CN114974214A - Voice keyword detection method and device, electronic equipment and storage medium - Google Patents

Abstract

The application relates to a method and a device for detecting a voice keyword, electronic equipment and a storage medium. The method comprises the following steps: acquiring a detection result of a voice detection model to-be-detected audio, wherein the detection result is used for representing the stage pronunciation duration characteristic of a first word, the first word is a word which is recognized by the voice detection model from the to-be-detected audio and is intended as a target keyword in a keyword set, and a plurality of keywords which are selected in advance are stored in the keyword set; and under the condition that the stage pronunciation duration characteristics of the first word are matched with the stage pronunciation duration characteristics of the target keyword, determining the first word as the target keyword. According to the method, under the condition that the target keyword possibly exists in the audio to be detected, secondary judgment is carried out according to the stage pronunciation duration characteristics of the first word to be the target keyword, so that the accuracy of the detection result is further improved, and the problem that the detection result of the keyword retrieval technology in the related technology is inaccurate is solved.

Description

Method, device, electronic device and storage medium for detecting speech keywords

technical field

The present application relates to the technical field of speech recognition, and in particular, to a method, device, electronic device and storage medium for detecting speech keywords.

Background technique

Voice keyword retrieval technology is a special voice recognition technology, which is to identify and determine some specific keywords in continuous speech, or search for keywords that users are interested in, and the retrieval system returns the specific keywords that appear in the audio. location or time.

At present, with regard to keyword retrieval technology, the method of machine learning or deep neural network is used to input the audio to be detected, so that the machine can automatically predict whether the keyword exists in the audio and the corresponding existence position (start and end time point) through the model, but the machine predicts The results are often affected by the environment or unclear pronunciation, resulting in errors in the prediction results.

For the above problems, no effective solution has been proposed yet.

SUMMARY OF THE INVENTION

The present application provides a method, device, electronic device and storage medium for detecting speech keywords, so as to solve the problem of inaccurate detection results of keyword retrieval technology in the related art.

In a first aspect, the present application provides a method for detecting speech keywords, including: acquiring a detection result of the audio to be detected by a speech detection model, wherein the detection result is used to represent the phased pronunciation duration feature of the first word, so The first word is a word recognized by the speech detection model from the audio to be detected and intended to be a target keyword in a keyword set, and the keyword set stores a plurality of pre-selected keywords; When the phased pronunciation duration feature of the first word matches the phased pronunciation duration characteristic of the target keyword, the first word is determined to be the target keyword.

In a second aspect, the present application provides an apparatus for detecting speech keywords, comprising: an acquisition unit configured to acquire a detection result of the audio to be detected by a speech detection model, wherein the detection result is used to indicate the phase of the first word Pronunciation duration feature, the first word is a word recognized by the speech detection model from the audio to be detected and intended to be a target keyword in a keyword set, and the keyword set stores a number of pre-selected words. A determination unit, configured to determine that the first word is the target keyword when the phased pronunciation duration feature of the first word matches the phased pronunciation duration characteristic of the target keyword .

In a third aspect, an electronic device is provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used for storing computer programs; processing The device is used to implement the steps of any one of the voice keyword detection methods when executing the program stored in the memory.

In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, implements the steps of any one of the voice keyword detection methods.

The technical solution of the present application can be applied to the field of deep learning technology based on speech recognition and processing. Compared with the prior art, the above-mentioned technical solutions provided in the embodiments of the present application have the following advantages:

In the method provided by the embodiment of the present application, first, the detection result of the audio to be detected by the speech detection model is obtained, wherein the detection result is used to represent the staged pronunciation duration feature of the first word in the audio to be detected that is to be the target keyword, and then In the case that the staged pronunciation duration feature of the first word matches the staged pronunciation duration feature of the target keyword, the first word is determined as the target keyword. The method first uses a speech detection model to detect the audio to be detected, and in the case that the target keyword may exist in the audio to be detected, it is judged whether the first word is the target according to the phased pronunciation duration feature of the first word to be the target keyword. keyword, the method of this embodiment performs a secondary judgment on the detection of speech keywords, thereby further increasing the accuracy of the detection result, and further solving the problem of inaccurate detection results of the keyword retrieval technology in the related art.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description serve to explain the principles of the invention.

