CN106531158A - Method and device for recognizing answer voice - Google Patents

Abstract

The invention relates to the field of computer paralinguistic information, in particular to a method and device for recognizing a response voice, which is used to solve the problem that the current response voice recognition method is not accurate enough in recognizing the response voice. The embodiment of the present invention obtains the response voice to be recognized; the response mode corresponding to the response voice to be recognized is determined using the response mode recognition model; if the response mode is a formal response mode, the response voice to be recognized is input into the first voice recognition system; if the response mode is In the informal response mode, the response voice to be recognized is input into the second voice recognition system. When the embodiment of the present invention recognizes the response voice, it first recognizes the response voice as a formal response mode or an informal response mode, and inputs different voice recognition systems for the formal response mode and the informal response mode for recognition, thereby improving the overall voice recognition performance .

Description

Recognition method and device for voice response

technical field

The invention relates to the field of computer paralanguage, in particular to a method and device for recognizing a response voice.

Background technique

In recent years, computer paralinguistics has become a research hotspot in the field of speech and language processing. The development of speech recognition technology plays an important role in promoting the development and application of intelligent and humanized new human-computer interaction technology.

Speech recognition is a technology that uses computers to automatically convert speech into text. Speech has always been an important medium for interaction in human life, so it is a crucial step for machines to recognize speech. Currently, voice recorders are used to record voices in many occasions, and the voices recorded in the voice recorders need to be analyzed. For example, in a flight scene, the cabin voice recorder is used to record the voice on the plane, and the flight quality is evaluated by recognizing the voice in the cabin voice recorder after the flight. At present, when the voice information recorded in the voice recorder is recognized, the method of automatic machine recognition is adopted. Specifically, the voice recorded in the voice recorder is divided into a sentence to be recognized and answered by using the endpoint recognition technology. The speech to be recognized is input into the speech recognition system, and the speech to be recognized is recognized by the speech recognition system. Since the speech to be recognized is divided into formal speech and informal speech according to different speaking objects and environments, the speech environments corresponding to formal speech and informal speech are different, and the tone and intonation of the speaker are not the same; while the existing The method of directly inputting the obtained answering voice into the voice recognition system for recognition often cannot accurately recognize the answering voice.

To sum up, the current response speech recognition methods are not accurate enough in recognizing the response speech.

Contents of the invention

The invention provides a method and device for recognizing a response voice, which is used to solve the problem that the current response voice recognition method is not accurate enough in recognizing the response voice.

Based on the above problems, an embodiment of the present invention provides a method for recognizing a response voice, including:

Obtain the response voice to be recognized;

Using the response mode recognition model to determine the response mode corresponding to the response voice to be recognized; wherein, the response mode recognition model is a supervised machine learning model;

If the response mode is a formal response mode, input the response voice to be recognized into the first voice recognition system, so that the first voice recognition system recognizes the response voice to be recognized and output the response voice to be recognized Corresponding text information;

If the response mode is an informal response mode, input the response voice to be recognized into the second voice recognition system, so that the second voice recognition system can recognize the response voice to be recognized and output the response to be recognized Text information corresponding to voice;

Wherein, the first speech recognition system and the second speech recognition system are configured with different parameters.

Since the embodiment of the present invention recognizes the response voice, after obtaining the response voice to be recognized, the response mode recognition model is used to determine the response mode corresponding to the response voice to be recognized, and input different voice recognition systems for the formal response mode and the informal response mode. identify. Since the first voice recognition system is used to recognize formal response voices, the second voice recognition system is used to recognize informal response voices, and the first voice recognition system and the second voice recognition system are configured with different parameters for different response methods Different voice recognition systems perform recognition, so that the recognition of the response voice to be recognized is more accurate.

Optionally, the use of the response mode recognition model to determine the response mode corresponding to the to-be-recognized response voice specifically includes:

Inputting the voice features extracted from the response voice to be recognized into the response mode recognition model;

Acquiring a response mode corresponding to the to-be-recognized response voice output by the response mode recognition model.

After the embodiment of the present invention extracts the features of the response voice to be recognized, the extracted voice features are input into the response mode recognition model, and the response mode corresponding to the response voice to be recognized is determined through the response mode recognition model.

Optionally, the speech features include frame-level features, slice-level features and segment-level features;

Speech features are extracted from the answering speech in the following way:

Using a feature extraction tool to extract the frame-level features of the response speech to be recognized according to the preset frame length and frame shift;

Perform smoothing and filtering processing on the frame-level features, and perform a differential operation on the smoothed frame-level features to determine the slice-level features of the speech to be recognized;

According to preset statistical parameters, the slice-level features are analyzed and processed to determine the segment-level features of the response speech to be recognized.

