CN109545193A

CN109545193A - Method and apparatus for generating model

Info

Publication number: CN109545193A
Application number: CN201811550086.XA
Authority: CN
Inventors: 李超
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2018-12-18
Filing date: 2018-12-18
Publication date: 2019-03-29
Anticipated expiration: 2038-12-18
Also published as: CN109545193B

Abstract

The embodiment of the present application discloses the method and apparatus for generating model, and the method and apparatus for generating information.One specific embodiment of the method for being used to generate model includes: to obtain the training sample set for being directed to target audio set, wherein, target audio set includes executing the audio that truncation obtains to initial audio, training sample and target audio correspond, training sample in training sample set includes the characteristic and identification information of the target audio in target audio set, whether it includes speech audio that identification information is used to indicate in the audio frame that target audio includes, and initial audio includes speech audio；Using machine learning algorithm, the characteristic for including using the training sample in training sample set is as input, and using identification information corresponding with the characteristic of input as desired output, training obtains speech recognition modeling.The embodiment enriches the training method of model, helps to improve the accuracy of speech terminals detection.

Description

Method and apparatus for generating model

Technical field

The invention relates to field of computer technology, and in particular to the method and apparatus for generating model.

Background technique

It is important that in audio, can judge position of the beginning and end of voice in audio in interactive voice It sets.In the prior art, voice activity detection (Voice Activity Detection, VAD) Lai Jinhang voice is generallyd use End-point detection.Voice activity detection, also known as speech terminals detection, speech endpoint detection, refer to and detect voice in noise circumstance Presence or absence.In general, voice activity detection can be used in the speech processing systems such as voice coding, speech enhan-cement, drop is played Low speech encoding rate saves communication bandwidth, reduces energy consumption of mobile equipment, improves the effects of discrimination.

The training sample of existing voice activity detection model is normally based on what the higher audio of quality obtained.

Summary of the invention

The embodiment of the present application proposes the method and apparatus for generating model, and method and dress for generating information It sets.

In a first aspect, the embodiment of the present application provides a kind of method for generating model, it is directed to this method comprises: obtaining The training sample set of target audio set, wherein target audio set includes executing truncation to initial audio to obtain Audio, the target audio in training sample and target audio set in training sample set correspond, training sample set In training sample include the target audio in target audio set characteristic and identification information, identification information is used to indicate It whether include speech audio in the audio frame that target audio includes, initial audio includes speech audio；Using machine learning algorithm, The characteristic for including using the training sample in training sample set is as input, by mark corresponding with the characteristic of input Information obtains speech recognition modeling as desired output, training.

In some embodiments, target audio set includes handling obtained target audio as follows to initial audio execution: Truncation is carried out to initial audio, obtains consonant frequency sequence；Delete at least one sub-audio in consonant frequency sequence；It will delete The combination of the sub-audio retained afterwards, is determined as target audio.

In some embodiments, at least one sub-audio in consonant frequency sequence is deleted, comprising: delete in consonant frequency sequence Preceding first quantity sub-audio, wherein the first quantity is less than the quantity of the consonant frequency sequence sub-audio that includes.

In some embodiments, at least one sub-audio in consonant frequency sequence is deleted, comprising: delete in consonant frequency sequence Rear second quantity sub-audio, wherein the second quantity is less than the quantity of the consonant frequency sequence sub-audio that includes.

In some embodiments, target audio set is by executing such as the initial audio in initial audio set What lower step obtained: random to generate the first random number and the second random number, wherein the first random number and the second random number are 0 Number between to 1；In response to determining that the first random number is less than predetermined first predetermined value, include to the initial audio Preceding third quantity sub-audio executes truncation and delete processing, and the combination of at least one sub-audio retained after deletion determines For target audio, wherein third quantity is less than the half of the quantity for the sub-audio that initial audio includes, and the first predetermined value is used for Characterize the audio that in predetermined audio set, first frame audio frame includes speech audio quantity and predetermined sound The ratio of the quantity of frequency set sound intermediate frequency；In response to determining that the second random number is less than predetermined second predetermined value, to this The rear 4th quantity sub-audio that initial audio includes executes truncation and delete processing, at least one consonant that will retain after deletion The combination of frequency, is determined as target audio, wherein and the 4th quantity is less than the half of the quantity for the sub-audio that initial audio includes, the Two predetermined values are for characterizing the number that in predetermined audio set, last frame audio frame includes the audio of speech audio The ratio of amount and the quantity of predetermined audio set sound intermediate frequency.

In some embodiments, using machine learning algorithm, the feature for including by the training sample in training sample set Data are as input, and using identification information corresponding with the characteristic of input as desired output, training obtains speech recognition mould Type, comprising: choose training sample from training sample set, and execute following training step: by the training sample packet of selection The characteristic included is input to initial speech identification model, obtains reality output, wherein reality output is initial speech identification mould The output of type；Based on reality output, determine whether initial speech identification model meets predetermined termination condition；In response to true Surely meet termination condition, the initial speech identification model of termination condition will be met, be determined as the speech recognition modeling that training obtains.

In some embodiments, this method further include: be unsatisfactory for termination condition in response to determination, be based on obtained reality Output and desired output corresponding with obtained reality output adjust the parameter of the model parameter of initial speech identification model It is worth, and chooses the training sample of unselected mistake from training sample set, is known based on parameter value initial speech adjusted Other model, continues to execute training step.

In some embodiments, the activation primitive for the output layer that initial speech identification model includes is normalization index letter Number, the cost function for the output layer that initial speech identification model includes are cross entropy cost function.

In some embodiments, speech recognition modeling is the Recognition with Recurrent Neural Network model with gating cycle unit.

Second aspect, the embodiment of the present application provide a kind of for generating the device of model, which includes: the first acquisition Unit is configured to obtain the training sample set for being directed to target audio set, wherein target audio set includes to initial sound Frequency executes the audio that truncation obtains, the target audio one in training sample and target audio set in training sample set One is corresponding, and the training sample in training sample set includes the characteristic and mark letter of the target audio in target audio set Breath, whether it includes speech audio that identification information is used to indicate in the audio frame that target audio includes, and initial audio includes voice sound Frequently；Training unit is configured to using machine learning algorithm, the characteristic for including by the training sample in training sample set As input, using identification information corresponding with the characteristic of input as desired output, training obtains speech recognition modeling.

In some embodiments, training unit includes: training module, is configured to choose training from training sample set Sample, and execute following training step: the characteristic that the training sample of selection includes is input to initial speech identification mould Type obtains reality output, wherein reality output is the output of initial speech identification model；Based on reality output, initial language is determined Whether sound identification model meets predetermined termination condition；Meet termination condition in response to determination, termination condition will be met Initial speech identification model is determined as the speech recognition modeling that training obtains.

In some embodiments, device further include: adjustment unit is configured in response to determination and is unsatisfactory for terminating item Part, is based on obtained reality output and desired output corresponding with obtained reality output, and adjustment initial speech identifies mould The parameter value of the model parameter of type, and the training sample of unselected mistake is chosen from training sample set, it is based on parameter value Initial speech identification model adjusted, continues to execute training step.

The third aspect, the embodiment of the present application provide a kind of method for generating information, this method comprises: obtaining target Audio, wherein target audio includes speech audio；For the audio frame that target audio includes, which is input in advance Trained speech recognition modeling obtains the probability that the audio frame includes speech audio, wherein speech recognition modeling is according to as above The method training for stating any embodiment in the method for generating information obtains；Based on obtained probability and predetermined The size relation of probability threshold value generates the speech terminals detection result of target audio.

In some optional implementations of the present embodiment, predetermined threshold value includes predetermined first threshold With predetermined second threshold, first threshold is greater than second threshold, and speech terminals detection result includes what target audio included The location information of starting point of the speech audio in target audio and the location information of terminal；And based on obtained probability and in advance The size relation of first determining threshold value, generates the speech terminals detection result of target audio, comprising: based on obtained probability and The size relation of predetermined first threshold, the position of starting point of the speech audio that generation target audio includes in target audio Confidence breath；Size relation based on obtained probability and predetermined second threshold generates the voice that target audio includes The location information of terminal of the audio in target audio.