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. In other words, on the premise of no creative labor, other drawings can also be obtained from these drawings.

1 is a schematic flowchart of a method for detecting a speech keyword according to an embodiment of the present application;

2 to 4 are schematic flowcharts of a method for obtaining a detection result by using a deep neural network according to an embodiment of the present application;

5 is a schematic flowchart of a method for detecting speech keywords provided by an embodiment of the present application;

6 is a schematic structural diagram of a device for detecting speech keywords provided by an embodiment of the present application;

FIG. 7 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments It is a part of the embodiments of this application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present application.

According to an aspect of the embodiments of the present application, a method for detecting speech keywords is provided. Optionally, in this embodiment, the foregoing method for detecting speech keywords may be applied to a hardware environment composed of a terminal and a server. The server is connected to the terminal through the network, which can be used to provide services for the terminal or the client installed on the terminal, and a database can be set on the server or independent of the server to provide data storage services for the server.

The above network may include, but is not limited to, at least one of the following: wired network, wireless network. The wired network may include but is not limited to at least one of the following: a wide area network, a metropolitan area network, and a local area network, and the wireless network may include but not limited to at least one of the following: WIFI (Wireless Fidelity, wireless fidelity), Bluetooth. The terminal may not be limited to a PC, a mobile phone, a tablet computer, or the like.

The method for detecting speech keywords in this embodiment of the present application may be executed by a server, a terminal, or jointly executed by a server and a terminal. Wherein, the terminal executing the voice keyword detection method of the embodiment of the present application may also be executed by a client installed on the terminal.

Taking the method for detecting speech keywords in this embodiment performed by a server as an example, FIG. 1 is a schematic flowchart of a method for detecting speech keywords according to an embodiment of the present application. As shown in Figure 1, the method includes the following steps:

Step S201, obtaining the detection result of the audio to be detected by the voice detection model, wherein the above-mentioned detection result is used to represent the phased pronunciation duration feature of the first word, and the above-mentioned first word is identified by the above-mentioned voice detection model from the above-mentioned audio to be detected. A word intended to be a target keyword in a keyword set, where multiple pre-selected keywords are stored in the keyword set;

In this embodiment, the above-mentioned phased pronunciation duration feature may be the duration of initials and finals of each character in the first word. For example, the initials and finals corresponding to potatoes are "t-u d-ou", and the phased pronunciation duration features of potatoes are The pronunciation duration of t, the pronunciation duration of u, the pronunciation duration of d, and the pronunciation duration of ou. In practical applications, due to the different frequencies of people speaking, the pronunciation duration of the initials and finals of the same word may be quite different, but the difference in the ratio of the durations of the initials and finals is small. In order to further improve the accuracy of the detection results, the above phased The pronunciation duration feature may be the ratio of the duration of the initial consonant to the duration of the final, for example, the ratio of the pronunciation duration of t to the pronunciation duration of u. The above-mentioned staged pronunciation duration feature may also be the normalized initial consonant duration and final consonant duration, thereby reducing errors caused by different speaking frequencies.

The above-mentioned first word is the word that the above-mentioned speech detection model recognizes from the above-mentioned audio to be detected and is intended to be the target keyword in the keyword set, which means that the above-mentioned first word is recognized as the target keyword by the speech detection model, but due to the speech detection model. Misrecognition occurs, and the first word may not actually be the target keyword. For example, the target keyword is potatoes, and the first word is actually copper beans, but is recognized as potatoes by the speech detection model.

The above keyword set includes a plurality of keywords, such as open, close, mute and so on. The above target keyword may be one or more than one.

Step S202, in the case that the phased pronunciation duration feature of the first word matches the phased pronunciation duration feature of the target keyword, determine that the first word is the target keyword.

In this embodiment, the staged pronunciation duration feature of the target keyword may be the duration of the initial consonant and the duration of the final consonant of each character, or the ratio of the duration of the initial consonant to the duration of the final vowel, or the normalized duration of the initial consonant and the duration of the final consonant. Rhyme duration. The phased pronunciation duration feature of the above target keyword is obtained by calculating the average duration of each type of initials or the duration of each type of finals of a large number of marked speeches in the speech database. If the difference between the staged pronunciation duration feature of the first word and the staged pronunciation duration feature of the target keyword is within a certain range, it means that the difference between the above-mentioned first word and the above-mentioned target keyword is small, and the first word is the target key. word.