Since the embodiment of the present invention extracts frame-level, slice-level, and segment-level speech features from the response speech to be recognized, it is ensured that the response method recognition model accurately recognizes the response method corresponding to the response speech to be recognized.

Optionally, the response mode identification model is obtained according to the following manner:

Determine a training set comprising a plurality of response voices, and a test set comprising a plurality of response voices; wherein, the response voices in the training set are different from the response voices in the test set;

For any answering voice in the training set, the voice features extracted from the answering voice are input into the response mode recognition model before training for training;

For any response voice in the test set, input the voice features extracted from the response voice into the trained response mode recognition model, and obtain the response corresponding to the response voice output by the response mode recognition model Way;

According to the response mode corresponding to each response voice in the test set output by the trained response mode recognition model, determine the correct recognition rate of the trained response mode recognition model, if the correct recognition rate is greater than the set threshold, determine The training of the trained response mode recognition model is completed, and the trained response mode recognition model is saved.

Because the embodiment of the present invention uses a plurality of response voices in the training set to train the response mode recognition model, use the response voices in the test set to judge whether the trained response mode recognition model meets the requirements after training, and identify the test mode in the response mode recognition model. When the correct recognition rate of the concentrated response speech is greater than the set threshold, it is determined that the response recognition model training is completed, and the trained response recognition model is saved; if the correct recognition rate is less than the preset threshold, then the response speech in the training set is used again The training is performed until the correct recognition rate of the response mode recognition model is greater than the set threshold, so as to ensure that the obtained response mode recognition model can more accurately identify the response mode corresponding to the response speech to be recognized.

Optionally, the response mode identification model is a support vector machine (SVM) model.

On the other hand, the embodiment of the present invention also provides a recognition device for answering voice, including:

An acquisition module, configured to acquire the response voice to be recognized;

The identification module is used to determine the response mode corresponding to the response voice to be recognized by using the response mode recognition model; wherein, the response mode recognition model is a supervised machine learning model;

A judging module, configured to input the response voice to be recognized into the first voice recognition system if the response mode is a formal response mode, so that the first voice recognition system can recognize the response voice to be recognized and output the response voice to be recognized. The text information corresponding to the response voice to be recognized; if the response mode is an informal response mode, then input the response voice to be recognized into the second voice recognition system, so that the second voice recognition system recognizes the response voice to be recognized Response voice, and output text information corresponding to the response voice to be recognized; wherein, the first voice recognition system and the second voice recognition system are configured with different parameters.

Optionally, the identification module is specifically used for:

Inputting the speech features extracted from the response speech to be recognized into the response mode recognition model; obtaining the response mode corresponding to the response speech to be recognized outputted by the response mode recognition model.

Optionally, the speech features include frame-level features, slice-level features and segment-level features;

The identification module is specifically used for:

Speech features are extracted from the answering speech in the following way:

Use a feature extraction tool to extract the frame-level features of the response speech to be recognized according to the preset frame length and frame shift; perform smoothing and filtering on the frame-level features, and perform a differential operation on the smoothed frame-level features Determining slice-level features of the speech to be recognized; analyzing and processing the slice-level features according to preset statistical parameters to determine segment-level features of the speech to be recognized.

Optionally, the acquisition module is also used for:

Obtain the response mode recognition model according to the following manner:

Determine the training set that includes a plurality of answering voices, and the test set that includes a plurality of answering voices; wherein, the answering voices in the training set are different from the answering voices in the test set; for any one answering voice in the training set, the The voice feature extracted from the answering voice is input into the response mode recognition model before training for training; for any answering voice in the test set, the voice feature extracted from the answering voice is input into the post-training In the response mode recognition model, and obtain the response mode corresponding to the said response voice output by the response mode recognition model; according to the response mode corresponding to each response voice in the test set output by the response mode recognition model after training, determine If the recognition accuracy rate of the trained response mode recognition model is greater than a set threshold, it is determined that the training of the trained response mode recognition model is completed, and the trained response mode recognition model is saved.

Optionally, the response mode identification model is a support vector machine (SVM) model.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For Those of ordinary skill in the art can also obtain other drawings based on these drawings without any creative effort.

