CN110516104A - Song recommendations method, apparatus and computer storage medium - Google Patents
Song recommendations method, apparatus and computer storage medium Download PDFInfo
- Publication number
- CN110516104A CN110516104A CN201910797899.7A CN201910797899A CN110516104A CN 110516104 A CN110516104 A CN 110516104A CN 201910797899 A CN201910797899 A CN 201910797899A CN 110516104 A CN110516104 A CN 110516104A
- Authority
- CN
- China
- Prior art keywords
- audio
- song
- trained
- sound quality
- quality score
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/60—Information retrieval; Database structures therefor; File system structures therefor of audio data
- G06F16/63—Querying
- G06F16/635—Filtering based on additional data, e.g. user or group profiles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/60—Information retrieval; Database structures therefor; File system structures therefor of audio data
- G06F16/68—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/683—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
This application discloses a kind of song recommendations method, apparatus and computer storage mediums, belong to multimedia technology field.The described method includes: the audio frequency characteristics for obtaining every song in number of songs to be recommended determine recommendation list according to the audio frequency characteristics of song every in number of songs.Wherein, audio frequency characteristics are used to indicate the tone, pitch and tone color of audio, it can thus be appreciated that, the embodiment of the present application is to recommend song according to the tone of song itself, pitch and tone color, the song for recommending high quality to user so may be implemented, and can be avoided and influenced during recommending song by the factor and individual subjective factor of the user, or influenced by extraneous factor, so that the song recommended meets the needs of user is to high quality song.In addition, song recommendations model is configured in the application, for quickly determining the sound quality score of every song, to improve the efficiency of determining recommendation list.
Description
Technical field
This application involves multimedia technology field, in particular to a kind of song recommendations method, apparatus and computer storage are situated between
Matter.
Background technique
After user logs in some music class application program by terminal, the corresponding server of music class application program is logical
The Chang Huixiang terminal recommends song, to promote the experience that the user uses the music class application program.
In the related technology, server can listen song data according to the history of the user, alternatively, according to the history of all songs
Data are listened to, a recommendation list is determined, and the recommendation list is sent to terminal, which is shown by terminal.
The process of above-mentioned recommendation song is easy to be influenced by the factor and individual subjective factor of the user, such as user preference, or by outer
The influence of boundary's factor, such as clicking rate or playback volume, so that the song recommended be caused possibly can not to be commented in terms of musical qualities
Sentence, and is unfavorable for excavating the musical qualities newly sung.
Summary of the invention
The embodiment of the present application provides a kind of song recommendations method, apparatus and computer storage medium, and recommendation can be improved
Song quality so that recommend song meet the needs of user is to high quality song.The technical solution is as follows:
In a first aspect, providing a kind of song recommendations method, the method includes song recommendations model, the song recommendations
Model is obtained by the neural network model training of initialization, which comprises
Obtain number of songs to be recommended；
Determine the audio frequency characteristics of every song in the number of songs, the audio frequency characteristics be used to indicate audio tone,
Pitch and tone color；
According to the audio frequency characteristics of every song and the song recommendations model, the sound quality score of every song is determined；
According to the sound quality score of song every in the number of songs, recommendation list is determined.
Optionally, the method also includes: obtain multiple trained audios and with the multiple trained audio it is one-to-one
Sound quality score and multiple verifying audios and with the multiple one-to-one sound quality score of verifying audio；It determines the multiple
The audio frequency characteristics of each trained audio and the multiple audio for verifying each verifying audio in audio are special in training audio
Sign；According to each trained audio in the one-to-one sound quality score of the multiple trained audio, the multiple trained audio
Audio frequency characteristics, and with each verifying in the one-to-one sound quality score of the multiple verifying audio, the multiple verifying audio
The audio frequency characteristics of audio are trained the neural network model of the initialization, obtain the song recommendations model.
Optionally, the basis and the one-to-one sound quality score of the multiple trained audio, the multiple trained audio
In each trained audio audio frequency characteristics, and with the one-to-one sound quality score of the multiple verifying audio, the multiple test
The audio frequency characteristics for demonstrate,proving each verifying audio in audio, are trained the neural network model of the initialization, obtain the song
Bent recommended models, comprising: according to in the one-to-one sound quality score of the multiple trained audio, the multiple trained audio
During the audio frequency characteristics of each trained audio are trained the neural network model of the initialization, training every time is obtained
Neural network model later；According to the one-to-one sound quality score of the multiple verifying audio and the multiple verifying audio
In it is each verifying audio audio frequency characteristics and the neural network model after training, determine the value of loss function, it is described
Loss function is used to indicate the gap between the predicted value of neural network model and true value；If the value of the loss function reaches
To minimum value, then the neural network model after training is determined as the song recommendations model；If the loss letter
Several values does not reach minimum value, then return execution according to the one-to-one sound quality score of the multiple trained audio and described
The step that the audio frequency characteristics of each trained audio are trained the neural network model after training in multiple trained audios
Suddenly.
Optionally, it is described the neural network model after training is determined as the song recommendations model before, also
It include: according to the sound with each testing audio in the one-to-one sound quality score of multiple testing audios, the multiple testing audio
Frequency feature tests the neural network model after training；Test result is sent to back-stage management terminal；If
The termination training instruction that the back-stage management terminal is sent is received, then is executed the neural network model after training is true
It is set to the operation of the song recommendations model.
Optionally, it is described obtain multiple trained audios and with the multiple trained one-to-one sound quality score of audio, with
And multiple verifying audios and with the multiple one-to-one sound quality score of verifying audio, comprising: for the first audio, obtain needle
To the score of first audio mark, the sound quality score of first audio is obtained；Alternatively, going through according to first audio
History collects number and history broadcasting time, determines the sound quality score of first audio；Wherein, first audio is described more
Any audio in a trained audio and the multiple verifying audio.