In order to obtain the detection result quickly, in an embodiment, in the case that the phased pronunciation duration feature of the above-mentioned first word matches the phased pronunciation duration characteristic of the above-mentioned target keyword, the above-mentioned first word is determined to be the above-mentioned target keyword. , comprising: determining a first difference between the staged pronunciation duration feature of the first word and the staged pronunciation duration feature of the target keyword; and determining the first word as the target keyword according to the first difference.

In this embodiment, it is first necessary to determine the first difference between the staged pronunciation duration feature of the first word and the staged pronunciation duration feature of the target keyword, and it can be determined whether the first word is the target according to the first difference Key words, this detection method is relatively simple and does not require complicated calculation process, so that the detection result can be obtained quickly.

In an embodiment, determining the first word as the target keyword according to the first difference value includes: in the case that the first difference value is smaller than a first predetermined value, determining the first word as the target keyword or, inputting the first difference into a two-class network to obtain the probability of the first word, wherein the probability of the first word is used to represent the probability that the first word is the target keyword, the above The two-class network is obtained by using the first sample to train the initial two-class network. The above-mentioned first sample is the phased pronunciation duration feature and reference stage of the words in the speech database to be the target keywords in the keyword set obtained after marking the difference between the sexual pronunciation duration features; when the probability of the first word is greater than the second predetermined value, it is determined that the first word is the target keyword.

In this embodiment, the comparison can be made directly according to the size of the first difference. If the first difference is smaller than the first predetermined value, it means that the phased pronunciation duration feature of the first word is different from the phased pronunciation duration characteristic of the target keyword. If it is smaller, the first word is the target keyword. It is also possible to input the first difference into a two-class network, and use the two-class network to obtain the probability that the first word is the target keyword. When the probability is greater than the second predetermined value, the above-mentioned first word is the above-mentioned target keyword, thereby further. Improves the accuracy of speech keyword detection results.

The labels of the above two-category network are 0 and 1, which correspond to the two categories of non-keywords and time-keywords, respectively. The above-mentioned binary classification network generally adopts machine learning algorithms such as Logistic Regression (LR), Support Vector Machine (SVM) and the like to train the secondary classification network. The parameters of the Logistic binary classification network are obtained through the training of the first sample, and the above-mentioned first sample is the phased pronunciation duration feature and the reference phased pronunciation of the words in the speech database to be the target keywords in the keyword set The difference between duration features, marked as 0 and 1.

得到上述第一差值，其中，MSE表示上述第一差值，

表示上述目标关键词的声母时长均值或韵母时长均值，y_i表示上述第一词语的声母时长或韵母时长，n表示上述第一词语的声母数量和韵母数量之和，i为大于等于1且小于等于n的整数。In an embodiment, determining the first difference between the phased pronunciation duration feature of the first word and the phased pronunciation duration characteristic of the target keyword includes: using

Obtain the above-mentioned first difference, wherein MSE represents the above-mentioned first difference,

Represents the average duration of initials or finals of the target keyword, y _i represents the duration of initials or finals of the first word, n represents the sum of the number of initials and the number of finals of the first word, i is greater than or equal to 1 and less than An integer equal to n.

Because the loss function usually adopts the mean square error, in this embodiment, the above formula is used to calculate the mean square error between the initial and final duration of the first word and the initial and final duration of the target keyword, which is conducive to the subsequent adoption of the model method to judge.

For example, if the target keyword is potatoes, the corresponding combination of initials and finals is "t-u d-ou", the average duration of these four initials and finals is x, y, z, m, and the speech detection model detects that the first word is actually "tongdou" , the corresponding combination of consonants and finals is "t-ong d-ou", and the durations a, b, c, and d of the four consonants are obtained, and the mean square error of the four consonants is calculated.

In one embodiment, determining the first difference between the phased pronunciation duration feature of the first word and the phased pronunciation duration characteristic of the target keyword includes: determining the first difference between the duration of the initials and the duration of the finals of the first word A ratio and a second ratio of the duration of the initial consonant and the duration of the final of the target keyword; the first difference is obtained according to the difference between the first ratio and the second ratio.