Fourth aspect, the embodiment of the present application provide a kind of for generating the device of model, which includes: the second acquisition Unit is configured to obtain target audio, wherein target audio includes speech audio；Input unit is configured to for target The audio frame is input to speech recognition modeling trained in advance by the audio frame that audio includes, and obtaining the audio frame includes voice The probability of audio, wherein speech recognition modeling is the method according to any embodiment in such as above-mentioned method for being used to generate information What training obtained；Generation unit is configured to the size relation based on obtained probability and predetermined probability threshold value, raw At the speech terminals detection result of target audio.

In some optional implementations of the present embodiment, predetermined threshold value includes predetermined first threshold With predetermined second threshold, first threshold is greater than second threshold, and speech terminals detection result includes what target audio included The location information of starting point of the speech audio in target audio and the location information of terminal；And generation unit includes: first raw At module, it is configured to the size relation based on obtained probability and predetermined first threshold, generates target audio packet The location information of starting point of the speech audio included in target audio；Second generation module is configured to based on obtained general The size relation of rate and predetermined second threshold, terminal of the speech audio that generation target audio includes in target audio Location information.

5th aspect, the embodiment of the present application provide a kind of electronic equipment, comprising: one or more processors；Storage dress It sets, is stored thereon with one or more programs, when said one or multiple programs are executed by said one or multiple processors, make It obtains the one or more processors and realizes the method such as any embodiment in the above-mentioned method for being used to generate model, alternatively, making The method that the one or more processors realize any embodiment in the method as above-mentioned for generating information.

6th aspect, the embodiment of the present application provide a kind of computer-readable medium, are stored thereon with computer program, should The method that any embodiment in the method as above-mentioned for generating model is realized when program is executed by processor, alternatively, the program The method of any embodiment in the method as above-mentioned for generating information is realized when being executed by processor.

Method and apparatus provided by the embodiments of the present application for generating model, by obtaining for target audio set Training sample set, wherein target audio set includes executing the audio that truncation obtains, training sample set to initial audio The target audio in training sample and target audio set in conjunction corresponds, and the training sample in training sample set includes The characteristic and identification information of target audio in target audio set, identification information are used to indicate in target audio whether wrap Include speech audio, then, using machine learning algorithm, the characteristic for including using the training sample in training sample set as Input, using identification information corresponding with the characteristic of input as desired output, training obtains speech recognition modeling, thus rich The rich training method of model, helps to improve the accuracy of speech terminals detection.

Detailed description of the invention

By reading a detailed description of non-restrictive embodiments in the light of the attached drawings below, the application's is other Feature, objects and advantages will become more apparent upon:

Fig. 1 is that one embodiment of the application can be applied to exemplary system architecture figure therein；

Fig. 2 is the flow chart according to one embodiment of the method for generating model of the application；

Fig. 3 A is the waveform signal according to the initial audio of one embodiment of the method for generating model of the application Figure；

Fig. 3 B- Fig. 3 D be for Fig. 3 A initial audio carry out truncation obtain the operation chart of target audio；

Fig. 4 is the schematic diagram according to an application scenarios of the method for generating model of the application；

Fig. 5 is the flow chart according to another embodiment of the method for generating model of the application；

Fig. 6 is the structural schematic diagram according to one embodiment of the device for generating model of the application；

Fig. 7 is the flow chart according to one embodiment of the method for generating information of the application；

Fig. 8 is the structural schematic diagram according to one embodiment of the device for generating information of the application；

Fig. 9 is adapted for the structural schematic diagram for the computer system for realizing the electronic equipment of the embodiment of the present application.

Specific embodiment

The application is described in further detail with reference to the accompanying drawings and examples.It is understood that this place is retouched The specific embodiment stated is used only for explaining related invention, rather than the restriction to the invention.It also should be noted that in order to Convenient for description, part relevant to related invention is illustrated only in attached drawing.

It should be noted that in the absence of conflict, the features in the embodiments and the embodiments of the present application can phase Mutually combination.The application is described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

Fig. 1 is shown can the method for generating model using the embodiment of the present application or the dress for generating model It sets, alternatively, the exemplary system architecture 100 for generating the method for information or the embodiment of the device for generating information.

As shown in Figure 1, system architecture 100 may include terminal device 101,102,103, network 104 and server 105. Network 104 between terminal device 101,102,103 and server 105 to provide the medium of communication link.Network 104 can be with Including various connection types, such as wired, wireless communication link or fiber optic cables etc..

User can be used terminal device 101,102,103 and be interacted by network 104 with server 105, to receive or send out Send message etc..Various telecommunication customer end applications can be installed, such as speech recognition class is answered on terminal device 101,102,103 With, web browser applications, shopping class application, searching class application, instant messaging tools, mailbox client, social platform software Deng.

Terminal device 101,102,103 can be hardware, be also possible to software.When terminal device 101,102,103 is hard When part, it can be the various electronic equipments with audio frequency transmission function, including but not limited to smart phone, tablet computer, electronics Book reader, MP3 player (Moving Picture Experts Group Audio Layer III, dynamic image expert Compression standard audio level 3), (Moving Picture Experts Group Audio Layer IV, dynamic image are special by MP4 Family's compression standard audio level 4) player, pocket computer on knee and desktop computer etc..When terminal device 101, 102,103 when being software, may be mounted in above-mentioned cited electronic equipment.Multiple softwares or software mould may be implemented into it Block (such as providing the software of Distributed Services or software module), also may be implemented into single software or software module.? This is not specifically limited.

Server 105 can be to provide the server of various services, such as to the sound that terminal device 101,102,103 is sent Frequency provides the background server supported.Background server can carry out the processing such as audio feature extraction to the audio received, and It generates processing result (such as audio frequency characteristics of extraction).

It should be noted that the method provided by the embodiment of the present application for generating model can be held by server 105 Row, can also be executed, correspondingly, the device for generating model can be set in server by terminal device 101,102,103 In 105, also it can be set in terminal device 101,102,103.In addition, for generating information provided by the embodiment of the present application Method can be executed by server 105, can also be executed by terminal device 101,102,103, correspondingly, for generating information Device can be set in server 105, also can be set in terminal device 101,102,103.Herein, above-mentioned to be used for The executing subject for generating the method and the method for generating information of model may be the same or different.

It should be noted that server can be hardware, it is also possible to software.When server is hardware, may be implemented At the distributed server cluster that multiple servers form, individual server also may be implemented into.It, can when server is software To be implemented as multiple softwares or software module (such as providing the software of Distributed Services or software module), also may be implemented At single software or software module.It is not specifically limited herein.

It should be understood that the number of terminal device, network and server in Fig. 1 is only schematical.According to realization need It wants, can have any number of terminal device, network and server.For example, when being used to generate model method operation thereon Electronic equipment when not needing to carry out data transmission with other electronic equipments, which can only include for generating model The electronic equipment of method operation thereon.

With continued reference to Fig. 2, the process of one embodiment of the method for generating model according to the application is shown 200.The method for being used to generate model, comprising the following steps:

Step 201, the training sample set for being directed to target audio set is obtained.

In the present embodiment, (such as server shown in FIG. 1 or terminal are set the executing subject for generating the method for model It is standby) target audio can be directed to from other electronic equipments or local obtain by wired connection mode or radio connection The training sample set of set.Wherein, target audio set includes executing the audio that truncation obtains to initial audio.Training The target audio in training sample and target audio set in sample set corresponds.Training sample in training sample set This includes the characteristic and identification information of the target audio in target audio set.Identification information is used to indicate target audio packet It whether include speech audio in the audio frame included, initial audio includes speech audio.

Above-mentioned initial audio can be the various audios including speech audio, for example, above-mentioned initial audio may include but It is not limited to following any one: the speech audio with noise, including mute and speech audio audio etc..Initial audio can be with It is the audio of random length, for example, in short；It is also possible to audio frame, wherein the length of audio frame, which can be, to be preset , such as frame length can be 32 milliseconds, 25 milliseconds etc..