Fig. 1 is the flow chart of the recognition method of response speech of the embodiment of the present invention;

Fig. 2 is the flow chart of the embodiment of the present invention extracting speech feature;

FIG. 3 is a flow chart of a method for obtaining a response mode identification model according to an embodiment of the present invention;

FIG. 4 is an overall flowchart of a method for obtaining a response mode recognition model according to an embodiment of the present invention;

Fig. 5A is a schematic diagram of the accuracy rate of the recognition result corresponding to the SVM kernel function according to the embodiment of the present invention;

Fig. 5B is a performance comparison diagram of the SVM kernel function of the embodiment of the present invention;

FIG. 6 is a schematic structural diagram of a recognition device for answering speech according to an embodiment of the present invention.

detailed description

The embodiment of the present invention acquires the response voice to be recognized; uses the response mode recognition model to determine the response mode corresponding to the response voice to be recognized; wherein, the response mode recognition model is a supervised machine learning model; if the response mode is formal response mode, then input the response voice to be recognized into the first voice recognition system, so that the first voice recognition system can recognize the response voice to be recognized, and output the text information corresponding to the response voice to be recognized; if the If the response mode is an informal response mode, the response voice to be recognized is input into the second voice recognition system, so that the second voice recognition system recognizes the response voice to be recognized, and outputs the corresponding voice of the response voice to be recognized. text information; wherein, the first speech recognition system and the second speech recognition system are configured with different parameters.

Since the embodiment of the present invention recognizes the response voice, after obtaining the response voice to be recognized, the response mode recognition model is used to determine the response mode corresponding to the response voice to be recognized, and input different voice recognition systems for the formal response mode and the informal response mode. identify. Since the first speech recognition system is used to recognize formal response speech, the second speech recognition system is used to recognize informal reply speech, and the first speech recognition system and the second speech recognition system are configured with different parameters, the embodiment of the present invention first recognizes The answering voice is a formal answering method or an informal answering method, and different speech recognition systems are used to recognize different answering methods, so as to improve the overall speech recognition performance and make the recognition of the answering voice to be recognized more accurate.

It should be noted that the method for identifying the response mode of the response voice in the embodiment of the present invention can not only be used to improve the effect of the voice recognition system, but also can be applied to other high-level systems, such as speaker recognition systems, abnormal sound monitoring systems, etc. .

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments . Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

As shown in Figure 1, the recognition method of answering voice in the embodiment of the present invention comprises:

Step 101, obtaining the response voice to be recognized;

Step 102, using the response mode recognition model to determine the response mode corresponding to the response speech to be recognized; wherein, the response mode recognition model is a supervised machine learning model;

Step 103: If the answering method is a formal answering method, input the speech to be recognized into the first speech recognition system, so that the first speech recognition system can recognize the speech to be recognized and output the speech to be recognized. Recognize the text information corresponding to the response voice; if the response mode is an informal response mode, input the response voice to be recognized into the second voice recognition system, so that the second voice recognition system can recognize the response voice to be recognized , and output text information corresponding to the response voice to be recognized; wherein, the first voice recognition system and the second voice recognition system are configured with different parameters.

The response modes corresponding to the to-be-recognized response voice in the embodiment of the present invention include formal response modes and informal response modes;

The embodiment of the present invention can be applied in a flight scene to identify the response mode of the response voice in the flight scene, and identify whether the response mode of the response voice in flight is a formal response mode or an informal response mode. Among them, the recognized voice of the formal answering method is an indicative dialogue between the driver and the ground control center; Reply confirmation etc.

The recognized voice of the informal response mode is a dialogue between the pilot and the co-pilot, or a dialogue between the pilot and the ground tower; for example, the voice of the chat between the pilot and the co-pilot, the guidance between the pilot and the co-pilot during the flight Voice, the pilot reports the status of the aircraft to the ground tower center.

It should be noted that the embodiment of the present invention is not limited to the flight scene, and the response method recognition method of the embodiment of the present invention can be used in any language scene, and, in different language scenes, the formal response method and non- Definitions of formal responses also vary. For example, A and B are football game narrators. When determining the response mode of the dialogue information between A and B, the dialogue about the football match between A and B is defined as the dialogue of the formal response mode, and A and B The dialogue between B that has nothing to do with the football game is defined as the dialogue in the form of informal response.

In the embodiment of the present invention, when using the response mode recognition model to determine the response mode corresponding to the response voice to be recognized, the following methods are specifically adopted:

Optionally, input the speech features extracted from the response speech to be recognized into the response mode recognition model; and acquire the response mode corresponding to the response speech to be recognized output by the response mode recognition model.