As mentioned in the above content, the different frequencies of people's speech lead to large differences in the characteristics of the stage pronunciation duration. In order to avoid this difference and further improve the accuracy of the detection results, in this embodiment, the difference between the duration of the initial consonant and the duration of the final vowel is calculated. The ratio is then determined according to the difference between the first ratio of the first word and the second ratio of the target keyword.

The ratio of the duration of the initial consonant and the duration of the final vowel of the above-mentioned first word or target keyword can be the ratio of the duration of the initial consonant and the duration of the final vowel of each character in the first word or the target keyword, or the ratio of the duration of each character. The ratio of the duration of initials to the duration of finals. If it is the ratio of the duration of the initials to the duration of the finals of each character, there will be multiple ratios. Calculate the error of the ratio between the first word and each character in the target keyword, and then calculate the mean squared error to get the first difference. .

In one embodiment, acquiring the detection result of the audio to be detected by the speech detection model includes: using a speech recognition model to decode the audio to be detected to obtain a sequence of candidate words, wherein the speech recognition model is used to decode the audio to be detected. The audio is converted into text data; the word grid is generated according to the above-mentioned candidate word sequence and the corresponding backtracking path and matching score; Perform retrieval to obtain the above-mentioned first word; according to the pinyin of the above-mentioned first word, obtain the staged pronunciation duration feature of the above-mentioned first word.

In this embodiment, a speech recognition model is used to first decode the audio to be detected to obtain a word grid, and then perform a quick search on the word grid to retrieve the first word and the corresponding phased pronunciation duration feature. The accuracy of the method is higher, but the running time is longer because of the need for decoding.

In one embodiment, acquiring the detection result of the audio to be detected by the speech detection model includes: extracting the speech feature of each frame of the audio to be detected, and obtaining the speech feature vector of each frame; inputting the speech feature vector of each frame into a Perform classification in the deep neural network model to obtain multiple labels and the probability of each label, wherein the above labels are identified from the above-mentioned audio to be detected and are intended to be the target keywords in the keyword set or the part of the above-mentioned target keywords. word; when the probability of the label is greater than the third predetermined value, the label is determined as the first word; according to the initial speech frame and the end speech frame of the first word, the phased pronunciation duration feature of the first word is determined.

In this embodiment, a neural network method is used for keyword retrieval, which does not require decoding of the audio to be detected, thereby reducing the computation time and memory usage during runtime. Figures 2 to 4 are for the application A schematic flowchart of a method for obtaining a detection result by using a deep neural network provided by the embodiment, as shown in FIG. 2 , first, feature extraction is performed on the audio to be detected, and a speech feature vector of the audio to be detected is extracted, and the audio to be detected includes the key phrase "okay, google" ", then input the speech feature vector into the deep neural network as shown in Figure 3, output 3 labels "okay", "google" and ", filler", and generate the frame-level posterior probability as shown in Figure 4 , the post-processing module combines these probabilities to provide the final confidence score for that window.

The technical solution of the present application will be described in further detail below in conjunction with specific embodiments, as shown in Figure 5:

Step 1, using a speech recognition model to decode the audio to be detected to obtain a sequence of candidate words, wherein the speech recognition model is used to convert the audio to be detected into text data, and then according to the sequence of candidate words and the corresponding backtracking path and The matching score generates a word grid, and then uses the pinyin of a plurality of keywords in the above keyword set to retrieve the word pinyin in the above word grid to obtain the above-mentioned first word, and according to the pinyin of the above-mentioned first word , to obtain the duration of the initial consonant and the duration of the final vowel of each character in the first word above.

Step 2: Calculate the average duration of each type of initials or durations of each type of finals of a large number of marked voices in the speech database, and then find the durations of initials and finals of each character in the target keyword.

计算第一词语中的声母时长和韵母时长与目标关键词的声母时长和韵母时长的均方误差。Step three, use

Calculate the mean squared error between the initials and finals in the first word and the initials and finals of the target keyword.