Target audio in above-mentioned target audio set, which can be, executes the audio that truncation obtains to initial audio.

As an example, please referring to Fig. 3 A- Fig. 3 D.Fig. 3 A is a reality according to the method for generating model of the application Apply the waveform diagram of the initial audio of example.Fig. 3 B- Fig. 3 D be for Fig. 3 A initial audio carry out truncation obtain target The operation chart of audio.

As shown in Figure 3B, if above-mentioned executing subject or other electronic equipments pair with the communication connection of above-mentioned executing subject Initial audio carries out truncated position at point of cut-off (such as line segment 301 intersect with the waveform of initial audio point), to initial audio Reason, then, after executing truncation, two sub-audios of available initial audio (including sub-audio shown in Fig. 3 C With sub-audio shown in Fig. 3 D).In this scenario, acquired target audio may include at least one in above-mentioned two sub-audio A sub-audio.That is, target audio can be sub-audio shown in Fig. 3 C, it is also possible to sub-audio shown in Fig. 3 D, can also wraps Include sub-audio shown in sub-audio shown in Fig. 3 C and Fig. 3 D.

Herein, the quantity of point of cut-off can be one, be also possible to multiple.The target obtained for an initial audio The quantity of audio can be one, be also possible to multiple.The embodiment of the present application does not limit this.

It should be noted that when target audio only includes the multiple consonants for obtain after truncation to initial audio When a part (not including carrying out obtained all sub-audios after truncation to initial audio) in frequency, truncated position at this time Reason can be referred to as truncation and delete processing.It is appreciated that truncation and delete processing be to initial audio carry out truncation after, At least one sub-audio in obtained multiple sub-audios is deleted, so that the combination of the sub-audio retained after deletion be determined For the processing of the target audio for the initial audio.

Features described above data can include but is not limited to the data of at least one following feature of audio: amplitude, frame per second, mistake Zero rate, short-time energy etc..As an example, features described above data can be Meier spectrum signature (the Mel Bank of 64 dimensions Features)。

In some optional implementations of the present embodiment, target audio set includes that following cut is executed to initial audio The target audio that disconnected processing obtains (including the first step to third step).Execution truncation can be as here below is held by above-mentioned Row main body is performed, and is also possible to be performed by the electronic equipment communicated to connect with above-mentioned executing subject.

The first step carries out truncation to initial audio, obtains consonant frequency sequence.

Second step deletes at least one sub-audio in consonant frequency sequence.

The combination of the sub-audio retained after deletion is determined as target audio by third step.

In some optional implementations of the present embodiment, above-mentioned steps two can specifically include following steps: delete Preceding first quantity sub-audio in consonant frequency sequence.Wherein, the first quantity is less than the number for the sub-audio that consonant frequency sequence includes Amount.Herein, above-mentioned first quantity, which can be, at random determines, is also possible to artificially to specify.

In some optional implementations of the present embodiment, above-mentioned steps two can specifically include following steps: delete Rear second quantity sub-audio in consonant frequency sequence.Wherein, the second quantity is less than the number for the sub-audio that consonant frequency sequence includes Amount.Herein, above-mentioned second quantity, which can be, at random determines, is also possible to artificially to specify.

In some optional implementations of the present embodiment, above-mentioned target audio set be by above-mentioned executing subject or The electronic equipment of person and the communication connection of above-mentioned executing subject execute following steps for the initial audio in initial audio set It obtains:

Firstly, generating the first random number and the second random number at random.Wherein, the first random number and the second random number are 0 Number between to 1.

It should be noted that first, second in above-mentioned first random number and the second random number is used only as distinguishing random number, The particular determination to random number is not constituted.Above-mentioned first random number and the second random number can be equal or differ.

Then, in response to determining that the first random number is less than predetermined first predetermined value, include to the initial audio Preceding third quantity sub-audio execute truncation and delete processing, by the combination of at least one sub-audio retained after deletion, really It is set to target audio.

Wherein, third quantity is less than the half of the quantity for the sub-audio that initial audio includes.First predetermined value is used for table Levy the quantity and predetermined audio of the audio that in predetermined audio set, first frame audio frame includes speech audio Gather the ratio of the quantity of sound intermediate frequency.

Above-mentioned third quantity, which can be, at random to be determined, is also possible to artificially to specify.

It is appreciated that not including the preceding third quantity that the initial audio includes in obtained target audio in this step A sub-audio only includes other sub-audios in addition to above-mentioned preceding third quantity sub-audio that the initial audio includes.

Herein, the first predetermined value is for characterizing the probability that first frame audio frame in audio includes speech audio.

Finally, including to the initial audio in response to determining that the second random number is less than predetermined second predetermined value Rear 4th quantity sub-audio execute truncation and delete processing, by the combination of at least one sub-audio retained after deletion, really It is set to target audio.

Wherein, the 4th quantity is less than the half of the quantity for the sub-audio that initial audio includes, and the second predetermined value is used for table Levy the audio that in predetermined audio set, last frame audio frame includes speech audio quantity and predetermined sound The ratio of the quantity of frequency set sound intermediate frequency.

Above-mentioned 4th quantity, which can be, at random to be determined, is also possible to artificially to specify.

It is appreciated that not including rear 4th quantity that the initial audio includes in obtained target audio in this step A sub-audio only includes other sub-audios in addition to above-mentioned rear 4th quantity sub-audio that the initial audio includes.

Herein, the second predetermined value is for characterizing the probability that last frame audio frame in audio includes speech audio.

It should be noted that first, second, in above-mentioned first quantity, the second quantity, third quantity and the 4th quantity Three, it the 4th is used only as distinguishing quantity, does not constitute the particular determination to quantity.Above-mentioned first quantity, the second quantity, third quantity It can be equal with the 4th quantity or differs.

In some optional implementations of the present embodiment, above-mentioned target audio set may include first frame audio frame Audio and last frame audio frame including speech audio include the audio of speech audio.

Above-mentioned first frame audio frame includes that the audio of speech audio can be cutting at random to initial audio execution first half The audio that disconnected and delete processing obtains.The Random Truncation Data and delete processing of above-mentioned first half refer to, if one section of initial audio Length is L sampled point, then one sampled point Q1 of selection that can be random in 0 to this sampled point segment of L/2,0 It is deleted to the data between Q1.

Above-mentioned last frame audio frame include the audio of speech audio can be initial audio is executed it is latter half of random The audio that truncation and delete processing obtain.Above-mentioned latter half of Random Truncation Data and delete processing refer to, if one section of initial audio Length be L sampled point, then one sampled point Q2 of selection that can be random in this sampled point segment of L/2 to L, 0 deletes to the data between Q2.

Optionally, the first frame audio frame that above-mentioned target audio set includes includes the audio and last frame of speech audio Audio frame is that the quantity of the audio of speech audio can be arbitrary；Be also possible to it is predetermined, for example, target audio set Including first frame audio frame be speech audio audio quantity, the quantity for the target audio for including with target audio set Ratio can be above-mentioned first predetermined value；The last frame audio frame that target audio set includes is the audio of speech audio Quantity, the ratio of the quantity for the target audio for including with target audio set can be above-mentioned second predetermined value.

Step 202, using machine learning algorithm, the characteristic for including using the training sample in training sample set as Input, using identification information corresponding with the characteristic of input as desired output, training obtains speech recognition modeling.

In the present embodiment, above-mentioned executing subject can use machine learning algorithm, by training accessed by step 201 The characteristic that training sample in sample set includes makees identification information corresponding with the characteristic of input as input For desired output, training obtains speech recognition modeling.

Specifically, above-mentioned executing subject can use machine learning algorithm, the training sample set that step 201 is got In the training sample characteristic that includes as input, identification information corresponding with the characteristic of input is defeated as expectation Out, initial model (such as Recognition with Recurrent Neural Network, convolutional neural networks) is trained, for the characteristic of each training input According to available reality output.Wherein, reality output is initial model reality output.Then, above-mentioned executing subject can adopt With gradient descent method, it is based on reality output and desired output, the parameter of initial model is adjusted, by what is obtained after each adjusting parameter Initial model of the model as training next time, and in the case where meeting preset trained termination condition, terminate training, to instruct Get speech recognition modeling.