Wherein, the response mode recognition model in the embodiment of the present invention is a supervised machine learning model, specifically, the response mode recognition model in the embodiment of the present invention is an SVM (Support Vector Machine) model.

In the embodiment of the present invention, after the speech to be recognized is acquired, a feature extraction tool is used to extract speech features in the speech to be recognized.

During implementation, the embodiment of the present invention extracts the speech features in the response speech to be recognized in a layered extraction manner when extracting the speech features in the response speech to be recognized.

The speech features in the embodiment of the present invention include frame level features, segment level features and segment level features.

Specifically, the embodiment of the present invention uses the openSMILE feature extraction tool to perform hierarchical extraction of the response speech to be recognized, and extract the speech features in the response speech to be recognized.

Optionally, use a feature extraction tool to extract the frame-level features of the response speech to be recognized according to the preset frame length and frame shift; perform smoothing filtering on the frame-level features, and smooth the frame-level features after smoothing Perform differential operation on the features to determine slice-level features of the speech to be recognized; analyze and process the slice-level features according to preset statistical parameters to determine segment-level features of the speech to be recognized. .

The method for extracting speech features from the response speech to be recognized will be introduced in detail below in the embodiment of the present invention.

The first step is to extract the frame-level features in the speech to be recognized.

Among them, the frame-level features are the first-level speech features in the response speech to be recognized.

In the implementation, using the openSMILE feature extraction tool, the frame length is 20ms, and the frame shift is 10ms, which contains a total of 16-dimensional features. The specific frame-level feature parameters are shown in Table 1, including:

RMSenergy (Root Mean Square energy, energy root mean square), mfcc (Mel-FrequencyCepstral Coefficient, Mel frequency cepstral coefficient) 1-12 dimensions, zcr (zero-crossing rate, zero-crossing rate), Voice_prob (voiced sound ratio) , F0 (fundamental frequency calculated according to the cepstrum).

Table 1

English abbreviation for frame-level features Chinese interpretation of frame-level features RM Energy energy root mean square mfcc(1-12) Mel frequency cepstral coefficient 1-12 dimensions zcr Zero-crossing rate (frame level) Voice_prob Calculation of voiced speech proportion by autocorrelation F0 Fundamental frequency calculated from the cepstrum

The second step is to extract slice-level features in the speech to be recognized.

Among them, the slice-level feature is the second-level speech feature in the response speech to be recognized.

Specifically, smoothing and filtering is performed on the frame-level features, and a difference operation is performed on the smoothed frame-level features to determine slice-level features in the response speech to be recognized.

During implementation, the frame sequence obtained in the first step is carried out to a smoothing filter sma (smoothed by a moving average filter) with a window length of 3 frames;

After smoothing and filtering the frame sequence, the first-order difference de(deltacoefficient) is made on the smoothed features.

Among them, the slice-level feature analysis functions to be specified are shown in Table 2, specifically including:

sma (smooth filtering) and de (first difference).

Table 2

English abbreviation for chip-level feature analysis function Chinese explanation of chip-level feature analysis function sma smoothing filter de first difference

After the first step and the second step, a total of 16*2=32 dimensional speech features are obtained.

The third step is to extract the segment-level features in the speech to be recognized.

Among them, the segment-level features are the third-level speech features in the response speech to be recognized.

Specifically, the slice-level features are analyzed and processed according to preset statistical parameters to determine segment-level features in the response speech to be recognized.

During implementation, statistical analysis is performed on the features output in the second step, which mainly includes 12 statistical parameters. According to the 12 statistical parameters, the feature slice-level features output in the second step are analyzed and processed to obtain the segment-level features in the speech to be recognized. .

The specific preset 12 statistical parameters are shown in Table 3, including:

max(maximum, the maximum value of the envelope), min(minute, the minimum value of the envelope), range (the range of the envelope change), maxpos(maximum position, the maximum position), minpos(minute position, the absolute minimum value of the envelope Position), amean (Arithmetic mean, envelope arithmetic mean), linregc1 (linear approximation slope of the envelope), linregc2 (linear approximation offset of the envelope), linregerrQ (the root mean square of the linear prediction value and the actual value of the envelope ), stddev (standard deviation), skewness (third-order skewness), kurtosis (fourth-order kurtosis).

table 3

English abbreviation for segment-level feature statistics parameters Chinese interpretation of segment-level characteristic statistical parameters max Envelope takes the maximum value min Envelope takes minimum value range Envelope variation range max pos Maximum position minpos Envelope Minimum Absolute Position amean Envelope Arithmetic Mean linregc1 linear approximation slope of the envelope linregc2 Linear approximation offset of the envelope linregerrQ Envelope's linear predicted value and the root mean square of the actual value stddev standard deviation skewness third order skewness kurtosis fourth order kurtosis

As shown in Figure 2, when the embodiment of the present invention extracts segment-level features in the speech to be recognized in the third step, statistical analysis is performed on the slice-level features obtained in the second step, and includes 12 preset statistics parameters, after the third step of segment-level feature extraction, a total of 16*2*12=384-dimensional speech features are obtained.