Optionally, before acquisition number of songs to be recommended, further includes: the history for obtaining user listens to data；Root
Song data are listened to filter out the number of songs from song database according to the history of the user.
Second aspect, provides a kind of song recommendations device, and described device includes song recommendations model, the song recommendations
Model is obtained by the neural network model training of initialization, and described device includes:
First obtains module, for obtaining number of songs to be recommended；
First determining module, for determining that the audio frequency characteristics of every song in the number of songs, the audio frequency characteristics are used
In tone, pitch and the tone color of instruction audio；
Second determining module determines every for the audio frequency characteristics and the song recommendations model according to every song
The sound quality score of song determines recommendation list according to the sound quality score of song every in the number of songs.
Optionally, described device further include:
Second obtains module, for obtaining multiple trained audios and dividing with the multiple trained one-to-one sound quality of audio
Several and multiple verifying audios and with the multiple one-to-one sound quality score of verifying audio；
Third determining module, for determining the audio frequency characteristics of each trained audio in the multiple trained audio, Yi Jisuo
State the audio frequency characteristics of each verifying audio in multiple verifying audios；
Training module, for according to the one-to-one sound quality score of the multiple trained audio, the multiple trained sound
The audio frequency characteristics of each trained audio in frequency, and with the one-to-one sound quality score of the multiple verifying audio, the multiple
The audio frequency characteristics for verifying each verifying audio in audio, are trained the neural network model of the initialization, obtain described
Song recommendations model.
Optionally, the training module, is specifically used for:
According to each training in the one-to-one sound quality score of the multiple trained audio, the multiple trained audio
During the audio frequency characteristics of audio are trained the neural network model of the initialization, mind every time after training is obtained
Through network model；According to each tested in the one-to-one sound quality score of the multiple verifying audio and the multiple verifying audio
The audio frequency characteristics and the neural network model after training for demonstrate,proving audio, determine the value of loss function, the loss function
The gap being used to indicate between the predicted value of neural network model and true value；If the value of the loss function reaches minimum
Value, then be determined as the song recommendations model for the neural network model after training；If the value of the loss function
Do not reach minimum value, then returns to execution basis and the multiple trained one-to-one sound quality score of audio and the multiple instruction
Cultivate the voice the step of audio frequency characteristics of each trained audio are trained the neural network model after training in frequency.
Optionally, the training module, also particularly useful for:
According to each testing audio in the one-to-one sound quality score of multiple testing audios, the multiple testing audio
Audio frequency characteristics test the neural network model after training；Test result is sent to background management device；Such as
Fruit receives the termination training instruction that the background management device is sent, then executes the neural network model after training
It is determined as the operation of the song recommendations model.
Optionally, described second module is obtained, is specifically used for:
For the first audio, the score for first audio mark is obtained, obtains the sound quality point of first audio
Number；Alternatively, collecting number and history broadcasting time according to the history of first audio, the sound quality point of first audio is determined
Number；Wherein, first audio is any audio in the multiple trained audio and the multiple verifying audio.
Optionally, described device further include:
Third obtains module, and the history for obtaining user listens to data；
Screening module, for listening song data to filter out more first songs from song database according to the history of the user
It is bent.
The third aspect, provides a kind of song recommendations device, and described device includes:
Processor；
Memory for storage processor executable instruction；
Wherein, the processor is configured to the step of executing any one method described in above-mentioned first aspect.
Fourth aspect provides a kind of computer readable storage medium, finger is stored on the computer readable storage medium
The step of enabling, any one method described in above-mentioned first aspect realized when described instruction is executed by processor.
5th aspect, provides a kind of computer program product comprising instruction, when run on a computer, so that
Either described in the above-mentioned first aspect of computer execution the step of method.
Technical solution provided by the embodiments of the present application has the benefit that
In this application, the audio frequency characteristics for obtaining every song in number of songs to be recommended, according to every in number of songs
The audio frequency characteristics of song, determine recommendation list.Wherein, audio frequency characteristics are used to indicate the tone, pitch and tone color of audio, thus
It is found that the embodiment of the present application is to recommend song according to the tone of song itself, pitch and tone color, so may be implemented to
The song of high quality is recommended at family, and be can be avoided and influenced during recommending song by the factor and individual subjective factor of the user, or
Person is influenced by extraneous factor, so that the song recommended meets the needs of user is to high quality song.In addition, in the application
Configured with song recommendations model, for quickly determining the sound quality score of every song, to improve the efficiency of determining recommendation list.
Detailed description of the invention
In order to more clearly explain the technical solutions in the embodiments of the present application, make required in being described below to embodiment
Attached drawing is briefly described, it should be apparent that, the drawings in the following description are only some examples of the present application, for
For those of ordinary skill in the art, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing.
Fig. 1 is a kind of song recommendations system schematic provided by the embodiments of the present application；
Fig. 2 is a kind of song recommendations method flow diagram provided by the embodiments of the present application；
Fig. 3 is a kind of training method flow chart of trained song recommendations model provided by the embodiments of the present application；
Fig. 4 is a kind of song recommendations apparatus structure schematic diagram provided by the embodiments of the present application；
Fig. 5 is another song recommendations apparatus structure schematic diagram provided by the embodiments of the present application；
Fig. 6 is another song recommendations apparatus structure schematic diagram provided by the embodiments of the present application.