Herein, above-mentioned executing subject can be trained initial model using batch training algorithm, can also use Random training algorithm is trained initial model, and the embodiment of the present application does not limit this.

It should be noted that the training termination condition here preset at can include but is not limited to it is at least one of following: training Time is more than preset duration；Frequency of training is more than preset times；It is small to calculate resulting difference (such as functional value of loss function) In default discrepancy threshold.

In some optional implementations of the present embodiment, above-mentioned executing subject can in accordance with the following steps, Lai Zhihang The step 202:

Training sample is chosen from training sample set, and executes following training step:

Firstly, the characteristic that the training sample of selection includes is input to initial speech identification model, obtain practical defeated Out.

Herein, above-mentioned reality output can be the output of initial speech identification model.Above-mentioned initial speech identification model It can be indiscipline, or trained but do not met the model of termination condition.

Then, it is based on reality output, determines whether initial speech identification model meets predetermined termination condition.

Above-mentioned predetermined termination condition can include but is not limited at least one of following: when the training time is more than default It is long；Frequency of training is more than preset times；It calculates resulting difference (such as functional value of loss function) and is less than default discrepancy threshold.

As an example, when above-mentioned predetermined termination condition is " frequency of training is more than preset times ", above-mentioned execution The quantity of obtained reality output can be determined as frequency of training by main body, if the quantity of obtained reality output is (i.e. Frequency of training) it is more than preset times, then it can determine that initial speech identification model meets predetermined termination condition.When above-mentioned When predetermined termination condition is " functional value for calculating resulting loss function is less than default discrepancy threshold ", above-mentioned execution master Body can calculate predetermined according to obtained reality output, and the desired output of corresponding obtained reality output The functional value of loss function can determine just if the functional value for calculating resulting loss function is less than default discrepancy threshold Beginning speech recognition modeling meets predetermined termination condition.

Finally, meeting termination condition in response to determination, the initial speech identification model of termination condition will be met, be determined as instructing The speech recognition modeling got.

In some optional implementations of the present embodiment, above-mentioned executing subject can also be unsatisfactory for above-mentioned knot in determination In the case where beam condition, it is based on obtained reality output and desired output corresponding with obtained reality output, adjustment is just The parameter value of the model parameter of beginning speech recognition modeling, and choose from training sample set the training sample of unselected mistake This, is based on parameter value initial speech identification model adjusted, continues to execute above-mentioned training step.

Herein, above-mentioned executing subject can use back propagation, by calculate reality output and with obtained reality Border exports the gradient value of the gradient of corresponding desired output, come adjust initial speech identification model model parameter parameter value. Specifically, above-mentioned executing subject can use analytic calculation gradient value, can also calculate ladder using numerical gradient calculating method Angle value, and then gradient value obtained by calculation adjusts the parameter value of the model parameter of initial speech identification model.

It is art technology it should be noted that adjusting the mode of the parameter value of model parameter above by gradient value The well-known technique that personnel study extensively, details are not described herein.

In some optional implementations of the present embodiment, the activation letter for the output layer that initial speech identification model includes For number for normalization exponential function, the cost function for the output layer that initial speech identification model includes is cross entropy cost function.

It is appreciated that the available each identification information pair of activation primitive using normalization exponential function as output layer (i.e. the sub-audio probability that belongs to non-speech audio, sub-audio belongs to the probability of initial consonant audio to the probability answered and sub-audio belongs to rhythm The probability of vowel frequency).Better training effect, example can usually be brought using the cost function of cross entropy cost function output layer Such as, training speed is faster.

In some optional implementations of the present embodiment, above-mentioned initial speech identification model can be, and there is gate to follow The Recognition with Recurrent Neural Network model of ring element, the speech recognition modeling that above-mentioned training obtains as a result, can be for gating cycle lists The Recognition with Recurrent Neural Network model of member.

Herein, using the Recognition with Recurrent Neural Network model with gating cycle unit as initial speech identification model, instruction The speech recognition modeling got, the speech recognition modeling that other opposite models are obtained as initial speech identification model, training For, it can have faster computational efficiency.

Optionally, above-mentioned initial speech identification model is also possible to Recognition with Recurrent Neural Network, convolutional neural networks etc..

With continued reference to the signal that Fig. 4, Fig. 4 are according to the application scenarios of the method for generating model of the present embodiment Figure.In the application scenarios of Fig. 4, server 401 obtains the training sample set 4001 for target audio set first.Its In, target audio set includes that the audio that truncation obtains is executed to initial audio, the training sample in training sample set It is corresponded with the target audio in target audio set, the training sample in training sample set includes in target audio set Target audio characteristic and identification information, identification information be used to indicate in the audio frame that target audio includes whether include Speech audio, initial audio include speech audio.Then, server 401 utilizes machine learning algorithm, will be in training sample set The training sample characteristic that includes as (such as the Recognition with Recurrent Neural Network mould with gating cycle unit of initial model 4002 Type) input, using identification information corresponding with the characteristic of input as the desired output of initial model 4002, training is obtained Speech recognition modeling 4003.

The method provided by the above embodiment of the application is directed to the training sample set of target audio set by obtaining, Wherein, target audio set includes that the audio that truncation obtains is executed to initial audio, the training sample in training sample set This is corresponded with the target audio in target audio set, and the training sample in training sample set includes target audio set In target audio characteristic and identification information, whether identification information be used to indicate in the audio frame that target audio includes and wrap Speech audio is included, initial audio includes speech audio, then, using machine learning algorithm, by the training in training sample set The characteristic that sample includes is as input, using identification information corresponding with the characteristic of input as desired output, training Speech recognition modeling is obtained, thus using including the characteristic and correspondence for executing the audio that truncation obtains to initial audio Identification information training sample, Lai Xunlian speech recognition modeling enriches the training method of model, in addition, using trained The accuracy of speech terminals detection can be improved in the speech recognition modeling arrived.

With further reference to Fig. 5, it illustrates the processes 500 of another embodiment of the method for generating model.The use In the process 500 for the method for generating model, comprising the following steps:

Step 501, from initial audio set, the initial audio of unselected mistake is chosen.Later, step 502 is executed.

In the present embodiment, (such as server shown in FIG. 1 or terminal are set the executing subject for generating the method for model It is standby) initial audio of unselected mistake can be chosen from initial audio set.

Step 502, the first random number and the second random number are generated at random.Later, step 503 and step 504 are executed.

In the present embodiment, above-mentioned executing subject can generate the first random number and the second random number at random.Wherein, first Random number and the second random number are the number between 0 to 1.

Herein, first, second in the first random number and the second random number is used only as distinguishing random number, composition pair The particular determination of random number.Above-mentioned first random number and the second random number can be equal, can not also wait.

Step 503, determine whether the first random number is less than predetermined first predetermined value.Later, if so, executing Step 505；If it is not, thening follow the steps 507.

In the present embodiment, above-mentioned executing subject can determine whether the first random number is more than or equal to predetermined first Numerical value.Wherein, the first predetermined value includes speech audio for characterizing in predetermined audio set, first frame audio frame Audio quantity and predetermined audio set sound intermediate frequency quantity ratio.Herein, the first predetermined value is used for table First frame audio frame includes the probability of speech audio in sign audio.

Step 504, determine whether the second random number is less than predetermined second predetermined value.Later, if so, executing Step 506；If it is not, thening follow the steps 507.

In the present embodiment, above-mentioned executing subject can determine that the second random number is less than predetermined second value.Its In, the second predetermined value is for characterizing the sound that in predetermined audio set, last frame audio frame includes speech audio The ratio of the quantity of the quantity of frequency and predetermined audio set sound intermediate frequency.Herein, the second predetermined value is for characterizing sound Last frame audio frame includes the probability of speech audio in frequency.

