voice data sample. For the saved memo, it can also be regarded as data that has no meaningless words after adjustment and conforms to the user's expression habits, and can also be regarded as a standard voice data sample.

Pre-save the voice feature vector sequence of the voice data corresponding to the text content of the sent short message by voice input, and correspondingly save the voice data dictated by the user, and use the voice data dictated by the user as historical voice data. For example, taking the example of sending a short message by voice input, the voice data that the user dictates is "how to say about this problem, it is really difficult to solve", and the actual short message after processing is "about this problem, it is true Not easy to solve." Correspondingly store the voice feature vector sequence of the voice data that the user orally input, and the voice data corresponding to the short message actually sent out.

Step 403: Acquire a sequence of speech feature vectors of speech data corresponding to the text content of the short message and/or memo.

Obtain the speech feature vector sequence of the speech data in the body content of the sent SMS. Optionally, the voice feature vector sequence of the voice data in the text content of the stored memo may also be obtained.

Step 404: Acquire historical voice data of the short message and/or memo.

Obtain the content of the user's oral input corresponding to the sent short message as historical voice data. Optionally, the content of the user's oral input corresponding to the stored memo may also be obtained as historical voice data.

Step 405: Determine the personalized phoneme without actual meaning and the appearance position of the phoneme according to the historical speech data.

By analyzing historical voice data, it is possible to obtain the language habits of a user, that is, factors that have no actual meaning and where they appear. For example, during voice input, the user likes to say "how do you say" in the middle of a sentence without actual meaning.

Step 406: Add the phoneme to the speech feature vector sequence as a training sample according to the appearance position, and use the speech feature vector sequence as an expected output to train the screening model by using a supervised learning method.

Normalizing the training samples can eliminate the influence of different units and ranges of input data on speech recognition. At the same time, it is also beneficial to map the input data to the effective threshold of the activation function, reducing the network training error and network training time. .

Step 407: Adjust the parameters of the screening model according to the training results, where the parameters include connection weights and external bias values.

The network prediction error can be determined by analyzing the training samples and the expected output. According to the way the error is transmitted from the back (output layer) to the front (input layer) in the neural network model, the connection weight and external bias value of each neuron are modified respectively.

Step 408: Acquire the first voice data.

If the above-mentioned model update process has not ended when the first voice data is acquired, the first voice data is not recognized, and the user is prompted to update the screening model currently.

Step 409 : Input the first voice data into a pre-built screening model for screening, and obtain a voice segment output by the screening model for filtering out the set voice feature.

If the model update operation is not being performed when acquiring the first voice data, the first voice data is input into the screening model, and the first voice data is screened by the model to obtain voice fragments with voice features that have no actual meaning filtered out.

Step 410: Recognize the speech segment to obtain the corresponding text.

Step 411 , determine whether the text is command information, if yes, go to step 412 , otherwise go to step 413 .

The association between character combinations and command information is stored in advance through a whitelist. When the text corresponding to the speech segment is recognized, the whitelist is queried according to the text combination of the text. If the corresponding text combination is found in the whitelist, it is determined that the text corresponding to the voice segment represents command information, and step 412 is executed. If the corresponding text combination is not found in the whitelist, the user is prompted to select whether it is command information. If the user selects the text corresponding to the voice segment to represent the command information, the text combination determined by the user as the command information is added to the white list, and step 412 is executed. If the text corresponding to the voice segment selected by the user does not represent the command information, step 413 is executed.

Step 412: Execute the operation corresponding to the command information.

Step 413: Display the text in the user interface.

In the technical solution of this embodiment, when the update condition of the screening model is satisfied, the sent short message inputted by voice and/or the stored memo inputted by voice are used as training samples to train the screening model. The output of the screening model is adapted to the changing expression habits of users, and the false recognition rate and missed detection rate are effectively reduced.

FIG. 5 is a structural block diagram of a speech recognition apparatus provided by an embodiment of the present application. The apparatus may be implemented in software and/or hardware, and is typically integrated in electronic equipment. As shown in Figure 5, the apparatus may include:

The voice acquisition module 510 is configured to acquire first voice data.

The voice screening module 520 is configured to input the first voice data into a pre-built screening model for screening, and obtain a voice segment output by the screening model to filter out the set voice features, wherein the screening model is determined by adding The speech data samples of the meaning of the speech feature are obtained by training.

The speech recognition module 530 is used for recognizing the speech segment to obtain the corresponding text.