In the embodiment of the present invention, after the speech features in the response speech to be recognized are extracted by a feature extraction tool, the extracted speech features are input into the response mode recognition model, so that the response mode recognition model can recognize the speech features according to the speech features. The response mode corresponding to the response voice to be recognized; and the response mode corresponding to the response voice to be recognized outputted by the response mode recognition model according to the input voice features.

It should be noted that the response mode identification model in the embodiment of the present invention is a pre-trained model for identifying response modes.

Since the embodiment of the present invention recognizes the response mode corresponding to the response voice to be recognized, mainly by means of the response mode recognition model, and the response mode recognition model is a pre-trained model, therefore, the embodiment of the present invention also includes an important component , that is, to train the response pattern recognition model.

The process of training the response mode recognition model in the embodiment of the present invention will be described in detail below.

As shown in Figure 3, the method for obtaining the response mode recognition model in the embodiment of the present invention includes:

Step 301, determining a training set comprising a plurality of response voices, and a test set comprising a plurality of response voices; wherein, the response voices in the training set are different from the response voices in the test set;

Step 302, for any one of the response voices in the training set, input the voice features extracted from the response voice into the response mode recognition model before training for training;

Step 303, for any one of the response voices in the test set, input the voice features extracted from the response voice into the trained response pattern recognition model, and obtain the response speech output by the response pattern recognition model The corresponding response method;

Step 304, according to the response mode corresponding to each response voice in the test set output by the trained response mode recognition model, determine the recognition accuracy rate of the trained response mode recognition model, if the recognition accuracy rate is greater than the set The threshold value is used to determine that the training of the trained response mode recognition model is completed, and to save the trained response mode recognition model.

In step 301, the embodiment of the present invention selects a plurality of response voices from the corpus when determining the training set and the test set, and forms the training set or the test set from the selected multiple response voices.

The corpus in the embodiment of the present invention is a pre-recorded voice, and the pre-recorded voice includes multiple response voices in formal and informal response modes.

For example, the corpus can be 17.5 hours of speech recorded during the actual flight. After the recording, the 17.5 hours of speech are marked, assuming that the marking determines that the 17.5 hours of speech includes 18 speakers, including There are 4668 response voices in the formal response mode and 2257 response voices in the informal response mode, then the ratio of the response voices in the formal response mode to the response voices in the informal response mode is 2.07:1, and the voice sampling frequency of all the response voices Both are 16KHz, and the quantization precision is 16bit.

Select a plurality of response voices from all the response voices in the corpus to form a training set; preferably, the ratio of the response voices in the formal response mode to the response voices in the informal response mode in the training set is close to the response in the formal response mode in the corpus Ratio of speech to speech in response to informal responses.

For example, determine two training sets, which are respectively training set A and training set B, and determine a test set C, wherein, in the training set A, B and test set C, the response voice of the formal answering mode and the answering voice of the informal answering mode The number and proportion of voices are shown in Table 4:

Select 1580 formal response voices and 1580 informal response voices from the corpus to form training set A. The ratio of formal response voices to informal response voices in training set A is 1 : 1; from the corpus, select 3270 response voices of formal response mode and 1580 response voices of informal response mode to form training set B, the difference between the response voice of formal response mode and the response voice of informal response mode in training set B The ratio is 2.07:1; 1400 formal response voices and 677 informal response voices are selected from the corpus to form test set C. The ratio of voice to answer is 2.07:1.

Table 4

The following takes the training sets A, B and test set C shown in Table 4 as examples to illustrate the method of training the response mode recognition model.

Specifically, in the embodiment of the present invention, the response mode recognition model is trained through each response voice in the training set A and the training set B, and after the training is completed, each response voice in the test set C is input into the trained response For the mode recognition model, if the correct recognition rate of the response mode corresponding to the response voice in the test set C output by the response mode recognition model is greater than the set threshold, it is determined that the training of the response mode recognition model is completed, and the trained response mode recognition model is saved.