Specific embodiment
To keep the purposes, technical schemes and advantages of the application clearer, below in conjunction with attached drawing to the application embodiment party
Formula is described in further detail.
Fig. 1 is a kind of song recommendations system schematic provided by the embodiments of the present application, as shown in Figure 1, the song recommendations system
System 100 includes server 101, user terminal 102 and back-stage management terminal 103.Lead between user terminal 102 and server 101
It crosses wirelessly or non-wirelessly mode to connect to be communicated, also by wirelessly or non-wirelessly between back-stage management terminal 103 and server 101
Mode is connected to be communicated.
Wherein, server 101 is used to send recommendation list to user terminal 102, and user terminal 102 shows recommendation column
Table, includes number of songs in recommendation list, recommends song to realize to user.Back-stage management terminal 103 is used for server 101
It is managed.
In addition, user terminal 102 and back-stage management terminal 103 all can be mobile phone, tablet computer or desktop computers etc.
Equipment.Fig. 1 is only mobile phone with user terminal 102, back-stage management terminal 103 is that desktop computer is illustrated, not
Constitute the specific restriction to the embodiment of the present application.
In addition, server 101 can be connect with multiple user terminals 102, to realize to each of multiple user terminals
User terminal recommends song.Only with 1 user terminal 102 for example, not constituted the tool to the embodiment of the present application in Fig. 1
Body limits.
Fig. 2 is a kind of song recommendations method flow diagram provided by the embodiments of the present application, is applied to server shown in FIG. 1.
As shown in Fig. 2, this method comprises the following steps:
Step 201: obtaining number of songs to be recommended.
In one possible implementation, when user terminal detects song recommendations instruction, user terminal is to service
Device sends song recommendations request.The mark of the user terminal is carried in song recommendations request.When server receives the song
When recommendation request, number of songs are obtained from database, it is then true from number of songs according to following step 202 and step 203
Determine recommendation list.The recommendation list is sent to the user terminal, is shown by user terminal after determining recommendation list by server
Show the recommendation list.
Step 202: determine the audio frequency characteristics of every song in number of songs, audio frequency characteristics be used to indicate audio tone,
Pitch and tone color.
For any section audio, which usually all has three tone, pitch and tone color attributes.Wherein, tone is used for
Indicate the frequency of corresponding audio, pitch is used to indicate the intensity of corresponding audio, and tone color is used to indicate the humorous of corresponding audio
Wave component.Also, the audio of high quality has common feature usually on these three attributes.For example, each frequency point is corresponding in audio
Amplitude distribution it is uniform, harmonic components are abundant etc. in audio.Therefore, in the embodiment of the present application, recommend to realize to user
The song of high quality can recommend song to user according to the audio frequency characteristics of song every in number of songs to be recommended.
Wherein, the implementation for obtaining the audio frequency characteristics of every song in number of songs to be recommended can be with are as follows: for this
Any song in number of songs can be obtained the tone of the song by music pitch extraction technology respectively, extract skill by pitch
Art obtains the pitch of the song, and the tone color of the song is obtained by tone color extractive technique.It is extracted about music pitch extraction technology, pitch
Technology and tone color extractive technique can refer to the relevant technologies, just no longer repeat one by one herein.
Step 203: according to the audio frequency characteristics of every song and song recommendations model, determine the sound quality score of every song,
According to the sound quality score of song every in number of songs, recommendation list is determined.
In order to improve the efficiency of determining recommendation list, the neural network model for song recommendations is stored in server,
It is known as song recommendations model in the embodiment of the present application.Therefore, in one possible implementation, step 202 specifically can be with
Are as follows: according to the audio frequency characteristics and song recommendations model of song every in number of songs, determine the sound quality score of every song；According to
The sound quality score of every song, determines recommendation list in number of songs.
Wherein, according to the sound quality score of song every in number of songs, determine that the implementation of recommendation list can be with are as follows: from
Sound quality score is selected to be greater than the song with reference to score in number of songs, and using the song selected as the song in recommendation list.
Optionally, number of songs can also be ranked up according to the sound quality score of song every in number of songs, is then tied from sequence
At least one biggish song of sound quality score is selected in fruit, and using the song selected as the song in recommendation list.
In addition, in the embodiment of the present application, in order to enable the song recommended not only meets user to the need of high quality song
It asks, can also meet the needs of user itself is to music type, server is pushed away through the above steps 201 to step 203 determination
Data can also be listened to the history for obtaining user before recommending list, then listen song data from song data according to the history of user
Number of songs are filtered out in library, in order to determine recommendation list according to the number of songs filtered out.
Specifically, server is got after history listens to data, is listened to data according to history and is determined at least one music
Type, at least one music type can characterize the music type of user preferences.Then it can be screened from song database
Meet the number of songs of at least one music type, in order to according to the number of songs that filter out through the above steps 201 to
Step 203 determines recommendation list.
It that is to say, in the embodiment of the present application, server not only can only determine recommendation according to the audio frequency characteristics of song
List.Server can be combined with the audio frequency characteristics for listening to data and song of user to determine recommendation list.Certainly, server
The audio frequency characteristics of other data and song be can be combined with to determine recommendation list, for example, server can be combined with network matchmaker
Body recommending data and the audio frequency characteristics of song determine recommendation list.Not only meet user couple with the song for realizing that server is recommended
The demand of high quality song also meets the needs of user itself or meets the needs of current music platform, further improves and push away
Recommend the flexibility of song.
Above-mentioned song recommendations model is obtained by the neural network model training of initialization.Training is initialized at this
Neural network model obtain the detailed process of song recommendations model explanation be explained in detail.