An embodiment of the present application provides a speech recognition apparatus, which inputs the acquired first speech data into a screening model before speech recognition. Since the training samples of the screening model are speech data samples with speech features with or without actual meaning added, the first speech data is input into the screening model for calculation, and the phonemes without actual meaning contained in the first speech data can be filtered out, and the result is The actual meaning of the phoneme's speech fragment. Therefore, the data volume of the speech segment output by the screening model is smaller than the data volume of the first speech data. Recognizing the speech segment after the data amount is reduced can effectively reduce the calculation amount in the speech recognition process and improve the recognition speed.

Optionally, also include:

a user judgment module, configured to judge whether the user corresponding to the first voice data is a registered user when the first voice data is detected;

and, also includes:

The framing module is used to determine a corresponding framing strategy according to the judgment result before inputting the first voice data into a pre-built screening model for screening, and perform framing on the first voice data according to the framing strategy , obtain at least two second voice data;

Wherein, the framing strategy includes the selection of a window function, the value of frame length and the value of frame shift, and the framing strategy is associated with the language habits of different users.

Optionally, the framing module is specifically used for:

Obtain historical voice data of at least one registered user, and determine the speech rate and pause interval of each registered user according to the historical voice data;

According to the speech rate and the pause habit, a preset frame segmentation strategy set is inquired, and a frame segmentation strategy corresponding to the registered user is determined.

Optionally, the voice screening module 520 is specifically used for:

extracting the first speech feature vector sequence corresponding to the second speech data;

After normalizing the first speech feature vector sequence, input a pre-built cyclic neural network model for screening;

An output result of the cyclic neural network model is acquired, wherein the output result is a second speech feature vector sequence from which phonemes without actual meaning are filtered out.

Optionally, the speech recognition module 530 is specifically used for:

Determine whether the length of the second speech feature vector sequence is equal to the preset reference template;

When not equal, adopt the dynamic time warping algorithm to calculate the frame matching distance between the second speech feature vector sequence and the reference template;

The pronunciation corresponding to the minimum frame matching distance is determined, and the text whose pronunciation is matched is used as the speech recognition result.

Optionally, also include:

a word processing module, configured to determine whether the text is command information after recognizing the voice fragment to obtain the corresponding text;

If so, execute the operation corresponding to the command information;

If not, the text is displayed in the user interface.

Optionally, also include:

The model update module is used to obtain the sent short messages input by voice and/or stored memos input by voice when the model update conditions are met;

Obtain the speech feature vector sequence of the speech data corresponding to the text content of the short message and/or the memo;

Obtain the historical voice data of the text messages and/or memos;

Determine the personalized phoneme without actual meaning and the appearance position of the phoneme according to the historical voice data;

Add the phoneme to the speech feature vector sequence as a training sample according to the appearance position, and use the speech feature vector sequence as an expected output to train the screening model by using a supervised learning method;

The parameters of the screening model are adjusted according to the training results, and the parameters include connection weights and external bias values.

Embodiments of the present application further provide a storage medium containing computer-executable instructions, where the computer-executable instructions are used to execute a speech recognition method when executed by a computer processor, and the method includes:

obtain the first voice data;

Inputting the first voice data into a pre-built screening model for screening, and obtaining a voice segment output by the screening model to filter out the set voice features, wherein the screening model is composed of voice data with voice features that have no actual meaning added. sample training;

Recognize the speech segment to obtain the corresponding text.

storage medium - any of various types of memory devices or storage devices. The term "storage medium" is intended to include: installation media, such as CD-ROMs, floppy disks, or tape devices; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc. ; non-volatile memory, such as flash memory, magnetic media (eg hard disk or optical storage); registers or other similar types of memory elements, etc. The storage medium may also include other types of memory or combinations thereof. In addition, the storage medium may be located in the first computer system in which the program is executed, or may be located in a second, different computer system connected to the first computer system through a network such as the Internet. The second computer system may provide program instructions to the first computer for execution. The term "storage medium" may include two or more storage media that may reside in different locations (eg, in different computer systems connected by a network). The storage medium may store program instructions (eg, embodied as a computer program) executable by one or more processors.

Of course, a storage medium containing computer-executable instructions provided by the embodiments of the present application, the computer-executable instructions of which are not limited to the above-mentioned voice recognition operations, and can also execute any of the voice recognition methods provided in any of the embodiments of the present application. related operations.

