WO2019218462A1 - 一种歌曲清单生成方法、装置、终端设备及介质 - Google Patents

一种歌曲清单生成方法、装置、终端设备及介质 Download PDF

Info

Publication number
WO2019218462A1
WO2019218462A1 PCT/CN2018/097112 CN2018097112W WO2019218462A1 WO 2019218462 A1 WO2019218462 A1 WO 2019218462A1 CN 2018097112 W CN2018097112 W CN 2018097112W WO 2019218462 A1 WO2019218462 A1 WO 2019218462A1
Authority
WO
WIPO (PCT)
Prior art keywords
song
list
favorite
songs
user
Prior art date
Application number
PCT/CN2018/097112
Other languages
English (en)
French (fr)
Inventor
张辉
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2019218462A1 publication Critical patent/WO2019218462A1/zh

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
    • G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the present application belongs to the field of data processing technologies, and in particular, to a song list generation method, device, terminal device and medium.
  • the existing music software is simple to generate a list of recommended songs according to the song type label manually selected by the user. For example, if the user selects a Mandarin song/Cantonese/English song, the song list of the corresponding language is recommended, but actually the song library is selected. Huge, the number of songs corresponding to each type of tag is extremely large, that is, the number of recommended songs corresponding to each user is also very large. Therefore, in the prior art, the accuracy of generating a song list for a user is low, and it is impossible to implement a personalized song list generation for the user.
  • the embodiment of the present application provides a song list generation method and a terminal device, so as to solve the problem that the accuracy of generating a user-recommended song list in the prior art is low, and the user-specific recommended song list cannot be generated. problem.
  • a first aspect of the embodiments of the present application provides a method for generating a song list, including:
  • Each of said user are split into a list of favorite list of favorite type number M n, each corresponding to one kind of a list of favorite songs type type, M n is a positive integer, n ⁇ [1, N];
  • the song feature score matrix is used to record a song feature score of a feature of the song to be analyzed, and the feature to be analyzed is one or more of a time domain feature, a frequency domain feature, and a cepstrum feature of the song audio data.
  • the song feature score is a corresponding score obtained by analyzing the feature of the song to be analyzed, wherein P n is the number of songs included in the favorite type list, and H is the number of types of song features to be analyzed, P n and H is a positive integer;
  • a second aspect of the embodiment of the present application provides a song list generating apparatus, including:
  • a list extracting module configured to acquire a user favorite level corresponding to each song in the list of played songs of the user end, and extract N user favorite lists in the list of played songs, and each of the user favorite lists corresponds to one a user favorite level greater than a preset level, wherein N is a positive integer;
  • Listing difference module for each user a list of favorite were split number M n favorite genre list, each list corresponding to one kind of said favorite song genre type, M n is a positive integer, n ⁇ [1, N ];
  • a matrix dimension reduction module configured to perform dimension reduction on a first song feature score matrix of each of the favorite type lists corresponding to a dimension of P n ⁇ H, to obtain a dimension corresponding to each of the favorite type lists is 1 ⁇ H a second song feature score matrix, the song feature score matrix is used to record a song feature score of the song feature to be analyzed, and the song feature to be analyzed is a time domain feature, a frequency domain feature, and a cepstrum feature of the song audio data.
  • the song feature score is a corresponding score obtained by analyzing the feature of the song to be analyzed, wherein P n is the number of songs included in the favorite type list, and H is the feature of the song to be analyzed
  • P n is the number of songs included in the favorite type list
  • H is the feature of the song to be analyzed
  • the number of species, P n and H are positive integers;
  • a song screening module configured to acquire a number of song recommendation corresponding to each song type, and perform similarity matching on the third song feature score matrix of the song in the preset song library based on each of the second song feature score matrices And determining a matching successful song of the recommended number of songs corresponding to each of the favorite type lists, and obtaining a list of recommended songs.
  • a third aspect of the embodiments of the present application provides a terminal device, including a memory, a processor, and the computer storing computer readable instructions executable on the processor, where the processor executes the computer The following steps are implemented when reading the instruction:
  • Each of said user are split into a list of favorite list of favorite type number M n, each corresponding to one kind of a list of favorite songs type type, M n is a positive integer, n ⁇ [1, N];
  • the song feature score matrix is used to record a song feature score of a feature of the song to be analyzed, and the feature to be analyzed is one or more of a time domain feature, a frequency domain feature, and a cepstrum feature of the song audio data.