Step 505, truncation and delete processing are executed to the preceding third quantity sub-audio that initial audio includes, after deletion The combination of at least one sub-audio retained, is determined as target audio.Later, step 508 is executed.

In the present embodiment, the preceding third quantity sub-audio that above-mentioned executing subject can include to initial audio, which executes, cuts The combination of at least one sub-audio retained after deletion is determined as target audio by disconnected and delete processing.Wherein, third quantity The half of the quantity for the sub-audio for including less than initial audio.

Step 506, truncation and delete processing are executed to the rear 4th quantity sub-audio that initial audio includes, after deletion The combination of at least one sub-audio retained, is determined as target audio.Later, step 508 is executed.

In the present embodiment, the rear 4th quantity sub-audio that above-mentioned executing subject can include to initial audio, which executes, cuts The combination of at least one sub-audio retained after deletion is determined as target audio by disconnected and delete processing.Wherein, the 4th quantity The half of the quantity for the sub-audio for including less than initial audio.

Step 507, which is determined as target audio.Later, step 508 is executed.

In the present embodiment, which can be determined as target audio by above-mentioned executing subject.

Step 508, it determines in above-mentioned initial audio set, if there are the initial audios of unselected mistake.Later, if It is to then follow the steps 501；If it is not, thening follow the steps 509.

In the present embodiment, above-mentioned executing subject can determine in above-mentioned initial audio set, if there are unselected The initial audio crossed.

It is appreciated that in the above-mentioned initial audio set, there is no when the initial audio of unselected mistake to get having arrived mesh Mark audio set.

Step 509, the training sample set for being directed to target audio set is obtained.Later, step 510 is executed.

In the present embodiment, step 510 and the step 201 in Fig. 2 corresponding embodiment are almost the same, and which is not described herein again.

Step 510, using machine learning algorithm, the characteristic for including using the training sample in training sample set as Input, using identification information corresponding with the characteristic of input as desired output, training obtains speech recognition modeling.

In the present embodiment, step 510 and the step 202 in Fig. 2 corresponding embodiment are almost the same, and which is not described herein again.

From figure 5 it can be seen that the method for generating model compared with the corresponding embodiment of Fig. 2, in the present embodiment Process 500 highlight the step of obtaining target audio set.The scheme of the present embodiment description is for training speech recognition as a result, The audio of truncation forward for including in the training sample of model (executes truncation to the preceding third quantity sub-audio that initial audio includes And the target audio obtained after delete processing) quantity and the target audio set target audio that includes ratio of number, approach First frame audio frame includes the probability of speech audio in audio, and audio (rear 4th number for including to initial audio is truncated backward Measure sub-audio and execute the target audio obtained after truncation and delete processing) quantity and the target audio set target sound that includes The ratio of number of frequency includes the probability of speech audio close to last frame audio frame in audio, thus, the voice that training obtains Whether identification model can more accurately determine in audio comprising the position of speech audio and speech audio in audio.

With further reference to Fig. 6, as the realization to method shown in above-mentioned each figure, this application provides one kind for generating mould One embodiment of the device of type, the Installation practice is corresponding with embodiment of the method shown in Fig. 2, except following documented special Sign is outer, which can also include feature identical or corresponding with embodiment of the method shown in Fig. 2.The device specifically may be used To be applied in various electronic equipments.

As shown in fig. 6, the device 600 for generating model of the present embodiment includes: that first acquisition unit 601 and training are single Member 602.Wherein, first acquisition unit 601 is configured to obtain the training sample set for being directed to target audio set, wherein mesh Mark audio set includes that the audio that truncation obtains is executed to initial audio, training sample and target in training sample set Target audio in audio set corresponds, and the training sample in training sample set includes the target in target audio set The characteristic and identification information of audio, whether it includes voice sound that identification information is used to indicate in the audio frame that target audio includes Frequently, initial audio includes speech audio；Training unit 602 is configured to using machine learning algorithm, will be in training sample set The training sample characteristic that includes as input, identification information corresponding with the characteristic of input is defeated as expectation Out, training obtains speech recognition modeling.

It in the present embodiment, can be by wired connection side for generating the first acquisition unit 601 of the device 600 of model Formula or radio connection are directed to the training sample set of target audio set from other electronic equipments or local acquisition.Its In, target audio set includes executing the audio that truncation obtains to initial audio.Training sample in training sample set It is corresponded with the target audio in target audio set.Training sample in training sample set includes in target audio set Target audio characteristic and identification information.Identification information be used to indicate in the audio frame that target audio includes whether include Speech audio, initial audio include speech audio.

In the present embodiment, above-mentioned training unit 602 can use machine learning algorithm, by the instruction in training sample set Practice the characteristic that sample includes to instruct as input using identification information corresponding with the characteristic of input as desired output Get speech recognition modeling.

In some optional implementations of the present embodiment, target audio set includes that following place is executed to initial audio Manage obtained target audio:

Second step deletes at least one sub-audio in consonant frequency sequence.

In some optional implementations of the present embodiment, at least one sub-audio in consonant frequency sequence, packet are deleted It includes: deleting the preceding first quantity sub-audio in consonant frequency sequence, wherein the first quantity is less than the consonant that consonant frequency sequence includes The quantity of frequency.

In some optional implementations of the present embodiment, at least one sub-audio in consonant frequency sequence, packet are deleted It includes: deleting the rear second quantity sub-audio in consonant frequency sequence.Wherein, the second quantity is less than the consonant that consonant frequency sequence includes The quantity of frequency.

In some optional implementations of the present embodiment, target audio set is by in initial audio set Initial audio, execute following steps and obtain:

Then, in response to determining that the first random number is less than predetermined first predetermined value, include to the initial audio Preceding third quantity sub-audio execute truncation and delete processing, by the combination of at least one sub-audio retained after deletion, really It is set to target audio.Wherein, third quantity is less than the half of the quantity for the sub-audio that initial audio includes, and the first predetermined value is used In characterize the quantity of the audio that in predetermined audio set, first frame audio frame includes speech audio with it is predetermined The ratio of the quantity of audio set sound intermediate frequency.

Finally, including to the initial audio in response to determining that the second random number is less than predetermined second predetermined value Rear 4th quantity sub-audio execute truncation and delete processing, by the combination of at least one sub-audio retained after deletion, really It is set to target audio.Wherein, the 4th quantity is less than the half of the quantity for the sub-audio that initial audio includes, and the second predetermined value is used In the quantity for characterizing the audio that in predetermined audio set, last frame audio frame includes speech audio and predefine Audio set sound intermediate frequency quantity ratio.

In some optional implementations of the present embodiment, training unit includes: training module (not shown) quilt It is configured to choose training sample from training sample set, and executes following training step: including by the training sample of selection Characteristic be input to initial speech identification model, obtain reality output, wherein reality output is initial speech identification model Output；Based on reality output, determine whether initial speech identification model meets predetermined termination condition；In response to determination Meet termination condition, the initial speech identification model of termination condition will be met, is determined as the speech recognition modeling that training obtains.

In some optional implementations of the present embodiment, the device 600 further include: adjustment unit (not shown) It is configured in response to determination and is unsatisfactory for termination condition, based on obtained reality output and corresponding with obtained reality output Desired output, adjust the parameter value of the model parameter of initial speech identification model, and chosen not from training sample set The training sample being selected is based on parameter value initial speech identification model adjusted, continues to execute training step.

In some optional implementations of the present embodiment, speech recognition modeling is the circulation with gating cycle unit Neural network model.

The device provided by the above embodiment of the application is obtained by first acquisition unit 601 and is directed to target audio set Training sample set, wherein target audio set includes executing the obtained audio of truncation, training sample to initial audio The target audio in training sample and target audio set in set corresponds, the training sample packet in training sample set The characteristic and identification information of the target audio in target audio set are included, identification information is used to indicate what target audio included It whether include speech audio in audio frame, initial audio includes speech audio, and then, training unit 602 is calculated using machine learning Method, the characteristic for including using the training sample in training sample set, will be corresponding with the characteristic of input as input Identification information obtains speech recognition modeling as desired output, training, thus using including executing truncation to initial audio The training sample of the characteristic of obtained audio and corresponding identification information, Lai Xunlian speech recognition modeling, enriches model Training method, in addition, the accuracy of speech terminals detection can be improved using the obtained speech recognition modeling of training.