For example, if the target keyword is potatoes, the corresponding combination of initials and finals is "t-u d-ou", the average duration of these four initials and finals is x, y, z, m, and the speech detection model detects that the first word is actually "tongdou" , the corresponding combination of consonants and finals is "t-ong d-ou", and the durations a, b, c, and d of the four consonants are obtained, and the mean square error of the four consonants is calculated.

In step 4, the mean square error in step 3 is input into the binary classification model to obtain the probability of the first word, and the probability of the first word is used to represent the probability that the first word is the target keyword.

The training process of the binary classification network is as follows: the secondary classification network is trained by using machine learning algorithms such as Logistic Regression (LR) and Support Vector Machine (SVM). The parameters of the two-class network are obtained through the training of the first sample, and the above-mentioned first sample is the difference between the phased pronunciation duration feature and the reference phased pronunciation duration feature of the words in the speech database to be the target keywords in the keyword set. The difference between the values, marked as 0 and 1, corresponds to the two categories of non-keywords and keywords, respectively.

Step 5: Perform a second judgment according to the probability of the first word, and determine the first word as the target keyword when the probability of the first word is greater than the second predetermined value.

FIG. 6 is a schematic structural diagram of an apparatus for detecting speech keywords according to an embodiment of the present application. As shown in Figure 6, the device includes:

The obtaining unit 10 is used to obtain the detection result of the audio to be detected by the voice detection model, wherein the above-mentioned detection result is used to represent the phased pronunciation duration feature of the first word, and the above-mentioned first word is the above-mentioned audio to be detected by the above-mentioned voice detection model. The identified words that are intended to be the target keywords in the keyword set, and the above-mentioned keyword set saves a plurality of pre-selected keywords;

In this embodiment, the above-mentioned phased pronunciation duration feature may be the duration of initials and finals of each character in the first word. For example, the initials and finals corresponding to potatoes are "t-u d-ou", and the phased pronunciation duration features of potatoes are The pronunciation duration of t, the pronunciation duration of u, the pronunciation duration of d, and the pronunciation duration of ou. In practical applications, due to the different frequencies of people speaking, the pronunciation duration of the initials and finals of the same word may be quite different, but the difference in the ratio of the durations of the initials and finals is small. In order to further improve the accuracy of the detection results, the above phased The pronunciation duration feature may be the ratio of the duration of the initial consonant to the duration of the final, for example, the ratio of the pronunciation duration of t to the pronunciation duration of u. The above-mentioned staged pronunciation duration feature may also be the normalized initial consonant duration and final consonant duration, thereby reducing errors caused by different speaking frequencies.

The above-mentioned first word is the word that the above-mentioned speech detection model recognizes from the above-mentioned audio to be detected and is intended to be the target keyword in the keyword set, which means that the above-mentioned first word is recognized as the target keyword by the speech detection model, but due to the speech detection model. Misrecognition occurs, and the first word may not actually be the target keyword. For example, the target keyword is potatoes, and the first word is actually copper beans, but is recognized as potatoes by the speech detection model.

The above keyword set includes a plurality of keywords, such as open, close, mute and so on. The above target keyword may be one or more than one.

The determining unit 20 is configured to determine the first word as the target keyword when the phased pronunciation duration feature of the first word matches the phased pronunciation duration characteristic of the target keyword.

In this embodiment, the staged pronunciation duration feature of the target keyword may be the duration of the initial consonant and the duration of the final consonant of each character, or the ratio of the duration of the initial consonant to the duration of the final vowel, or the normalized duration of the initial consonant and the duration of the final consonant. Rhyme duration. The phased pronunciation duration feature of the above target keyword is obtained by calculating the average duration of each type of initials or the duration of each type of finals of a large number of marked speeches in the speech database. If the difference between the staged pronunciation duration feature of the first word and the staged pronunciation duration feature of the target keyword is within a certain range, it means that the difference between the above-mentioned first word and the above-mentioned target keyword is small, and the first word is the target key. word.

In order to obtain the detection result quickly, in an embodiment, the determining unit includes a first determining module and a second determining module, wherein the first determining module is used to determine the phased pronunciation duration feature of the first word and the target The first difference value of the phased pronunciation duration feature of the keyword; the above-mentioned second determination module is configured to determine the above-mentioned first word as the above-mentioned target keyword according to the above-mentioned first difference value.