  • the song feature score is a corresponding score obtained by analyzing the feature of the song to be analyzed, wherein P n is the number of songs included in the favorite type list, and H is the number of types of song features to be analyzed, P n and H is a positive integer;
  • a fourth aspect of the embodiments of the present application provides a computer readable storage medium storing computer readable instructions, wherein the computer readable instructions are implemented by at least one processor The following steps:
  • Each of said user are split into a list of favorite list of favorite type number M n, each corresponding to one kind of a list of favorite songs type type, M n is a positive integer, n ⁇ [1, N];
  • the song feature score matrix is used to record a song feature score of a feature of the song to be analyzed, and the feature to be analyzed is one or more of a time domain feature, a frequency domain feature, and a cepstrum feature of the song audio data.
  • the song feature score is a corresponding score obtained by analyzing the feature of the song to be analyzed, wherein P n is the number of songs included in the favorite type list, and H is the number of types of song features to be analyzed, P n and H is a positive integer;
  • the song list corresponding to the user's favorite degree is obtained, and then the song type is divided into each song favorite list, and the song feature scores are synthesized separately for different types of songs, and the song is determined.
  • the song characteristics of each type of song are guaranteed to provide subsequent screening accuracy.
  • the recommended number of recommended songs are filtered from the song library based on the synthesized song feature score matrix, which ensures that the user can fully fit the characteristics of the user's daily listening songs when recommending the songs to the user.
  • the songs that the user likes to listen to are more recommended, and cross-recommended songs recommended by users of different types and different tastes, so as to reduce the user's boredom with the recommended songs while satisfying the user's favorite song recommendation.
  • the user's preference for the song is precisely positioned, and then based on the analysis and matching of the song features of the favorite song, to achieve an accurate match to the user's favorite song. Therefore, the embodiment of the present application implements personalized generation of the user's favorite recommended song list generation, which greatly improves the accuracy of generating the user recommended song list.
  • FIG. 1 is a schematic flowchart of an implementation process of a song list generation method according to Embodiment 1 of the present invention
  • FIG. 2 is a schematic flowchart of an implementation process of a song list generating method according to Embodiment 2 of the present invention
  • FIG. 3 is a schematic flowchart of an implementation process of a song list generating method according to Embodiment 3 of the present invention.
  • FIG. 4 is a schematic flowchart showing an implementation process of a song list generating method according to Embodiment 4 of the present invention.
  • FIG. 5 is a schematic flowchart of an implementation process of a song list generation method according to Embodiment 5 of the present invention.
  • FIG. 6 is a schematic structural diagram of a song list generating apparatus according to Embodiment 6 of the present invention.
  • FIG. 7 is a schematic diagram of a song list generating terminal device according to Embodiment 7 of the present invention.
  • FIG. 1 is a flowchart showing an implementation of a song list generation method according to Embodiment 1 of the present application, which is described in detail as follows:
  • the user terminal is a device terminal for playing a song by a user, such as a mobile computer.
  • the list of played songs refers to the list of songs played by the user on the user side, such as the list of songs played on the mobile phone.
  • an application may not have the right to read the recording records of other applications, so When combined with the actual application, the list of played songs may refer to a list of songs played in a single or multiple applications in the client.
  • the user's taste for the song may change with time. For example, when the user has not played for a long time, the user may lose interest in the song.
  • the list of played songs in the embodiment of the present application It can be a list of all the songs played by the user, or it can only be a list of songs played in a certain period of time, such as a list of songs played in the last year, which can be set by the technician according to the requirements.
  • the user's favorite level that is, the level corresponding to the user's preference for a song, which may be obtained by the user's own evaluation of the song, or may be obtained according to the user's operation of the song, such as according to the user's
  • the number of times the song is played, whether it has been commented, whether it has been collected, and whether it has been marked or not, is scored, and the user's favorite rating for a song is obtained.
  • the related description of the second embodiment of the present application refer to the related description of the second embodiment of the present application.
  • the user may play songs that he does not like, such as listening to some online music stations, may play songs that the user does not like, so if you directly analyze all the songs that are played, the recommended songs are analyzed. It may result in songs that some users don't like and push to the user. Therefore, in order to ensure the accuracy and effectiveness of the list of recommended songs that are finally obtained, the embodiment of the present application will filter according to the user's favorite level of the song after obtaining the list of played songs, and only the songs higher than the preset level are used as the basis for the analysis of the recommended songs.
  • the data is extracted separately from the songs of each user's favorite level in the basic data to obtain a user favorite list corresponding to the user's favorite level, thereby performing analysis of the recommended songs and generation of the list. For example, assuming that the first user favorite level and the second user favorite level are two user favorite levels greater than the preset level, at this time, all the first user favorite level songs are extracted as one user favorite list, and all the second users are The favorite level song is extracted as another user favorite list, and a total of two corresponding user favorite lists are obtained.