Fig. 3 is a kind of training method flow chart of song recommendations model provided by the embodiments of the present application.The song recommendations mould
The training method of type can be applied in server shown in FIG. 1, naturally it is also possible to be realized by other servers or terminal, so
Obtained song recommendations model is stored into server afterwards.Specifically, as shown in figure 3, this method comprises the following steps:
Step 301: obtain multiple trained audios and with the one-to-one sound quality score of multiple trained audios and multiple test
Demonstrate,prove audio and with multiple one-to-one sound quality scores of verifying audio.
In one possible implementation, step 301 is specifically as follows: for the first audio, obtaining and is directed to the first sound
The score of frequency marking note, obtains the sound quality score of the first audio；Alternatively, being listened to data according to the history of the first audio, is determined
The sound quality score of one audio.Wherein, the first audio is any audio in multiple trained audios and multiple verifying audios.
It that is to say, for the first audio, the sound quality score of first audio can be obtained by way of manually marking.Also
Data can be listened to and determined the sound quality score of the first audio by the history of the first audio.Wherein, since manually mark can be with
It is completed by special musical expert, therefore, the sound quality score of first audio is obtained by way of manually marking, it can be with
So that height of the sound quality score more representative of the quality of the first audio.
But the mode manually marked needs to expend more human resources.So server can also be directly according to first
The history of audio is listened to data, determines the sound quality score of the first audio, to improve to training sample and verifying sample setting mark
The efficiency of label.
Wherein, since the history collection number of audio can characterize the quality height of the audio to a certain extent,
Data are listened to according to the history of the first audio, determine that the implementation of the sound quality score of the first audio is specifically as follows: according to
The history collection number and history broadcasting time of first audio, determine the sound quality score of the first audio.For example, server can incite somebody to action
The history collection number and history broadcasting time of first audio are divided by, and obtained quotient can be used as the sound quality point of first audio
Number.
Certainly, in the embodiment of the present application, server can also determine the sound quality of the first audio by other implementations
Score, as long as the sound quality score of first audio can characterize the height of the quality of the audio.For example, server can incite somebody to action
The number thumbed up and history broadcasting time of first audio are divided by, and obtained quotient also can be used as the sound quality point of first audio
Number.
In addition, in addition to training sample and verifying sample, usually also needing test specimens during training neural network
This, to be tested by test sample model after training after training.Therefore, in step 301, may be used also
With obtain multiple testing audios and with the one-to-one sound quality score of multiple testing audios.
Specifically, in one possible implementation, server can first obtain magnanimity audio sample, then according to one
Magnanimity audio sample is divided into training sample, verifying sample and test sample by fixed ratio, and passes through any of the above-described realization side
Formula determines the sound quality score of each sample.For example, magnanimity audio sample can be divided into instruction according to the ratio of 8:1:1 by server
Practice sample, verifying sample and test sample.
Optionally, not smart enough so as to cause the model trained in order to avoid the sample of certain sound quality scores is excessively concentrated
Really.Server after the sound quality score for determining training sample, verifying sample and test sample and each sample, for
Any sort sample counts the distribution histogram of the sound quality score of such sample, rejects the too high or too low sample of some scores, with
The distribution of the sound quality score of the sample after rejecting is set to meet normal distribution.
In addition, server is in the sound quality score determined training sample, verify sample and test sample and each sample
It later, can be directly using audio each in training sample as a trained audio, using each audio in verifying sample as one
A verifying audio, using audio each in test sample as a testing audio,
Certainly, server can also intercept each sample, so that a length of when the audio of each sample refer to duration.
It for example can be 1 minute.Using the sample after interception as corresponding trained audio, verifying audio and testing audio.
Step 302: determining in multiple trained audios every in the audio frequency characteristics of each trained audio and multiple verifying audios
The audio frequency characteristics of a verifying audio.
Wherein, the implementation of step 302 can be with reference to the step 202 in Fig. 2 embodiment, and details are not described herein.
Optionally, it if server is after training, needs to survey model after training by test sample
Examination, at this point, in step 302, the audio frequency characteristics of each testing audio in multiple testing audios can also be obtained
Step 303: according to each trained sound in the one-to-one sound quality score of multiple trained audios, multiple trained audios
The audio frequency characteristics of frequency, and with each verifying audio in the one-to-one sound quality score of multiple verifying audios, multiple verifying audios
Audio frequency characteristics, the neural network model of initialization is trained, song recommendations model is obtained.
In one possible implementation, step 303 is specifically as follows: corresponding according to multiple trained audios
Sound quality score, the audio frequency characteristics of each trained audio are trained the neural network model of initialization in multiple trained audios
During, obtain neural network model every time after training；According to multiple one-to-one sound quality scores of verifying audio
With each audio frequency characteristics for verifying audio in multiple verifying audios and neural network model after training, loss letter is determined
Several values.If the value of loss function reaches minimum value, neural network model after training is determined as song recommendations mould
Type.If the value of loss function does not reach minimum value, execution basis and multiple trained one-to-one sound quality of audio are returned
The audio frequency characteristics of each trained audio are trained neural network model after training in score and multiple trained audios
Step.
Wherein, loss function is used to indicate the gap between the predicted value of neural network model and true value, therefore, can be with
Determine whether to terminate current training by the value of loss function.Furthermore it is possible to preset valuation functions, pass through valuation functions
It can determine whether the value of current loss function reaches minimum value.In the embodiment of the present application, loss function can be mean square error
Mean absolute error function can be used in difference function, valuation functions.