In this embodiment, it is first necessary to determine the first difference between the staged pronunciation duration feature of the first word and the staged pronunciation duration feature of the target keyword, and it can be determined whether the first word is the target according to the first difference Key words, this detection method is relatively simple and does not require complicated calculation process, so that the detection result can be obtained quickly.

In an embodiment, the second determination module includes a first determination sub-module, an input sub-module and a second determination sub-module, wherein the first determination sub-module is configured to be used when the first difference value is smaller than a first predetermined value In the case of , determine that the above-mentioned first word is the above-mentioned target keyword; or, the above-mentioned input sub-module is used to input the above-mentioned first difference into the two-class network to obtain the probability of the above-mentioned first word, wherein the above-mentioned first word The probability of is used to characterize the probability that the first word is the target keyword. The above-mentioned two-class network is obtained by using the first sample to train the initial two-class network, and the above-mentioned first sample is the prediction in the speech database. Obtained after marking the difference between the phased pronunciation duration feature of the words of the target keyword in the keyword set and the reference phased pronunciation duration feature; the above-mentioned second determination sub-module is used when the probability of the above-mentioned first word is greater than that of the first word. In the case of two predetermined values, it is determined that the above-mentioned first word is the above-mentioned target keyword.

In this embodiment, the comparison can be made directly according to the size of the first difference. If the first difference is smaller than the first predetermined value, it means that the phased pronunciation duration feature of the first word is different from the phased pronunciation duration characteristic of the target keyword. If it is smaller, the first word is the target keyword. It is also possible to input the first difference into a two-class network, and use the two-class network to obtain the probability that the first word is the target keyword. When the probability is greater than the second predetermined value, the above-mentioned first word is the above-mentioned target keyword, thereby further. Improves the accuracy of speech keyword detection results.

The labels of the above two-category network are 0 and 1, which correspond to the two categories of non-keywords and time-keywords, respectively. The above-mentioned binary classification network generally adopts machine learning algorithms such as Logistic Regression (LR), Support Vector Machine (SVM) and the like to train the secondary classification network. The parameters of the Logistic binary classification network are obtained through the training of the first sample, and the above-mentioned first sample is the phased pronunciation duration feature and the reference phased pronunciation of the words in the speech database to be the target keywords in the keyword set The difference between duration features, marked as 0 and 1.

得到上述第一差值，其中，MSE表示上述第一差值，

表示上述目标关键词的声母时长均值或韵母时长均值，y_i表示上述第一词语的声母时长或韵母时长，n表示上述第一词语的声母数量和韵母数量之和，i为大于等于1且小于等于n的整数。In an embodiment, the above-mentioned first determination module includes a first calculation sub-module, wherein the above-mentioned first calculation sub-module is used for using

Obtain the above-mentioned first difference, wherein MSE represents the above-mentioned first difference,

Represents the average duration of initials or finals of the target keyword, y _i represents the duration of initials or finals of the first word, n represents the sum of the number of initials and the number of finals of the first word, i is greater than or equal to 1 and less than An integer equal to n.

Because the loss function usually adopts the mean square error, in this embodiment, the above formula is used to calculate the mean square error between the initial and final duration of the first word and the initial and final duration of the target keyword, which is conducive to the subsequent adoption of the model method to judge.

For example, if the target keyword is potatoes, the corresponding combination of initials and finals is "t-u d-ou", the average duration of these four initials and finals is x, y, z, m, and the speech detection model detects that the first word is actually "tongdou" , the corresponding combination of consonants and finals is "t-ong d-ou", and the durations a, b, c, and d of the four consonants are obtained, and the mean square error of the four consonants is calculated.

In an embodiment, the first determination module includes a second calculation submodule and a third calculation submodule, wherein the second calculation submodule is used to determine a first ratio between the duration of the initial consonant and the duration of the final vowel of the first word and the second ratio of the duration of the initial consonant and the duration of the final vowel of the target keyword; the third calculation sub-module is configured to obtain the first difference according to the difference between the first ratio and the second ratio.

As mentioned in the above content, the different frequencies of people's speech lead to large differences in the characteristics of the stage pronunciation duration. In order to avoid this difference and further improve the accuracy of the detection results, in this embodiment, the difference between the duration of the initial consonant and the duration of the final vowel is calculated. The ratio is then determined according to the difference between the first ratio of the first word and the second ratio of the target keyword.