  • the specific level of the preset level can be set by the technician according to the actual situation.
  • M n is a positive integer, n ⁇ [1, N].
  • the degree of preference for these different types of songs may be similar. For example, if the user likes rock and jazz-style songs is similar, then if you directly recommend the songs to the user's favorite list. Analysis, it is difficult to achieve the unification of the user's favorite list style, so that the accuracy of the recommended songs obtained is lower. Therefore, after obtaining N user favorite lists, the embodiment of the present application continues to classify songs for each user favorite list, and splits different types of songs into different favorite type lists, assume that the user favorite list A is included. After 10 songs, the songs are classified to determine that there are 3 songs of the first type, 4 songs of the second type, and 3 songs of the third type.
  • the user favorite list A is split into separate A list of three favorite types including 3 first type songs, 4 second type songs, and 3 third type songs.
  • the method for classifying the songs includes, but is not limited to, adaptively classifying the songs by using a neural network, or classifying the songs based on the type labels set by the technician in advance for the songs, which may be set by the technician according to actual conditions.
  • a list of favorite genre of each dimension corresponds to a first P n ⁇ H characteristic song score matrix dimensionality reduction, to give each of the favorite genre list corresponding dimension to a second 1 ⁇ H characteristic song score matrix, songs
  • the feature score matrix is used to record the song feature score of the feature to be analyzed.
  • the feature to be analyzed is one or more of the time domain feature, the frequency domain feature and the cepstrum feature of the song audio data, and the song feature score is to be analyzed.
  • Corresponding scores obtained by analyzing song features wherein P n is the number of songs included in the favorite type list, H is the number of types of song features to be analyzed, and P n and H are positive integers.
  • the analysis matching of the recommended songs is performed according to the song feature scores of the songs, but considering that the number of songs included in the favorite type list may be greater than 1, at this time
  • a favorite type list corresponds to a plurality of different song feature scores, such as when a favorite type list contains 3 songs, the favorite type list corresponds to three different sets of song feature scores, and each song in the song library to be matched Only one set of song feature scores is corresponding, so it is not possible to directly use the favorite type list to match the recommended songs at this time.
  • the plurality of sets of song feature scores included in the favorite type list are dimension-reduced to realize the synthesis of the plurality of sets of song feature scores, so that Each list of favorite types only corresponds to a set of song feature scores to ensure a normal match of subsequent recommended songs.
  • the type and quantity of the songs to be analyzed that are specifically required to be used in the embodiment of the present application, and the corresponding scoring method, can be set by the technician, for example, the short-term energy and the short-term average of the song audio data can be directly
  • the zero time domain feature is used as the feature of the song to be analyzed in the embodiment of the present application, and is scored by using some common audio analysis models, and the deep learning algorithm can also be used to analyze and learn the song audio data, and adaptively determine some to be analyzed.
  • the song features are scored and can be set by the technician according to actual needs.
  • the method for reducing the dimensionality of the song feature score matrix of the favorite type list is not limited herein, including but not limited to, averaging the scores of the same feature of the song to be analyzed for each song, and the obtained H pieces.
  • the average score of the song feature scores is a set of song feature scores and the like corresponding to the favorite type list. For example, the scores of the short-term energies of all the songs in the favorite type list are averaged to obtain a short-term energy score corresponding to the favorite type list.
  • the song feature score in the embodiment of the present application may itself be a vector of 1 ⁇ R, where R is a positive integer greater than 1.
  • R is a positive integer greater than 1.
  • the song features may change with time, such as the rhythm of the song, so that a song feature is represented by only one fractional value. It may be difficult to more accurately reflect the actual characteristics of the feature of the song. Therefore, in order to improve the accurate analysis of the feature of the song, and to provide guarantee for subsequent accurate matching, in the embodiment of the present application, the entire song is divided into A plurality of song segments are separately analyzed for each of the segments, so that the song features correspond to a plurality of different score values, and the corresponding song feature score vectors are obtained.
  • each song can be uniformly divided into the same number of song segments in a fixed proportion of the total duration of the songs, for example, The songs are divided into 4 segments by 25%, 25%, 25%, and 25%. In this case, no matter what the actual duration of the song is, it will be divided into four song segments for analysis. It should be noted that for some song features that are not related to time changes, such as the total number of zero crossings of a song, etc., the song feature scores may not be in a vector manner, but are still directly represented by a specific score value.
  • S104 Acquire a number of song recommendation corresponding to each song type, and perform similarity matching on the third song feature score matrix of the song in the preset song library based on each second song feature score matrix to determine each favorite.