Optionally, it if server is after training, needs to survey model after training by test sample
Examination, at this point, server is before being determined as song recommendations model for neural network model after training, can also according to it is more
The audio frequency characteristics of each testing audio in the one-to-one sound quality score of a testing audio, multiple testing audios, to after training
Neural network model tested；Test result is sent to back-stage management terminal；If receiving back-stage management terminal hair
The termination training instruction sent then executes the operation that neural network model after training is determined as to song recommendations model.
Correspondingly, if receiving the continuation training instruction of back-stage management terminal transmission, the mind of adjustable initialization
Through the relevant parameter in network model.And continued according to multiple trained audios and multiple verifying audios to the nerve after adjusting parameter
Network model is trained.It that is to say, return to step 303.
The neural network model of above-mentioned initialization can be one-dimensional convolutional neural networks model, or two-dimensional volume
Product neural network model, naturally it is also possible to be other neural network models, be not specifically limited herein.When the nerve net of initialization
It can also be initialization when network model is one-dimensional convolutional neural networks model or two-dimensional convolutional neural networks model
Neural network model allocation activation function, for example, activation primitive can be S sigmoid growth curve (sigmoid) function.
In the embodiment of the present application, the audio frequency characteristics for obtaining every song in number of songs to be recommended, according to more first songs
The audio frequency characteristics of every song in song determine the sound quality score of every song；According to the sound quality of song every in number of songs point
Number, determines recommendation list.Wherein, audio frequency characteristics are used to indicate the tone, pitch and tone color of audio, it follows that the application is real
Applying example is to recommend song according to the tone of song itself, pitch and tone color, so may be implemented to recommend high quality to user
Song, and can be avoided and influenced during recommending song by the factor and individual subjective factor of the user, or by extraneous factor
Influence so that recommend song meet the needs of user is to high quality song.In addition, being configured in the embodiment of the present application
Song recommendations model, for quickly determining the sound quality score of every song, to improve the efficiency of determining recommendation list.
Fig. 4 is a kind of song recommendations device 400 provided by the embodiments of the present application, which includes song recommendations model, song
Bent recommended models are obtained by the neural network model training of initialization, which includes:
First obtains module 401, for obtaining number of songs to be recommended；
First determining module 402, for determining that the audio frequency characteristics of every song in number of songs, audio frequency characteristics are used to indicate
Tone, pitch and the tone color of audio；
Second determining module 403 determines every song for the audio frequency characteristics and song recommendations model according to every song
Sound quality score recommendation list is determined according to the sound quality score of song every in number of songs.
Optionally, as shown in figure 5, device 400 further include:
Second obtains module 404, for obtaining multiple trained audios and dividing with multiple trained one-to-one sound quality of audio
Several and multiple verifying audios and with multiple one-to-one sound quality scores of verifying audio；
Third determining module 405, for determining in multiple trained audios audio frequency characteristics of each trained audio and multiple
Verify the audio frequency characteristics of each verifying audio in audio；
Training module 406, for according to it is every in the one-to-one sound quality score of multiple trained audios, multiple trained audios
The audio frequency characteristics of a trained audio, and with it is each in the one-to-one sound quality score of multiple verifying audios, multiple verifying audios
The audio frequency characteristics for verifying audio, are trained the neural network model of initialization, obtain song recommendations model.
Optionally, training module 406 are specifically used for:
According to the sound with each trained audio in the one-to-one sound quality score of multiple trained audios, multiple trained audios
During frequency feature is trained the neural network model of initialization, neural network model every time after training is obtained；
According to the sound with each verifying audio in the one-to-one sound quality score of multiple verifying audios and multiple verifying audios
Frequency feature and neural network model after training, determine the value of loss function, loss function is used to indicate neural network mould
Gap between the predicted value and true value of type；
If the value of loss function reaches minimum value, neural network model after training is determined as song recommendations mould
Type；
If the value of loss function does not reach minimum value, execution is returned according to one-to-one with multiple trained audios
The audio frequency characteristics of each trained audio instruct neural network model after training in sound quality score and multiple trained audios
Experienced step.
Optionally, training module 406, also particularly useful for:
According to the audio with each testing audio in the one-to-one sound quality score of multiple testing audios, multiple testing audios
Feature tests neural network model after training；
Test result is sent to background management device；
If receiving the termination training instruction of background management device transmission, execute neural network mould after training
Type is determined as the operation of song recommendations model.
Optionally, second module 404 is obtained, is specifically used for:
For the first audio, the score for the first audio mark is obtained, the sound quality score of the first audio is obtained；Alternatively,
Data are listened to according to the history of the first audio, determine the sound quality score of the first audio；
Wherein, the first audio is any audio in multiple trained audios and multiple verifying audios.
Optionally, device 400 further include:
Third obtains module, and the history for obtaining user listens to data；
Screening module listens song data to filter out number of songs from song database for the history according to user.
In the embodiment of the present application, the audio frequency characteristics for obtaining every song in number of songs to be recommended, according to more first songs
The audio frequency characteristics of every song in song determine the sound quality score of every song；According to the sound quality of song every in number of songs point
Number, determines recommendation list.Wherein, audio frequency characteristics are used to indicate the tone, pitch and tone color of audio, it follows that the application is real
Applying example is to recommend song according to the tone of song itself, pitch and tone color, so may be implemented to recommend high quality to user
Song, and can be avoided and influenced during recommending song by the factor and individual subjective factor of the user, or by extraneous factor
Influence so that recommend song meet the needs of user is to high quality song.In addition, being configured in the embodiment of the present application
Song recommendations model, for quickly determining the sound quality score of every song, to improve the efficiency of determining recommendation list.