The following describes the process of training the response mode recognition model for any response voice in the training set A:

1. Use a feature extraction tool to extract the speech features of the answering speech.

The specific method for extracting the speech features of the response speech adopts the above-mentioned method, which will not be described in detail here.

2. Input the speech features corresponding to the response speech into the response mode recognition model for training.

Specifically, the answering voice corresponding to the answering voice is input into the response mode recognition model, and the answering mode corresponding to the response voice is input into the response mode recognition model, so that the response mode recognition model learns the response mode corresponding to the voice feature.

The embodiment of the present invention adopts the above-mentioned method, and uses the response voice in the training set to train the response mode recognition model. Respond to the voice, and judge whether the training of the recognition model of the response method is completed.

Specifically, when the test set C is used to judge whether the response mode recognition model has been trained, for any answer voice in the test set C, the following operations are performed:

1. Use the feature extraction tool to extract the speech features of the response speech;

The specific method for extracting the speech features of the response speech adopts the above-mentioned method, which will not be described in detail here.

2. Input the speech features corresponding to the response speech into the trained response mode recognition model;

3. Acquiring the response mode corresponding to the response voice output by the trained response mode recognition model.

Specifically, when the pre-set response mode recognition model determines that the response mode corresponding to the response voice is a formal response mode, the response mode recognition model outputs "1"; when determining that the response mode corresponding to the response voice is an informal response mode, the response mode The recognition model outputs "0".

In the embodiment of the present invention, after using the trained response mode recognition model to judge each response voice in the test set C, determine the corresponding recognition result of each response voice in the test set C; the test set determined by the response mode recognition model The recognition result corresponding to each response voice in C is compared with the response method corresponding to each response voice to determine the correct recognition rate of the recognition result corresponding to the test set C. If the correct recognition rate is greater than the set threshold, the response is determined. After the mode recognition model training is completed, save the trained response mode recognition model; if the correct recognition rate is not greater than the set threshold, then re-select the training set and test set, and continue training the response mode recognition model until the response mode recognition model The correct recognition rate corresponding to the recognition result of the response speech in the test set is greater than the set threshold.

As shown in FIG. 4 , an overall flow chart of a method for obtaining a response mode recognition model according to an embodiment of the present invention.

Step 401, determining a training set comprising a plurality of response voices, and a test set comprising a plurality of response voices; wherein, the response voices in the training set are different from the response voices in the test set;

The following steps 402 and 403 are for any answering speech in the training set.

Step 402, using a feature extraction tool to extract speech features in the response speech;

Step 403, input the extracted speech features and the corresponding response mode of the response voice into the response mode recognition model for training;

The following steps 404 and 405 are for any answering speech in the training set.

Step 404, using a feature extraction tool to extract speech features in the response speech;

Step 405, input the extracted speech features into the response mode recognition model for recognition;

Step 406, determining the recognition result of each response voice in the test set;

Step 407: Comparing the recognition result of each response voice in the test set with the corresponding answer mode of each response voice in the test set, and determining the correct recognition rate of the recognition result corresponding to the test set;

Step 408, judging whether the correct recognition rate is greater than the set threshold, if so, execute step 409, if not, return to step 401;

Step 409, after determining that the training of the response mode recognition model is completed, save the trained response mode recognition model.

In the embodiment of the present invention, in the binary classification problem of identifying the response mode, the support vector machine SVM classifier suitable for small data volumes is used as the response mode identification model, and the following kernel functions are compared: linear kernel function, polynomial kernel function, Gaussian Radial basis kernel function and arctangent kernel function.

The embodiment of the present invention is based on the training set shown in Table 4, respectively using linear kernel function, polynomial kernel function, Gaussian radial basis kernel function and arctangent kernel function to conduct experiments, and the accuracy of the obtained recognition results is shown in Figure 5A , wherein, when the SVM kernel function is a linear kernel function, the accuracy rate of the recognition result corresponding to the training set A is 80.30, and the accuracy rate of the recognition result corresponding to the training set B is 81.02; the SVM kernel function is a polynomial kernel function, and d=2 , the accuracy rate of the recognition result corresponding to the training set A is 77.95, and the accuracy rate of the recognition result corresponding to the training set B is 79.25; the SVM kernel function is a polynomial kernel function, and when d=3, the recognition result corresponding to the training set A is The accuracy rate is 76.17, the accuracy rate of the recognition result corresponding to the training set B is 81.13; the SVM kernel function is a polynomial kernel function, and when d=4, the accuracy rate of the recognition result corresponding to the training set A is 63.79, and the accuracy rate of the recognition result corresponding to the training set B is 63.79. The accuracy rate of the recognition result is 63.94; when the SVM kernel function is a Gaussian radial basis kernel function, the accuracy rate of the recognition result corresponding to the training set A is 90.71, and the accuracy rate of the recognition result corresponding to the training set B is 91.62; the SVM kernel function is When the arctangent kernel function is used, the accuracy rate of the recognition result corresponding to the training set A is 84.45, and the accuracy rate of the recognition result corresponding to the training set B is 89.56;

Moreover, the performance comparison of the SVM model using linear kernel function, polynomial kernel function, Gaussian radial basis kernel function and arctangent kernel function is shown in Fig. 5B.