With continued reference to Fig. 7, the process of one embodiment of the method for generating information according to the application is shown 700.The method for being used to generate information, comprising the following steps:

Step 701, target audio is obtained.

In the present embodiment, (such as server shown in FIG. 1 or terminal are set the executing subject for generating the method for information It is standby) target audio can be obtained by wired connection mode or radio connection from other electronic equipments or local.Its In, above-mentioned target audio can be the various audios to carry out speech terminals detection to it including speech audio.

Step 702, which is input to speech recognition mould trained in advance by the audio frame for including for target audio Type obtains the probability that the audio frame includes speech audio.

In the present embodiment, the audio frame at least one audio frame for including for target audio, above-mentioned executing subject The characteristic of the audio frame can be input to speech recognition modeling trained in advance, obtaining the audio frame includes speech audio Probability.Wherein, above-mentioned speech recognition modeling can be above-mentioned executing subject or the electricity with the communication connection of above-mentioned executing subject Sub- equipment is obtained according to method training described in any embodiment in the method for generating model as shown in Figure 2.

It is appreciated that in general, according to the speech recognition modeling that above-mentioned training method obtains, during actual use, The probability in audio frame comprising speech audio can be exported, in turn, above-mentioned executing subject can be by comparing obtained probability With the size relation between predetermined probabilities threshold value, so that it is determined that in audio frame whether include speech audio.

Step 703, the size relation based on obtained probability and predetermined probability threshold value, generates target audio Speech terminals detection result.

In the present embodiment, above-mentioned executing subject can be based on the big of obtained probability and predetermined probability threshold value Small relationship generates the speech terminals detection result of target audio.

Above-mentioned speech terminals detection result can serve to indicate that the start bit for the speech audio for including in above-mentioned target audio It sets and final position.

For example, if obtained probability is greater than predetermined probability threshold value, above-mentioned executing subject can determine The corresponding audio frame of the probability is speech audio, to obtain the speech terminals detection result of target audio.

In some optional implementations of the present embodiment, predetermined threshold value includes predetermined first threshold With predetermined second threshold, first threshold is greater than second threshold, and speech terminals detection result includes what target audio included The location information of starting point of the speech audio in target audio and the location information of terminal.Above-mentioned executing subject can be by as a result, Above-mentioned steps 703 are executed according to following steps:

Firstly, the size relation based on obtained probability and predetermined first threshold, generating target audio includes Starting point of the speech audio in target audio location information.

For example, above-mentioned executing subject can determine in the probability sequence for the audio frame sequence for including for target audio, the The corresponding audio frame of the probability is determined as the language that target audio includes by one probability for being greater than predetermined first threshold Starting point of the sound audio in target audio, to obtain the location information of starting point of the speech audio in target audio.

Then, the size relation based on obtained probability and predetermined second threshold, generating target audio includes Terminal of the speech audio in target audio location information.

For example, above-mentioned executing subject can determine in the probability sequence for the audio frame sequence for including for target audio, most The corresponding audio frame of the probability is determined as what target audio included by the probability that the latter is greater than predetermined second threshold Terminal of the speech audio in target audio, to obtain the location information of terminal of the speech audio in target audio.

After determining whether each audio frame that target audio includes includes speech audio, above-mentioned executing subject can be true First in the audio frame sequence that the audio that sets the goal includes and the last one include the audio frame of speech audio, and will determine The first audio frame including speech audio, be determined as the initial position of speech audio for including in target audio, will determine The last one include the audio frame of speech audio, be determined as the final position of speech audio for including in target audio, thus Speech terminals detection result is obtained.

Optionally, whether each audio frame that above-mentioned target audio can also directly be included by above-mentioned executing subject includes language The result of sound audio is determined as speech terminals detection result.For example, if above-mentioned target audio is made of 10 frame audio frames.Its In, the 2nd frame to the 9th frame audio frame includes speech audio, and the 1st frame and the 10th frame audio frame do not include speech audio.So, above-mentioned The sequence { 0,1,1,1,1,1,1,1,1,0 } of characterization the above results can be generated in executing subject, wherein first in above-mentioned sequence A element is used to indicate whether first audio frame that target audio includes includes speech audio, the 2nd in above-mentioned sequence member Whether the 2nd audio frame that element can serve to indicate that target audio includes includes speech audio, and so on." 0 " can characterize It does not include speech audio, " 1 " can be characterized including speech audio.As a result, above-mentioned executing subject can by sequence 0,1,1,1,1, 1,1,1,1,0 } it is determined directly as speech terminals detection result.Under this application scenarios, by the speech terminals detection as a result, can To determine that target audio is made of 10 frame audio frames.Wherein, the 2nd frame to the 9th frame audio frame includes speech audio, the 1st frame and 10 frame audio frames do not include speech audio.

The method provided by the above embodiment of the application, by obtaining target audio, wherein target audio includes voice sound Frequently, then, which is input to speech recognition modeling trained in advance, obtained by the audio frame for including for target audio The audio frame includes the probability of speech audio, wherein speech recognition modeling is in the method according to such as above-mentioned for generating information What the method training of any embodiment obtained, finally, being closed based on the size of obtained probability and predetermined probability threshold value System generates the speech terminals detection of target audio as a result, speech recognition modeling is applied to speech terminals detection as a result, thus The order of accuarcy for improving speech terminals detection enriches the mode of speech terminals detection.

With further reference to Fig. 8, as the realization to method shown in above-mentioned each figure, this application provides one kind for generating letter One embodiment of the device of breath, the Installation practice is corresponding with embodiment of the method shown in Fig. 7, except following documented special Sign is outer, which can also include feature identical or corresponding with embodiment of the method shown in Fig. 7.The device specifically may be used To be applied in various electronic equipments.

As shown in figure 8, the device 800 for generating information of the present embodiment includes: second acquisition unit 801, input list Member 802 and generation unit 803.Wherein, second acquisition unit 801 is configured to obtain target audio, wherein target audio includes Speech audio；Input unit 802 is configured to the audio frame for including for target audio, which is input to preparatory training Speech recognition modeling, obtain the probability that the audio frame includes speech audio, wherein speech recognition modeling is as above-mentioned for giving birth to It is obtained at the method training of any embodiment in the method for model；Generation unit 803 is configured to based on obtained probability With the size relation of predetermined probability threshold value, the speech terminals detection result of target audio is generated.

It in the present embodiment, can be by wired connection side for generating the second acquisition unit 801 of the device 800 of information Formula or radio connection obtain target audio from other electronic equipments or local.

Above-mentioned target audio can be the various audios including speech audio.

In the present embodiment, at least one audio frame that the target audio got for second acquisition unit 801 includes In audio frame, the characteristic of the audio frame can be input in advance trained speech recognition mould by above-mentioned input unit 802 Type obtains the probability that the audio frame includes speech audio.Wherein, above-mentioned speech recognition modeling can be above-mentioned executing subject or Electronic equipment with the communication connection of above-mentioned executing subject is according to any implementation in the method for generating model as shown in Figure 2 The training of method described in example obtains.

In the present embodiment, it is closed based on the size of the obtained probability of input unit 802 and predetermined probability threshold value The speech terminals detection result of target audio can be generated in system, above-mentioned generation unit 803.Wherein, above-mentioned speech terminals detection knot Fruit can serve to indicate that initial position and the final position for the speech audio for including in above-mentioned target audio.

In some optional implementations of the present embodiment, predetermined threshold value includes predetermined first threshold With predetermined second threshold, first threshold is greater than second threshold, and speech terminals detection result includes what target audio included The location information of starting point of the speech audio in target audio and the location information of terminal.And generation unit includes: first raw It is configured to the size relation based on obtained probability and predetermined first threshold at module (not shown), is generated The location information of starting point of the speech audio that target audio includes in target audio；Second generation module (not shown) quilt It is configured to the size relation based on obtained probability and predetermined second threshold, generates the voice sound that target audio includes The location information of terminal of the frequency in target audio.