It should be understood that song recommendations device provided by the above embodiment is when carrying out song recommendations, only with above-mentioned each
The division progress of functional module can according to need and for example, in practical application by above-mentioned function distribution by different function
Energy module is completed, i.e., the internal structure of equipment is divided into different functional modules, to complete whole described above or portion
Divide function.In addition, song recommendations device provided by the above embodiment and song recommendations embodiment of the method belong to same design, have
Body realizes that process is detailed in embodiment of the method, and which is not described herein again.
Fig. 6 is a kind of structural representation of song recommendations device provided by the embodiments of the present application, and server shown in FIG. 1 can be with
It is realized by device shown in fig. 6.The server can be the server in background server cluster.Specifically:
Song recommendations device 600 includes 602 He of central processing unit (CPU) 601 including random access memory (RAM)
The system storage 604 of read-only memory (ROM) 603, and connection system storage 604 and central processing unit 601 be
System bus 605.Song recommendations device 600 further include help computer in each device between transmit information it is basic input/
Output system (I/O system) 606, and for the great Rong of storage program area 613, application program 614 and other program modules 615
Amount storage equipment 607.Upper central processing unit is referred to as processor.
Basic input/output 606 includes display 608 for showing information and inputs information for user
The input equipment 609 of such as mouse, keyboard etc.Wherein display 608 and input equipment 609 are all by being connected to system bus
605 input and output controller 610 is connected to central processing unit 601.Basic input/output 606 can also include defeated
Enter o controller 610 for receiving and handling from the defeated of multiple other equipment such as keyboard, mouse or electronic touch pen
Enter.Similarly, input and output controller 610 also provides output to display screen, printer or other kinds of output equipment.
Mass-memory unit 607 is connected by being connected to the bulk memory controller (not shown) of system bus 605
To central processing unit 601.Mass-memory unit 607 and its associated computer-readable medium are song recommendations device
600 provide non-volatile memories.That is, mass-memory unit 607 may include such as hard disk or CD-ROM driving
The computer-readable medium (not shown) of device etc.
Without loss of generality, computer-readable medium may include computer storage media and communication media.Computer storage
Medium includes any of the information such as computer readable instructions, data structure, program module or other data for storage
The volatile and non-volatile of method or technique realization, removable and irremovable medium.Computer storage medium include RAM,
ROM, EPROM, EEPROM, flash memory or other solid-state storages its technologies, CD-ROM, DVD or other optical storages, cassette, magnetic
Band, disk storage or other magnetic storage devices.Certainly, skilled person will appreciate that computer storage medium is not limited to
It states several.Above-mentioned system storage 604 and mass-memory unit 607 may be collectively referred to as memory.
According to the various embodiments of the application, song recommendations device 600 can also be connected to the network by internet etc.
Remote computer operation on to network.Namely song recommendations device 600 can be by the network that is connected on system bus 605
Interface unit 611 is connected to network 612, in other words, Network Interface Unit 611 can be used also to be connected to other kinds of net
Network or remote computer system (not shown).
Above-mentioned memory further includes one, and perhaps more than one program one or more than one program are stored in storage
In device, it is configured to be executed by CPU.The one or more programs include for carrying out song provided by the embodiments of the present application
The instruction of bent recommended method.
The embodiment of the present application also provides a kind of non-transitorycomputer readable storage mediums, when in the storage medium
When instruction is executed by the processor of song recommendations device 600, so that song recommendations device 600 is able to carry out above-described embodiment offer
Song recommendations method.
The embodiment of the present application also provides a kind of computer program products comprising instruction, when the instruction is transported on computers
When row, so that computer executes song recommendations method provided by the above embodiment.
Those of ordinary skill in the art will appreciate that realizing that all or part of the steps of above-described embodiment can pass through hardware
It completes, relevant hardware can also be instructed to complete by program, the program can store in a kind of computer-readable
In storage medium, storage medium mentioned above can be read-only memory, disk or CD etc..
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (14)
1. a kind of song recommendations method, which is characterized in that the method includes song recommendations model, the song recommendations model is
It is obtained by the neural network model training of initialization, which comprises
Obtain number of songs to be recommended；
Determine that the audio frequency characteristics of every song in the number of songs, the audio frequency characteristics are used to indicate the tone of audio, pitch
And tone color；
According to the audio frequency characteristics of every song and the song recommendations model, the sound quality score of every song is determined；
According to the sound quality score of song every in the number of songs, recommendation list is determined.
2. the method as described in claim 1, which is characterized in that the method also includes:
Obtain multiple trained audios and with the multiple trained one-to-one sound quality score of audio and multiple verifying audios and
With the multiple one-to-one sound quality score of verifying audio；
It determines in the multiple trained audio and is each tested in the audio frequency characteristics of each trained audio and the multiple verifying audio
Demonstrate,prove the audio frequency characteristics of audio；
According to each trained audio in the one-to-one sound quality score of the multiple trained audio, the multiple trained audio
Audio frequency characteristics, and with each verifying in the one-to-one sound quality score of the multiple verifying audio, the multiple verifying audio
The audio frequency characteristics of audio are trained the neural network model of the initialization, obtain the song recommendations model.