The ratio of the duration of the initial consonant and the duration of the final vowel of the above-mentioned first word or target keyword can be the ratio of the duration of the initial consonant and the duration of the final vowel of each character in the first word or the target keyword, or the ratio of the duration of each character. The ratio of the duration of initials to the duration of finals. If it is the ratio of the duration of the initials to the duration of the finals of each character, there will be multiple ratios. Calculate the error of the ratio between the first word and each character in the target keyword, and then calculate the mean squared error to get the first difference. .

In one embodiment, the obtaining unit includes a decoding module, a generating module, a retrieval module and a third determining module, wherein the decoding module is used to decode the audio to be detected by using a speech recognition model to obtain a sequence of candidate words, Wherein, the above-mentioned speech recognition model is used to convert the above-mentioned audio to be detected into text data; the above-mentioned generation module is used to generate a word grid according to the above-mentioned candidate word sequence and the corresponding backtracking path and matching score; the above-mentioned retrieval module is used to use the above-mentioned The pinyin of a plurality of keywords in the keyword set is used to retrieve the pinyin of the words in the word grid to obtain the first word; the third determination module is used for obtaining the first word according to the pinyin of the first word Periodic pronunciation duration characteristics of words.

In this embodiment, a speech recognition model is used to first decode the audio to be detected to obtain a word grid, and then perform a quick search on the word grid to retrieve the first word and the corresponding phased pronunciation duration feature. The accuracy of the method is higher, but the running time is longer because of the need for decoding.

In one embodiment, the obtaining unit includes an extraction module, an input module, a fourth determination module, and a fifth determination module, wherein the extraction module is used to extract the speech features of each frame of the audio to be detected, and obtain the Speech feature vector; the above-mentioned input module is used to input the above-mentioned speech feature vector of each frame into the deep neural network model for classification, and obtain a plurality of labels and the probability of each label, wherein, the above-mentioned labels are identified from the above-mentioned audio to be detected. The above-mentioned fourth determination module is used to determine that the above-mentioned label is the above-mentioned first word when the probability of the label is greater than the third predetermined value; the above-mentioned The fifth determining module is configured to determine the phased pronunciation duration feature of the first word according to the starting speech frame and the ending speech frame of the first word.

In this embodiment, a neural network method is used for keyword retrieval, which does not require decoding of the audio to be detected, thereby reducing the computation time and memory usage during runtime. Figures 2 to 4 are for the application A schematic flowchart of a method for obtaining a detection result by using a deep neural network provided by the embodiment, as shown in FIG. 2 , first, feature extraction is performed on the audio to be detected, and a speech feature vector of the audio to be detected is extracted, and the audio to be detected includes the key phrase "okay, google" ", then input the speech feature vector into the deep neural network as shown in Figure 3, output 3 labels "okay", "google" and ", filler", and generate the frame-level posterior probability as shown in Figure 4 , the post-processing module combines these probabilities to provide the final confidence score for that window.

As shown in FIG. 7 , an embodiment of the present application provides an electronic device, including a processor 111 , a communication interface 112 , a memory 113 and a communication bus 114 , wherein the processor 111 , the communication interface 112 , and the memory 113 are completed through the communication bus 114 communication with each other,

a memory 113 for storing computer programs;

In an embodiment of the present application, the processor 111 is configured to implement the control method for detecting speech keywords provided by any one of the foregoing method embodiments when executing the program stored in the memory 113 .

Embodiments of the present application further provide a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements the steps of the voice keyword detection method provided by any one of the foregoing method embodiments.

It should be noted that, in this document, relational terms such as "first" and "second" etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply these There is no such actual relationship or sequence between entities or operations. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass a non-exclusive inclusion such that a process, method, article or device that includes a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

The above descriptions are only specific embodiments of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims (10)

1. A method for detecting a voice keyword is characterized by comprising the following steps:

acquiring a detection result of a voice detection model to-be-detected audio, wherein the detection result is used for representing the stage pronunciation duration characteristic of a first word, the first word is a word which is recognized by the voice detection model from the to-be-detected audio and is intended as a target keyword in a keyword set, and a plurality of preselected keywords are stored in the keyword set;

and under the condition that the stage pronunciation duration characteristics of the first word are matched with the stage pronunciation duration characteristics of the target keyword, determining that the first word is the target keyword.