The embodiment of the present application provides an electronic device, in which the voice recognition apparatus provided by the embodiment of the present application can be integrated. Among them, electronic equipment includes smart phones, tablet computers, handheld game consoles, notebook computers and smart watches. FIG. 6 is a structural block diagram of an electronic device provided by an embodiment of the present application. As shown in FIG. 6 , the electronic device may include: a memory 601 , a central processing unit (Central Processing Unit, CPU) 602 (also called a processor, hereinafter referred to as CPU), a voice collector 606 and a touch screen 611 . The touch screen 611 is used to convert user operations into electrical signals and input them to the processor, and display a visual output signal; the voice collector 606 is used to collect the first voice data; the memory 601 is used to Store computer programs; the CPU 602 reads and executes the computer programs stored in the memory 601 . The CPU 602 implements the following steps when executing the computer program: obtaining first voice data; inputting the first voice data into a pre-built screening model for screening, and obtaining the output of the screening model to filter out the set voice features. A speech segment, wherein the screening model is obtained by training speech data samples with speech features without actual meaning added; the speech segment is recognized to obtain the corresponding text.

The electronic device further includes: a peripheral interface 603, an RF (Radio Frequency, radio frequency) circuit 605, a power management chip 608, an input/output (I/O) subsystem 609, other input/control devices 610, and an external port 604, These components communicate through one or more communication buses or signal lines 607 .

It should be understood that the illustrated electronic device 600 is only an example of an electronic device, and that the electronic device 600 may have more or fewer components than those shown in the figures, two or more components may be combined, Or can have different component configurations. The various components shown in the figures may be implemented in hardware, software, or a combination of hardware and software, including one or more signal processing and/or application specific integrated circuits.

The following describes the electronic device integrated with the voice recognition device provided in this embodiment in detail, and the electronic device takes a mobile phone as an example.

Memory 601, which can be accessed by CPU 602, peripheral interface 603, etc., said memory 601 can include high-speed random access memory, and can also include non-volatile memory, such as one or more disk storage devices, flash memory devices , or other volatile solid-state storage devices.

A peripheral interface 603 that can connect the input and output peripherals of the device to the CPU 602 and the memory 601 .

I/O subsystem 609 , which may connect input and output peripherals on the device, such as touch screen 611 and other input/control devices 610 , to peripherals interface 603 . The I/O subsystem 609 may include a display controller 6091 and one or more input controllers 6092 for controlling other input/control devices 610 . Wherein, one or more input controllers 6092 receive electrical signals from or send electrical signals to other input/control devices 610, which may include physical buttons (push buttons, rocker buttons, etc. ), dial pad, slide switch, joystick, click wheel. Notably, the input controller 6092 can be connected to any of the following: a keyboard, an infrared port, a USB interface, and a pointing device such as a mouse.

Display controller 6091 in I/O subsystem 609 receives electrical signals from touch screen 611 or sends electrical signals to touch screen 611 . The touch screen 611 detects the contact on the touch screen, and the display controller 6091 converts the detected contact into interaction with the user interface objects displayed on the touch screen 611, that is, to realize human-computer interaction, and the user interface objects displayed on the touch screen 611 can be run. Icons for games, icons for connecting to the corresponding network, etc. It is worth noting that the device may also include a light mouse, which is a touch-sensitive surface that does not display visual output, or an extension of the touch-sensitive surface formed by a touch screen module.

The RF circuit 605 is mainly used to establish the communication between the mobile phone and the wireless network (ie, the network side), and realize the data reception and transmission between the mobile phone and the wireless network. Such as sending and receiving text messages, e-mails, etc. Specifically, the RF circuit 605 receives and transmits RF signals, also known as electromagnetic signals, the RF circuit 605 converts electrical signals into electromagnetic signals or converts electromagnetic signals into electrical signals, and communicates with communication networks and other devices through the electromagnetic signals to communicate. RF circuitry 605 may include known circuitry for performing these functions including, but not limited to, antenna systems, RF transceivers, one or more amplifiers, tuners, one or more oscillators, digital signal processors, CODECs ( COder-DECoder, codec) chip set, subscriber identity module (Subscriber Identity Module, SIM) and so on.

The voice collector 606 includes a microphone, and wireless earphones such as Bluetooth earphones and infrared earphones, and is mainly used for receiving audio data and converting the audio data into electrical signals.

The power management chip 608 is used for power supply and power management for the hardware connected to the CPU 602, the I/O subsystem and the peripheral interface.

The electronic device provided by the embodiment of the present application inputs the acquired first voice data into the screening model before voice recognition. Since the training samples of the screening model are speech data samples with speech features with or without actual meaning added, the first speech data is input into the screening model for calculation, and the phonemes without actual meaning contained in the first speech data can be filtered out, and the result is obtained without any actual meaning. The actual meaning of the phoneme's speech fragment. Therefore, the data volume of the speech segment output by the screening model is smaller than the data volume of the first speech data. Recognizing the speech segment after the data amount is reduced can effectively reduce the calculation amount in the speech recognition process and improve the recognition speed.