3. method according to claim 2, which is characterized in that the basis and the multiple trained one-to-one sound of audio
The audio frequency characteristics of each trained audio in matter score, the multiple trained audio, and it is a pair of with the multiple verifying audio one
Sound quality score, the multiple audio frequency characteristics for verifying each verifying audio in audio answered, to the neural network of the initialization
Model is trained, and obtains the song recommendations model, comprising:
According to each trained audio in the one-to-one sound quality score of the multiple trained audio, the multiple trained audio
Audio frequency characteristics the neural network model of the initialization is trained during, obtain nerve net after training every time
Road model；
According to each verifying audio in the one-to-one sound quality score of the multiple verifying audio and the multiple verifying audio
Audio frequency characteristics and the neural network model after training, determine the value of loss function, the loss function is for referring to
Show the gap between the predicted value of neural network model and true value；
If the value of the loss function reaches minimum value, the neural network model after training is determined as the song
Bent recommended models；
If the value of the loss function does not reach minimum value, execution is returned according to a pair of with the multiple trained audio one
The audio frequency characteristics of each trained audio are to the nerve net after training in the sound quality score answered and the multiple trained audio
The step of network model is trained.
4. method as claimed in claim 3, which is characterized in that described to be determined as the neural network model after training
Before the song recommendations model, further includes:
According to the audio with each testing audio in the one-to-one sound quality score of multiple testing audios, the multiple testing audio
Feature tests the neural network model after training；
Test result is sent to back-stage management terminal；
If receiving the termination training instruction that the back-stage management terminal is sent, execute the nerve net after training
Road model is determined as the operation of the song recommendations model.
5. method according to claim 2, which is characterized in that it is described obtain multiple trained audios and with the multiple trained sound
The one-to-one sound quality score of frequency and multiple verifying audios and with the multiple one-to-one sound quality score of verifying audio,
Include:
For the first audio, the score for first audio mark is obtained, the sound quality score of first audio is obtained；Or
Person,
Data are listened to according to the history of first audio, determine the sound quality score of first audio；
Wherein, first audio is any audio in the multiple trained audio and the multiple verifying audio.
6. method as claimed in claim 1 to 5, which is characterized in that before acquisition number of songs to be recommended, also
Include:
The history for obtaining user listens to data；
Song data are listened to filter out the number of songs from song database according to the history of the user.
7. a kind of song recommendations device, which is characterized in that described device includes song recommendations model, and the song recommendations model is
It is obtained by the neural network model training of initialization, described device includes:
First obtains module, for obtaining number of songs to be recommended；
First determining module, for determining the audio frequency characteristics of every song in the number of songs, the audio frequency characteristics are for referring to
Show the tone, pitch and tone color of audio；
Second determining module determines every song for the audio frequency characteristics and the song recommendations model according to every song
Sound quality score determines recommendation list according to the sound quality score of song every in the number of songs.
8. device as claimed in claim 7, which is characterized in that described device further include:
Second obtain module, for obtain multiple trained audios and with the multiple trained one-to-one sound quality score of audio,
And multiple verifying audios and with the multiple one-to-one sound quality score of verifying audio；
Third determining module, for determining in the multiple trained audio audio frequency characteristics of each trained audio and described more
The audio frequency characteristics of each verifying audio in a verifying audio；
Training module, for according to in the one-to-one sound quality score of the multiple trained audio, the multiple trained audio
The audio frequency characteristics of each trained audio, and with the one-to-one sound quality score of the multiple verifying audio, the multiple verifying
The audio frequency characteristics of each verifying audio, are trained the neural network model of the initialization, obtain the song in audio
Recommended models.
9. device as claimed in claim 8, which is characterized in that the training module is specifically used for:
According to each trained audio in the one-to-one sound quality score of the multiple trained audio, the multiple trained audio
Audio frequency characteristics the neural network model of the initialization is trained during, obtain nerve net after training every time
Road model；
According to each verifying audio in the one-to-one sound quality score of the multiple verifying audio and the multiple verifying audio
Audio frequency characteristics and the neural network model after training, determine the value of loss function, the loss function is for referring to
Show the gap between the predicted value of neural network model and true value；
If the value of the loss function reaches minimum value, the neural network model after training is determined as the song
Bent recommended models；
If the value of the loss function does not reach minimum value, execution is returned according to a pair of with the multiple trained audio one
The audio frequency characteristics of each trained audio are to the nerve net after training in the sound quality score answered and the multiple trained audio
The step of network model is trained.
10. device as claimed in claim 9, which is characterized in that the training module, also particularly useful for:
According to the audio with each testing audio in the one-to-one sound quality score of multiple testing audios, the multiple testing audio
Feature tests the neural network model after training；
Test result is sent to background management device；
If receiving the termination training instruction that the background management device is sent, execute the nerve net after training
Road model is determined as the operation of the song recommendations model.
11. device as claimed in claim 8, which is characterized in that described second obtains module, is specifically used for:
For the first audio, the score for first audio mark is obtained, the sound quality score of first audio is obtained；Or
Person,
Data are listened to according to the history of first audio, determine the sound quality score of first audio；
Wherein, first audio is any audio in the multiple trained audio and the multiple verifying audio.
12. device a method according to any one of claims 7 to 11, which is characterized in that described device further include:
Third obtains module, and the history for obtaining user listens to data；
Screening module, for listening song data to filter out the number of songs from song database according to the history of the user.
13. a kind of song recommendations device, which is characterized in that described device includes:
Processor；
Memory for storage processor executable instruction；
Wherein, the processor is configured to perform claim require 1 to method any one of as claimed in claim 6 the step of.