Based on the same inventive concept, an identification device of a response mode is also provided in the embodiment of the present invention. Since the problem-solving principle of the device is similar to the identification method of the response mode in the embodiment of the present invention, the implementation of the device can refer to the method implementation, the repetition will not be repeated.

As shown in Figure 6, the recognition device of the response voice in the embodiment of the present invention includes:

Obtaining module 601, an obtaining module, used to obtain the response voice to be recognized;

The identification module 602 is configured to use a response mode recognition model to determine the response mode corresponding to the response speech to be recognized; wherein, the response mode recognition model is a supervised machine learning model;

Judging module 603, configured to input the response voice to be recognized into the first voice recognition system if the response mode is a formal response mode, so that the first voice recognition system can recognize the response voice to be recognized and output The text information corresponding to the response voice to be recognized; if the response mode is an informal response mode, input the response voice to be recognized into the second voice recognition system, so that the second voice recognition system can recognize the response voice to be recognized Recognizing a response voice, and outputting text information corresponding to the response voice to be recognized; wherein, the first voice recognition system and the second voice recognition system are configured with different parameters.

Optionally, the identification module 602 is specifically configured to:

Inputting the speech features extracted from the response speech to be recognized into the response mode recognition model; obtaining the response mode corresponding to the response speech to be recognized outputted by the response mode recognition model.

Optionally, the speech features include frame-level features, slice-level features and segment-level features;

The identification module 602 is specifically used for:

Speech features are extracted from the answering speech in the following way:

Use a feature extraction tool to extract the frame-level features of the response speech to be recognized according to the preset frame length and frame shift; perform smoothing and filtering on the frame-level features, and perform a differential operation on the smoothed frame-level features Determining slice-level features of the speech to be recognized; analyzing and processing the slice-level features according to preset statistical parameters to determine segment-level features of the speech to be recognized.

Optionally, the obtaining module 601 is also used for:

Obtain the response mode recognition model according to the following manner:

Determine the training set that includes a plurality of answering voices, and the test set that includes a plurality of answering voices; wherein, the answering voices in the training set are different from the answering voices in the test set; for any one answering voice in the training set, the The voice feature extracted from the answering voice is input into the response mode recognition model before training for training; for any answering voice in the test set, the voice feature extracted from the answering voice is input into the post-training In the response mode recognition model, and obtain the response mode corresponding to the said response voice output by the response mode recognition model; according to the response mode corresponding to each response voice in the test set output by the response mode recognition model after training, determine If the recognition accuracy rate of the trained response mode recognition model is greater than a set threshold, it is determined that the training of the trained response mode recognition model is completed, and the trained response mode recognition model is saved.

Optionally, the response mode identification model is a support vector machine (SVM) model.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special purpose computer, an embedded processor, or other programmable data processing equipment, so that the instructions executed by the processor of the computer or other programmable data processing equipment can realize the A process or processes and/or a function specified in a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in the flow or flows of the flowcharts and/or the block or blocks of the block diagrams.

While preferred embodiments of the present invention have been described, additional changes and modifications can be made to these embodiments by those skilled in the art once the basic inventive concept is appreciated. Therefore, it is intended that the appended claims be construed to cover the preferred embodiment as well as all changes and modifications which fall within the scope of the invention.

Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these modifications and variations.

Claims (10)

1. a kind of recognition methodss of response voice, it is characterised in that the method includes：

Obtain response voice to be identified；

The corresponding response mode of the response voice to be identified is determined using response mode identification model；Wherein, the answer party Formula identification model is the machine learning model for having supervision；

If the response mode is formal response mode, by first speech recognition system of response phonetic entry to be identified, So that first speech recognition system recognizes the response voice to be identified, and it is corresponding to export the response voice to be identified Text message；

If the response mode is unofficial response mode, by the second speech recognition of response phonetic entry system to be identified System, so that second speech recognition system recognizes the response voice to be identified, and exports the response voice pair to be identified The text message answered；

Wherein, first speech recognition system and second speech recognition system are configured with different parameters.