  • the type list corresponds to the recommended number of songs matching the successful songs, and a list of recommended songs is obtained.
  • the song library is matched based on the obtained song feature score matrix of the favorite type list to determine the required recommended song.
  • the number of the songs included in the preset song library is generally large.
  • the setting method of the recommended number of songs for each song type can be set by the technician, including but not limited to, if a fixed number is directly set for each type of song, or a total is set for the final recommended song list.
  • a fixed number, and the number of song recommendations is proportionally distributed according to the number of specific songs included in each favorite type list.
  • the songs when generating a list of recommended songs, are preferably sorted from top to bottom in order of highest priority to the user's favorite level corresponding to the songs to obtain a list of recommended songs to ensure user preference.
  • the exposure of high songs increases the effectiveness of the list of recommended songs.
  • the song list corresponding to the user's favorite degree is obtained by dividing the favorite level of the song played by the user, and then the song type is divided into each song favorite list, and the different types of songs are separately performed.
  • the song feature scores are synthesized to determine the song characteristics of each type of song, so as to provide guarantee for the subsequent song screening accuracy.
  • the recommended number of recommended songs are filtered from the song library based on the synthesized song feature score matrix, which ensures that the user can fully fit the characteristics of the user's daily listening songs when recommending the songs to the user.
  • the songs that the user likes to listen to are more recommended, and cross-recommended songs recommended by users of different types and different tastes, so as to reduce the user's boredom with the recommended songs while satisfying the user's favorite song recommendation.
  • the user's preference for the song is precisely positioned, and then based on the analysis and matching of the song features of the favorite song, to achieve an accurate match to the user's favorite song. Therefore, the embodiment of the present application implements personalized generation of the user's favorite recommended song list generation, which greatly improves the accuracy of generating the user recommended song list.
  • the second embodiment of the present application includes: S201, acquiring operation data of each song in the list of played songs by the user terminal, and based on the pre- The operation score table is set to score the operation data to obtain the user favorite score of each song in the list of played songs.
  • the operation data refers to data obtained by recording, for example, a user's operation, such as playing, commenting, collecting, and marking the song.
  • the operation score table is a data table obtained by the technician in advance analyzing the user's operation behavior and setting corresponding scores for each operation behavior. Since different operation behaviors can reflect whether the user likes the song and likes or hates the song, if the user has collected the song, it means that the song is very favorite, therefore, in order to achieve accurate quantification of the user's song,
  • the technician analyzes and quantifies the possible operational behaviors of the user in advance and sets corresponding score values, and then analyzes and evaluates the operation data of each song based on the obtained operation score table, and obtains corresponding correspondence for each song. Users love scores to achieve effective quantification of user songs. After the user's favorite score is obtained, the song is divided into user favorite levels according to the set favorite score threshold to determine the specific user favorite level of the first song.
  • the operation data records the number of times the user plays the song, the duration of the play, whether it has been commented, whether it has been collected, and whether the tag has been liked
  • the operation score table is as follows:
  • the method when performing song classification, includes: extracting song features to be analyzed from song audio data, and processing the obtained feature of the song to be analyzed by using a neural network to perform songs on the songs.
  • Adaptive classification when performing song classification, includes: extracting song features to be analyzed from song audio data, and processing the obtained feature of the song to be analyzed by using a neural network to perform songs on the songs.
  • the feature of the song to be analyzed is first extracted, and then the obtained feature of the song to be analyzed is processed based on the neural network, so as to adaptively classify the song, the type of the song obtained at this time is not simple.
  • the type of artificial division of "rock” or "jazz” is a type of song that is more suitable for the characteristics of the song features to be analyzed. Since the type of the division is closer to the characteristics of the song itself, the song can be realized. For precise classification, it provides the basis for the exact match of the recommended songs, ensuring the accuracy of the final recommended songs.
  • the time domain feature such as the short-term energy of the song audio data
  • the short-term average zero-crossing is used as the feature of the song to be analyzed, and the feature data is extracted
  • the feature data of the obtained time domain features are obtained at this time.
  • the songs are adaptively classified to obtain different song types corresponding to different time domain features. For example, after adaptive analysis, the songs are divided into short-time energy highs and short-time average zero-crosses as the first type of songs, which will be short-term energy. High and short-time average zero-crossing as the second type of song, the short-time energy is low and the short-term average is more than zero as the third type of song, and the short-term energy is low and the short-term average is zero.
  • Type songs to divide the songs in the user's favorite list into four different types and split the list of favorite types.
  • the method before performing the dimensionality reduction synthesis on the song feature score matrix of the favorite type list, the method further includes:
  • S302 performing Fourier transform on the song audio data of the song in the user favorite list and performing dimensionality reduction processing to obtain processed song audio data.