14. a kind of computer readable storage medium, it is stored with instruction on the computer readable storage medium, described instruction is located
Manage the step of claim 1 to any one method as claimed in claim 6 is realized when device executes.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910797899.7A CN110516104A (en) | 2019-08-27 | 2019-08-27 | Song recommendations method, apparatus and computer storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910797899.7A CN110516104A (en) | 2019-08-27 | 2019-08-27 | Song recommendations method, apparatus and computer storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110516104A true CN110516104A (en) | 2019-11-29 |
Family
ID=68628266
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910797899.7A Pending CN110516104A (en) | 2019-08-27 | 2019-08-27 | Song recommendations method, apparatus and computer storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110516104A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110913240A (en) * | 2019-12-02 | 2020-03-24 | 广州酷狗计算机科技有限公司 | Video interception method, device, server and computer readable storage medium |
| CN111488485A (en) * | 2020-04-16 | 2020-08-04 | 北京雷石天地电子技术有限公司 | Music recommendation method based on convolutional neural network, storage medium and electronic device |
| CN113032616A (en) * | 2021-03-19 | 2021-06-25 | 腾讯音乐娱乐科技（深圳）有限公司 | Audio recommendation method and device, computer equipment and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110184948A1 (en) * | 2010-01-22 | 2011-07-28 | National Cheng Kung University | Music recommendation method and computer readable recording medium storing computer program performing the method |
| CN106528813A (en) * | 2016-11-18 | 2017-03-22 | 腾讯科技（深圳）有限公司 | Multimedia recommendation method and apparatus |
| CN108111908A (en) * | 2017-12-25 | 2018-06-01 | 深圳Tcl新技术有限公司 | Audio quality determines method, equipment and computer readable storage medium |
| CN108735192A (en) * | 2018-04-09 | 2018-11-02 | 华南理工大学 | A kind of piano performance assessment of acoustics system and method for combination style of song |
| CN108766451A (en) * | 2018-05-31 | 2018-11-06 | 腾讯音乐娱乐科技（深圳）有限公司 | A kind of audio file processing method, device and storage medium |
-
2019
- 2019-08-27 CN CN201910797899.7A patent/CN110516104A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110184948A1 (en) * | 2010-01-22 | 2011-07-28 | National Cheng Kung University | Music recommendation method and computer readable recording medium storing computer program performing the method |
| CN106528813A (en) * | 2016-11-18 | 2017-03-22 | 腾讯科技（深圳）有限公司 | Multimedia recommendation method and apparatus |
| CN108111908A (en) * | 2017-12-25 | 2018-06-01 | 深圳Tcl新技术有限公司 | Audio quality determines method, equipment and computer readable storage medium |
| CN108735192A (en) * | 2018-04-09 | 2018-11-02 | 华南理工大学 | A kind of piano performance assessment of acoustics system and method for combination style of song |
| CN108766451A (en) * | 2018-05-31 | 2018-11-06 | 腾讯音乐娱乐科技（深圳）有限公司 | A kind of audio file processing method, device and storage medium |
Non-Patent Citations (2)
| Title |
|---|
| 佟雪娜: "《音乐产业运营与管理》", 30 October 2016 * |
| 李欣雯: ""基于双神经网络的钢琴琴键音质评价方法研究"", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110913240A (en) * | 2019-12-02 | 2020-03-24 | 广州酷狗计算机科技有限公司 | Video interception method, device, server and computer readable storage medium |
| CN111488485A (en) * | 2020-04-16 | 2020-08-04 | 北京雷石天地电子技术有限公司 | Music recommendation method based on convolutional neural network, storage medium and electronic device |
| CN111488485B (en) * | 2020-04-16 | 2023-11-17 | 北京雷石天地电子技术有限公司 | Music recommendation method based on convolutional neural network, storage medium and electronic device |
| CN113032616A (en) * | 2021-03-19 | 2021-06-25 | 腾讯音乐娱乐科技（深圳）有限公司 | Audio recommendation method and device, computer equipment and storage medium |
| CN113032616B (en) * | 2021-03-19 | 2024-02-20 | 腾讯音乐娱乐科技（深圳）有限公司 | Audio recommendation method, device, computer equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104090912B (en) | Information-pushing method and device | |
| CN110516104A (en) | Song recommendations method, apparatus and computer storage medium | |
| US11069361B2 (en) | System and method for validating natural language content using crowdsourced validation jobs | |
| CN111901626B (en) | Background audio determining method, video editing method, device and computer equipment | |
| US20130238592A1 (en) | Application store tastemaker recommendations | |
| Contreras et al. | A flexible modelling of galaxy assembly bias | |
| CN103455538B (en) | Information processing unit, information processing method and program | |
| CN106991163A (en) | A kind of song recommendations method based on singer's sound speciality | |
| CN103400593B (en) | A kind of method of audition audio frequency and device | |
| CN106294794A (en) | A kind of content recommendation method and device | |
| CN103440330A (en) | Music program information acquisition method and equipment | |
| CN108764332A (en) | A kind of Channel Quality analysis method, computing device and storage medium | |
| CN108280179B (en) | Method and system, terminal and the computer readable storage medium of audio advertisement detection | |
| TWI791176B (en) | Method, system, device and computer program carrier for automatically identifying effective data collection modules | |
| Steppat et al. | What news users perceive as ‘alternative media’varies between countries: How media fragmentation and polarization matter | |
| CN102075366A (en) | Method and equipment for processing data in communication network | |
| CN108764114A (en) | A kind of signal recognition method and its equipment, storage medium, terminal | |
| CN112598245A (en) | Government affair service experience improving method and device without user feedback | |
| CN108629047A (en) | A kind of song list generation method and terminal device | |
| CN107810638A (en) | By the transmission for skipping redundancy fragment optimization order content | |
| CN107562847A (en) | Information processing method and related product | |
| CN108446370A (en) | Voice data statistical method and system | |
| CN111192170A (en) | Topic pushing method, device, equipment and computer readable storage medium | |
| CN113077815A (en) | Audio evaluation method and component | |
| US10681402B2 (en) | Providing relevant and authentic channel content to users based on user persona and interest |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191129 |