The device provided by the above embodiment of the application obtains target audio by second acquisition unit 801, wherein mesh Mark with phonetic symbols frequency includes speech audio, and then, input unit 802 is directed to the audio frame that target audio includes, which is input to Trained speech recognition modeling in advance, obtains the probability that the audio frame includes speech audio, wherein speech recognition modeling is as above The method training for stating any embodiment in the method for generating model obtains, finally, generation unit 803 be based on it is obtained The size relation of probability and predetermined probability threshold value generates the speech terminals detection of target audio as a result, as a result, by voice Identification model is applied to speech terminals detection, to improve the order of accuarcy of speech terminals detection, enriches sound end inspection The mode of survey.

Below with reference to Fig. 9, it illustrates the computer systems 900 for the electronic equipment for being suitable for being used to realize the embodiment of the present application Structural schematic diagram.Electronic equipment shown in Fig. 9 is only an example, function to the embodiment of the present application and should not use model Shroud carrys out any restrictions.

As shown in figure 9, computer system 900 includes central processing unit (CPU) 901, it can be read-only according to being stored in Program in memory (ROM) 902 or be loaded into the program in random access storage device (RAM) 903 from storage section 908 and Execute various movements appropriate and processing.In RAM 903, also it is stored with system 900 and operates required various programs and data. CPU901, ROM 902 and RAM 903 is connected with each other by bus 904.Input/output (I/O) interface 905 is also connected to always Line 904.

I/O interface 905 is connected to lower component: the importation 906 including keyboard, mouse etc.；It is penetrated including such as cathode The output par, c 907 of spool (CRT), liquid crystal display (LCD) etc. and loudspeaker etc.；Storage section 908 including hard disk etc.； And the communications portion 909 of the network interface card including LAN card, modem etc..Communications portion 909 via such as because The network of spy's net executes communication process.Driver 910 is also connected to I/O interface 905 as needed.Detachable media 911, such as Disk, CD, magneto-optic disk, semiconductor memory etc. are mounted on as needed on driver 910, in order to read from thereon Computer program be mounted into storage section 908 as needed.

Particularly, in accordance with an embodiment of the present disclosure, it may be implemented as computer above with reference to the process of flow chart description Software program.For example, embodiment of the disclosure includes a kind of computer program product comprising be carried on computer-readable medium On computer program, which includes the program code for method shown in execution flow chart.In such reality It applies in example, which can be downloaded and installed from network by communications portion 909, and/or from detachable media 911 are mounted.When the computer program is executed by central processing unit (CPU) 901, limited in execution the present processes Above-mentioned function.

It should be noted that computer-readable medium described herein can be computer-readable signal media or meter Calculation machine readable storage medium storing program for executing either the two any combination.Computer readable storage medium for example can be --- but not Be limited to --- electricity, magnetic, optical, electromagnetic, infrared ray or semiconductor system, device or device, or any above combination.Meter The more specific example of calculation machine readable storage medium storing program for executing can include but is not limited to: have the electrical connection, just of one or more conducting wires Taking formula computer disk, hard disk, random access storage device (RAM), read-only memory (ROM), erasable type may be programmed read-only storage Device (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), light storage device, magnetic memory device, Or above-mentioned any appropriate combination.In this application, computer readable storage medium can be it is any include or storage journey The tangible medium of sequence, the program can be commanded execution system, device or device use or in connection.And at this In application, computer-readable signal media may include in a base band or as carrier wave a part propagate data-signal, Wherein carry computer-readable program code.The data-signal of this propagation can take various forms, including but unlimited In electromagnetic signal, optical signal or above-mentioned any appropriate combination.Computer-readable signal media can also be that computer can Any computer-readable medium other than storage medium is read, which can send, propagates or transmit and be used for By the use of instruction execution system, device or device or program in connection.Include on computer-readable medium Program code can transmit with any suitable medium, including but not limited to: wireless, electric wire, optical cable, RF etc. are above-mentioned Any appropriate combination.

The calculating of the operation for executing the application can be write with one or more programming languages or combinations thereof Machine program code, described program design language include object-oriented programming language-such as Python, Java, Smalltalk, C++ further include conventional procedural programming language-such as " C " language or similar program design language Speech.Program code can be executed fully on the user computer, partly be executed on the user computer, as an independence Software package execute, part on the user computer part execute on the remote computer or completely in remote computer or It is executed on server.In situations involving remote computers, remote computer can pass through the network of any kind --- packet It includes local area network (LAN) or wide area network (WAN)-is connected to subscriber computer, or, it may be connected to outer computer (such as benefit It is connected with ISP by internet).

Flow chart and block diagram in attached drawing are illustrated according to the system of the various embodiments of the application, method and computer journey The architecture, function and operation in the cards of sequence product.In this regard, each box in flowchart or block diagram can generation A part of one module, program segment or code of table, a part of the module, program segment or code include one or more use The executable instruction of the logic function as defined in realizing.It should also be noted that in some implementations as replacements, being marked in box The function of note can also occur in a different order than that indicated in the drawings.For example, two boxes succeedingly indicated are actually It can be basically executed in parallel, they can also be executed in the opposite order sometimes, and this depends on the function involved.Also it to infuse Meaning, the combination of each box in block diagram and or flow chart and the box in block diagram and or flow chart can be with holding The dedicated hardware based system of functions or operations as defined in row is realized, or can use specialized hardware and computer instruction Combination realize.

Being described in unit involved in the embodiment of the present application can be realized by way of software, can also be by hard The mode of part is realized.Described unit also can be set in the processor, for example, can be described as: a kind of processor packet Include first acquisition unit and training unit.Wherein, the title of these units is not constituted to the unit itself under certain conditions Restriction, for example, first acquisition unit be also described as " obtain for target audio set training sample set list Member ".

As on the other hand, present invention also provides a kind of computer-readable medium, which be can be Included in electronic equipment described in above-described embodiment；It is also possible to individualism, and without in the supplying electronic equipment. Above-mentioned computer-readable medium carries one or more program, when said one or multiple programs are held by the electronic equipment When row, so that the electronic equipment: obtaining the training sample set for being directed to target audio set, wherein target audio set includes The audio that truncation obtains is executed to initial audio, the mesh in training sample and target audio set in training sample set Mark with phonetic symbols frequency corresponds, and the training sample in training sample set includes the characteristic of the target audio in target audio set And identification information, whether it includes speech audio, initial audio packet that identification information is used to indicate in the audio frame that target audio includes Include speech audio；Using machine learning algorithm, the characteristic for including using the training sample in training sample set as input, Using identification information corresponding with the characteristic of input as desired output, training obtains speech recognition modeling.

Above description is only the preferred embodiment of the application and the explanation to institute's application technology principle.Those skilled in the art Member is it should be appreciated that invention scope involved in the application, however it is not limited to technology made of the specific combination of above-mentioned technical characteristic Scheme, while should also cover in the case where not departing from foregoing invention design, it is carried out by above-mentioned technical characteristic or its equivalent feature Any combination and the other technical solutions formed.Such as features described above has similar function with (but being not limited to) disclosed herein Can technical characteristic replaced mutually and the technical solution that is formed.

Claims

1. a kind of method for generating model, comprising:

Obtain the training sample set for being directed to target audio set, wherein the target audio set includes holding to initial audio The audio that row truncation obtains, the target sound in training sample and the target audio set in the training sample set Frequency corresponds, and the training sample in the training sample set includes the feature of the target audio in the target audio set Data and identification information, whether it includes speech audio, initial sound that identification information is used to indicate in the audio frame that target audio includes Frequency includes speech audio；

Using machine learning algorithm, the characteristic for including using the training sample in the training sample set, will as input Identification information corresponding with the characteristic of input obtains speech recognition modeling as desired output, training.

2. according to the method described in claim 1, wherein, the target audio set includes executing following processing to initial audio Obtained target audio:

Truncation is carried out to initial audio, obtains consonant frequency sequence；

Delete at least one sub-audio in the consonant frequency sequence；

By the combination of the sub-audio retained after deletion, it is determined as target audio.