2. the method for claim 1, it is characterised in that the use response mode identification model determines described to be identified The corresponding response mode of response voice, specifically includes：

The phonetic feature extracted from the response voice to be identified is input into into the response mode identification model；

Obtain the corresponding response mode of response voice described to be identified of the response mode identification model output.

3. method as claimed in claim 2, it is characterised in that the phonetic feature includes frame level feature, chip level feature and section Level feature；

Phonetic feature is extracted from response voice according to following manner：

Using feature extraction tools, moved according to default frame length and frame, extract the frame level feature of the response voice to be identified；

The frame level feature is done into the disposal of gentle filter, and calculus of differences is done to the frame level feature after smoothing processing, it is determined that described The chip level feature of response voice to be identified；

According to default statistical parameter, process is analyzed to the chip level feature, determines the section of the response voice to be identified Level feature.

4. the method for claim 1, it is characterised in that the response mode identification model is obtained according to following manner：

It is determined that the training set comprising multiple response voices, and the test set comprising multiple response voices；Wherein, the training set In response voice it is different from the response voice in the test set；

For any one response voice in the training set, the phonetic feature extracted from the response voice is input to It is trained in response mode identification model before training；

For any one response voice in the test set, the phonetic feature extracted from the response voice is input to In response mode identification model after training, and obtain the response language of the output of the response mode identification model after the training The corresponding response mode of sound；

In the test set according to the response mode identification model output after the training, each response voice is corresponding should Mode is answered, the recognition correct rate of the response mode identification model after the training is determined, if the recognition correct rate is more than setting Threshold value, determines that the training of the response mode identification model after the training is completed, and preserves the response mode identification mould after the training Type.

5. the method as described in Claims 1 to 4 is arbitrary, it is characterised in that the response mode identification model is supporting vector Machine SVM models.

6. a kind of identifying device of response voice, it is characterised in that include：

Acquisition module, for obtaining response voice to be identified；

Identification module, for determining the corresponding response mode of the response voice to be identified using response mode identification model；Its In, the response mode identification model is the machine learning model for having supervision；

Judge module, if being formal response mode for the response mode, by the response phonetic entry to be identified first Speech recognition system, so that first speech recognition system recognizes the response voice to be identified, and exports described to be identified The corresponding text message of response voice；If the response mode is unofficial response mode, by the response voice to be identified The second speech recognition system is input into, so that second speech recognition system recognizes the response voice to be identified, and institute is exported State the corresponding text message of response voice to be identified；Wherein, first speech recognition system and the second speech recognition system It is under unified central planning to be equipped with different parameters.

7. device as claimed in claim 6, it is characterised in that the identification module, specifically for：

The phonetic feature extracted from the response voice to be identified is input into into the response mode identification model；Obtain described answering Answer the corresponding response mode of response voice described to be identified of mode identification model output.

8. device as claimed in claim 7, it is characterised in that the phonetic feature includes frame level feature, chip level feature and section Level feature；

The identification module, specifically for：

Phonetic feature is extracted from response voice according to following manner：

Using feature extraction tools, moved according to default frame length and frame, extract the frame level feature of the response voice to be identified；Will The frame level feature does the disposal of gentle filter, and does calculus of differences to the frame level feature after smoothing processing, determines described to be identified The chip level feature of response voice；According to default statistical parameter, process is analyzed to the chip level feature, it is determined that described wait to know The Utterance level feature of voice is not replied.

9. device as claimed in claim 6, it is characterised in that the acquisition module, is additionally operable to：

The response mode identification model is obtained according to following manner：

It is determined that the training set comprising multiple response voices, and the test set comprising multiple response voices；Wherein, the training set In response voice it is different from the response voice in the test set；For any one response voice in the training set, will The phonetic feature extracted from the response voice is input in the response mode identification model before training and is trained；For Any one response voice in the test set, the phonetic feature extracted from the response voice is input to after training In response mode identification model, and it is corresponding to obtain the response voice of the output of the response mode identification model after the training Response mode；In the test set according to the response mode identification model output after training, each response voice is corresponding should Mode is answered, the recognition correct rate of the response mode identification model after the training is determined, if the recognition correct rate is more than setting Threshold value, determines that the training of the response mode identification model after the training is completed, and preserves the response mode identification mould after the training Type.

10. the device as described in claim 6～9 is arbitrary, it is characterised in that the response mode identification model is supporting vector Machine SVM models.