The speech recognition apparatus, storage medium, and electronic device provided in the above embodiments can execute the speech recognition method provided by any embodiment of the present application, and have corresponding functional modules and beneficial effects for executing the method. For technical details not described in detail in the foregoing embodiments, reference may be made to the speech recognition method provided by any embodiment of the present application.

Note that the above are only preferred embodiments of the present application and applied technical principles. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the protection scope of the present application. Therefore, although the present application has been described in detail through the above embodiments, the present application is not limited to the above embodiments, and can also include more other equivalent embodiments without departing from the concept of the present application. The scope is determined by the scope of the appended claims.

Claims (9)

1. a speech recognition method, is characterized in that, comprises: obtain the first voice data; When detecting the first voice data, determine whether the user corresponding to the first voice data is a registered user; A corresponding framing strategy is determined according to the judgment result, and the first voice data is divided into frames according to the framing strategy to obtain at least two second voice data, wherein the framing strategy includes selection of a window function, frame The long value and the frame shift value, and the framing strategy is associated with the language habits of different users; Extracting the first voice feature vector sequence corresponding to the second voice data, inputting the first voice feature vector sequence into a pre-built screening model for screening, and obtaining a voice segment output by the screening model to filter out the set voice features , wherein the screening model is obtained by adding voice data samples without actual meaning of voice features; Recognize the speech segment to obtain the corresponding text. 2. method according to claim 1, is characterized in that, according to judgment result to determine corresponding framing strategy, comprising: Obtain historical voice data of at least one registered user, and determine the speech rate and pause interval of each registered user according to the historical voice data; According to the speech rate and the pause interval, a preset framing strategy set is inquired, and a framing strategy corresponding to the registered user is determined. 3. The method according to claim 1, wherein, extracting the first voice feature vector sequence corresponding to the second voice data, and inputting the first voice feature vector sequence into a pre-built screening model for screening, comprising: : extracting the first speech feature vector sequence corresponding to the second speech data; After normalizing the first speech feature vector sequence, input a pre-built cyclic neural network model for screening; An output result of the cyclic neural network model is acquired, wherein the output result is a second speech feature vector sequence from which phonemes without actual meaning are filtered out. 4. The method according to claim 3, characterized in that, comprising: recognizing the speech segment to obtain corresponding text, comprising: Determine whether the length of the second speech feature vector sequence is equal to the preset reference template; When not equal, adopt the dynamic time warping algorithm to calculate the frame matching distance between the second speech feature vector sequence and the reference template; The pronunciation corresponding to the minimum frame matching distance is determined, and the text whose pronunciation is matched is used as the speech recognition result. 5. The method according to claim 1, characterized in that, after recognizing the speech segment to obtain the corresponding text, the method further comprises: Determine whether the text is command information; If so, execute the operation corresponding to the command information; If not, the text is displayed in the user interface. 6. The method according to any one of claims 1 to 5, characterized in that, further comprising: When the model update conditions are met, obtain the sent text messages input by voice, and/or stored memos input by voice; Obtain the speech feature vector sequence of the speech data corresponding to the text content of the short message and/or the memo; Obtain the historical voice data of the text messages and/or memos; Determine the personalized phoneme without actual meaning and the appearance position of the phoneme according to the historical voice data; Add the phoneme to the speech feature vector sequence as a training sample according to the appearance position, and use the speech feature vector sequence as an expected output to train the screening model by using a supervised learning method; The parameters of the screening model are adjusted according to the training results, and the parameters include connection weights and external bias values. 7. A voice recognition device, comprising: a voice acquisition module for acquiring the first voice data; a user judgment module, configured to judge whether the user corresponding to the first voice data is a registered user when the first voice data is detected; A framing module, configured to determine a corresponding framing strategy according to the judgment result, and perform framing on the first voice data according to the framing strategy to obtain at least two second voice data, wherein the framing strategy includes The selection of window function, the value of frame length and the value of frame shift, and the framing strategy is associated with the language habits of different users; The voice screening module is used for extracting the first voice feature vector sequence corresponding to the second voice data, inputting the first voice feature vector sequence into a pre-built screening model for screening, and obtaining the filtering settings output by the screening model. A voice segment with a defined voice feature, wherein the screening model is obtained by adding voice data samples with no actual meaning of voice feature training; The speech recognition module is used for recognizing the speech segment to obtain the corresponding text. 8. A computer-readable storage medium on which a computer program is stored, characterized in that, when the program is executed by a processor, the speech recognition method according to any one of claims 1 to 6 is implemented. 9. An electronic device, comprising a voice collector, a memory, a processor and a computer program that is stored on the memory and can be run on the processor for collecting the first voice data, wherein the processor executes all The voice recognition method according to any one of claims 1 to 6 is realized when the computer program is implemented.