3. according to the method described in claim 2, wherein, described at least one sub-audio deleted in the consonant frequency sequence, Include:

Delete the preceding first quantity sub-audio in the consonant frequency sequence, wherein first quantity is less than the sub-audio The quantity for the sub-audio that sequence includes.

4. according to the method described in claim 2, wherein, described at least one sub-audio deleted in the consonant frequency sequence, Include:

Delete the rear second quantity sub-audio in the consonant frequency sequence, wherein second quantity is less than the sub-audio The quantity for the sub-audio that sequence includes.

5. according to the method described in claim 1, wherein, the target audio set is by in initial audio set Initial audio executes what following steps obtained:

It is random to generate the first random number and the second random number, wherein first random number and second random number are 0 to arrive Number between 1；

It is less than predetermined first predetermined value in response to determination first random number, the before including to the initial audio Three quantity sub-audios execute truncation and delete processing and the combination of at least one sub-audio retained after deletion are determined as mesh Mark with phonetic symbols frequency, wherein the third quantity is less than the half of the quantity for the sub-audio that initial audio includes, first predetermined value The quantity for the audio for including speech audio for characterizing in predetermined audio set, first frame audio frame and it is described in advance The ratio of the quantity of determining audio set sound intermediate frequency；

It is less than predetermined second predetermined value in response to determination second random number, include to the initial audio rear the Four quantity sub-audios execute truncation and delete processing and the combination of at least one sub-audio retained after deletion are determined as mesh Mark with phonetic symbols frequency, wherein the 4th quantity is less than the half of the quantity for the sub-audio that initial audio includes, second predetermined value The quantity for the audio for including speech audio for characterizing in predetermined audio set, last frame audio frame with it is described pre- The first ratio of the quantity of determining audio set sound intermediate frequency.

6. it is described to utilize machine learning algorithm according to the method described in claim 1, wherein, it will be in the training sample set The training sample characteristic that includes as input, identification information corresponding with the characteristic of input is defeated as expectation Out, training obtains speech recognition modeling, comprising:

Training sample is chosen from the training sample set, and executes following training step: by the training sample packet of selection The characteristic included is input to initial speech identification model, obtains reality output, wherein reality output is initial speech identification mould The output of type；Based on reality output, determine whether initial speech identification model meets predetermined termination condition；In response to true Surely meet the termination condition, the initial speech identification model of the termination condition will be met, be determined as the voice that training obtains Identification model.

7. according to the method described in claim 6, wherein, the method also includes:

It is unsatisfactory for the termination condition in response to determination, based on obtained reality output and corresponding with obtained reality output Desired output, adjust the parameter value of the model parameter of initial speech identification model, and selected from the training sample set The training sample of unselected mistake is taken, parameter value initial speech identification model adjusted is based on, continues to execute the training step Suddenly.

8. method according to claim 6 or 7, wherein the activation primitive for the output layer that initial speech identification model includes To normalize exponential function, the cost function for the output layer that initial speech identification model includes is cross entropy cost function.

9. method described in one of -8 according to claim 1, wherein the speech recognition modeling is with gating cycle unit Recognition with Recurrent Neural Network model.

10. a kind of method for generating information, comprising:

Obtain target audio, wherein the target audio includes speech audio；

For the audio frame that the target audio includes, which is input to speech recognition modeling trained in advance, is obtained The audio frame includes the probability of speech audio, wherein the speech recognition modeling is according to as claimed in one of claims 1 to 9 Method training obtains；

Size relation based on obtained probability and predetermined probability threshold value, generates the sound end of the target audio Testing result.

11. according to the method described in claim 10, wherein, the predetermined threshold value includes predetermined first threshold With predetermined second threshold, the first threshold is greater than the second threshold, and the speech terminals detection result includes institute State the location information of the starting point of speech audio that target audio includes in the target audio and the location information of terminal；And

The size relation based on obtained probability and predetermined threshold value, generates the sound end of the target audio Testing result, comprising:

Size relation based on obtained probability and predetermined first threshold generates the voice that the target audio includes The location information of starting point of the audio in the target audio；

Size relation based on obtained probability and predetermined second threshold generates the voice that the target audio includes The location information of terminal of the audio in the target audio.

12. a kind of for generating the device of model, comprising:

First acquisition unit is configured to obtain the training sample set for being directed to target audio set, wherein the target audio Set includes that the audio that truncation obtains is executed to initial audio, training sample and the mesh in the training sample set The target audio marked in audio set corresponds, and the training sample in the training sample set includes the target audio collection The characteristic and identification information of target audio in conjunction, identification information be used to indicate in the audio frame that target audio includes whether Including speech audio, initial audio includes speech audio；

Training unit is configured to using machine learning algorithm, the spy for including by the training sample in the training sample set Data are levied as input, using identification information corresponding with the characteristic of input as desired output, training obtains speech recognition Model.

13. device according to claim 12, wherein the target audio set includes executing following place to initial audio Manage obtained target audio:

Delete at least one sub-audio in the consonant frequency sequence；

14. device according to claim 13, wherein described at least one consonant deleted in the consonant frequency sequence Frequently, comprising:

15. device according to claim 13, wherein described at least one consonant deleted in the consonant frequency sequence Frequently, comprising:

16. device according to claim 12, wherein the target audio set is by in initial audio set Initial audio, execute following steps and obtain:

17. device according to claim 12, wherein the training unit includes:

Training module is configured to choose training sample from the training sample set, and executes following training step: will The characteristic that the training sample of selection includes is input to initial speech identification model, obtains reality output, wherein reality output It is the output of initial speech identification model；Based on reality output, it is predetermined to determine whether initial speech identification model meets Termination condition；Meet the termination condition in response to determination, the initial speech identification model of the termination condition will be met, determines The speech recognition modeling obtained for training.

18. device according to claim 17, wherein described device further include:

Adjustment unit, be configured in response to determination be unsatisfactory for the termination condition, based on obtained reality output and with institute The corresponding desired output of obtained reality output adjusts the parameter value of the model parameter of initial speech identification model, and from institute The training sample for choosing unselected mistake in training sample set is stated, parameter value initial speech identification model adjusted is based on, Continue to execute the training step.

19. device described in 7 or 18 according to claim 1, wherein the activation letter for the output layer that initial speech identification model includes For number for normalization exponential function, the cost function for the output layer that initial speech identification model includes is cross entropy cost function.

20. device described in one of 2-19 according to claim 1, wherein the speech recognition modeling is with gating cycle list The Recognition with Recurrent Neural Network model of member.

21. a kind of for generating the device of information, comprising:

Second acquisition unit is configured to obtain target audio, wherein the target audio includes speech audio；

Input unit is configured to the audio frame for including for the target audio, which is input to training in advance Speech recognition modeling obtains the probability that the audio frame includes speech audio, wherein the speech recognition modeling is according to such as right It is required that the training of method described in one of 12-20 obtained；

Generation unit, is configured to the size relation based on obtained probability and predetermined probability threshold value, described in generation The speech terminals detection result of target audio.

22. device according to claim 21, wherein the predetermined threshold value includes predetermined first threshold With predetermined second threshold, the first threshold is greater than the second threshold, and the speech terminals detection result includes institute State the location information of the starting point of speech audio that target audio includes in the target audio and the location information of terminal；And

The generation unit includes:

First generation module is configured to the size relation based on obtained probability and predetermined first threshold, generates The location information of starting point of the speech audio that the target audio includes in the target audio；

Second generation module is configured to the size relation based on obtained probability and predetermined second threshold, generates The location information of terminal of the speech audio that the target audio includes in the target audio.

23. a kind of electronic equipment, comprising:

One or more processors；

Storage device is stored thereon with one or more programs,

When one or more of programs are executed by one or more of processors, so that one or more of processors are real The now method as described in any in claim 1-11.

24. a kind of computer-readable medium, is stored thereon with computer program, wherein real when described program is executed by processor The now method as described in any in claim 1-11.