2. The method of claim 1, wherein determining that the first word is the target keyword if the stage pronunciation time characteristics of the first word match the stage pronunciation time characteristics of the target keyword comprises:

determining a first difference value of the stage pronunciation duration characteristic of the first word and the stage pronunciation duration characteristic of the target keyword;

and determining the first word as the target keyword according to the first difference.

3. The method of claim 2, wherein determining the first term as the target keyword according to the first difference comprises:

determining that the first word is the target keyword under the condition that the first difference is smaller than a first preset value; or the like, or, alternatively,

inputting the first difference into a two-class network to obtain the probability of the first word, wherein the probability of the first word is used for representing the probability that the first word is the target keyword, the two-class network is obtained by training an initial two-class network by using a first sample, and the first sample is obtained by labeling the difference between the stage pronunciation time length characteristic and the reference stage pronunciation time length characteristic of the word which is supposed to be the target keyword in the keyword set in the voice database; determining that the first word is the target keyword when the probability of the first word is greater than a second predetermined value.

4. The method of claim 2, wherein determining a first difference between the staged pronunciation time feature of the first word and the staged pronunciation time feature of the target keyword comprises:

by using

Obtaining the first difference, wherein MSE represents the first difference,

an initial consonant duration mean or a final sound duration mean, y, representing the target keyword _i The time length of the initial consonant or the time length of the final of the first word is represented, n represents the sum of the initial consonant number and the final number of the first word, and i is an integer which is greater than or equal to 1 and less than or equal to n.

5. The method of claim 2, wherein determining a first difference between the staged pronunciation time feature of the first word and the staged pronunciation time feature of the target keyword comprises:

determining a first ratio of the initial duration and the final duration of the first word and a second ratio of the initial duration and the final duration of the target keyword;

and obtaining the first difference value according to the difference value of the first ratio value and the second ratio value.

6. The method of claim 1, wherein obtaining the detection result of the voice detection model on the audio to be detected comprises:

decoding the audio to be detected by adopting a voice recognition model to obtain a candidate word sequence, wherein the voice recognition model is used for converting the audio to be detected into text data;

generating a word grid according to the candidate word sequence, the corresponding backtracking path and the matching score;

retrieving word pinyin in the word grid by using pinyin of a plurality of keywords in the keyword set to obtain the first word;

and obtaining the stage pronunciation duration characteristics of the first word according to the pinyin of the first word.

7. The method of claim 1, wherein obtaining the detection result of the voice detection model on the audio to be detected comprises:

extracting the voice characteristics of each frame of the audio to be detected to obtain the voice characteristic vector of each frame;

inputting the voice feature vectors of each frame into a deep neural network model for classification to obtain a plurality of labels and the probability of each label, wherein the labels are words which are identified from the audio to be detected and are intended as target keywords in a keyword set or parts of the target keywords;

determining that the label is the first word under the condition that the probability of the label is greater than a third preset value;

and determining the stage pronunciation duration characteristics of the first word according to the starting speech frame and the ending speech frame of the first word.

8. A detection apparatus for a speech keyword, comprising:

the device comprises an acquisition unit, a detection unit and a processing unit, wherein the acquisition unit is used for acquiring a detection result of a voice detection model to-be-detected audio, the detection result is used for representing the stage pronunciation duration characteristic of a first word, the first word is a word which is recognized by the voice detection model from the to-be-detected audio and is intended as a target keyword in a keyword set, and a plurality of preselected keywords are stored in the keyword set;

a determining unit, configured to determine that the first word is the target keyword when the stage pronunciation duration feature of the first word matches the stage pronunciation duration feature of the target keyword.

9. An electronic device is characterized by comprising a processor, a communication interface, a memory and a communication bus, wherein the processor and the communication interface are used for realizing mutual communication by the memory through the communication bus;

a memory for storing a computer program;

a processor for implementing the steps of the method for detecting a speech keyword according to any one of claims 1 to 7 when executing a program stored in the memory.

10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method for detecting a speech keyword according to any one of claims 1 to 7.

Images