  • the main component analysis algorithm can be used to complete the dimensionality reduction processing of the song audio data.
  • each hidden layer of the network represents the identification of the input data in a certain feature space. These features are different from the manually selected features.
  • the intermediate layer features are obtained from the training data, and are implicit.
  • the softmax classifier is used for classification, so that the processing of the song audio data by using the deep belief network can better reflect the essence of the song audio data, and the characteristics obtained by the network learning have better classification and prediction effects, and therefore, in the present application
  • the deep belief network by using the deep belief network to adaptively perform feature extraction evaluation and classification on the song audio data, the effectiveness of classifying the songs is greatly improved, and the classification matching accuracy of the subsequent songs is guaranteed.
  • the third song feature score matrix of all the songs in the preset song library may be matched one by one with the second song feature score of the favorite type list, and the former song with the highest similarity will be matched.
  • the recommended number of songs are recommended songs, and each song is matched, thus ensuring that the final recommended song can satisfy the user's actual song preference.
  • the fourth embodiment of the present application includes:
  • the randomly selected song is recorded as the recommended song corresponding to the favorite type list, and returns to perform the operation of randomly selecting the song from the preset song library until the recording obtains the recommended number of songs corresponding to the favorite type list. So far from the song. Considering that the number of songs included in the preset database is large in the actual situation, if the matching is performed, it takes a lot of processing resources and time, so that the cost of the processing increases sharply. Therefore, in the embodiment of the present application, the matching is randomly matched.
  • the method is to select the recommended songs, as long as the similarity between the song feature score matrix of the song and the song feature score matrix of the favorite type list reaches a preset threshold, the song can be considered as belonging to the user's favorite song, and can be recorded as the recommended song.
  • the matching process can be completed until the number of recorded recommended songs reaches the requirement, which can greatly reduce the cost of matching song matching screening and improve matching efficiency.
  • the method further includes:
  • This step is the same as the processing method of the first embodiment of the present application, and details are not described herein.
  • the song feature score of the blacklist of the obtained song will be utilized in the embodiment of the present application.
  • the matrix matches the generated recommended songs and removes songs that are not preferred by the successful user.
  • the list of recommended songs compiled by other users may be pushed to the user, or there may be It may be a list of recommended songs generated according to some music rankings. Therefore, the recommended songs referred to in the embodiments of the present application refer to all recommended songs to the user, and not only the recommended songs in the first embodiment of the present application.
  • FIG. 6 is a structural block diagram of a song list generating apparatus provided by an embodiment of the present application. For the convenience of description, only parts related to the embodiment of the present application are shown.
  • the song list generating device exemplified in FIG. 6 may be the execution body of the song list generating method provided in the first embodiment.
  • the song list generating apparatus includes:
  • the list extracting module 61 is configured to obtain a user favorite level corresponding to each song in the list of played songs of the user end, and extract N user favorite lists in the list of played songs, and each of the user favorite lists corresponds to a user favorite level greater than a preset level, wherein N is a positive integer.
  • Listing difference module 62 for each of the user's favorite list are split into a list of favorite type M n, each corresponding to one kind of a list of favorite songs type type, M n is a positive integer, n ⁇ [1, N].
  • the matrix dimension reduction module 63 is configured to perform dimension reduction on the first song feature score matrix of each of the favorite type lists corresponding to the dimension P n ⁇ H, and obtain a dimension corresponding to each of the favorite type lists respectively.
  • a second song feature score matrix of H wherein the song feature score matrix is used to record a song feature score of a feature to be analyzed, wherein the feature to be analyzed is a time domain feature, a frequency domain feature, and a cepstrum feature of the song audio data.
  • the song feature score is a corresponding score obtained by analyzing the feature of the song to be analyzed, wherein P n is the number of songs included in the favorite type list, and H is the song to be analyzed
  • P n and H are positive integers.
  • the song screening module 64 is configured to obtain a number of song recommendation corresponding to each song type, and perform similarity on the third song feature score matrix of the song in the preset song library based on each of the second song feature score matrices. Matching, determining a matching successful song of the recommended number of songs corresponding to each of the favorite type lists, and obtaining a list of recommended songs.
  • the list extraction module 61 includes:
  • a song scoring module configured to acquire operation data of each song in the list of played songs by the user terminal, and perform scoring processing on the operation data based on a preset operation score table to obtain the list of played songs The user's favorite score for each song in the song.
  • the rating module is configured to divide the user favorite score based on the preset plurality of favorite level score thresholds, and determine the user favorite level corresponding to each song in the list of played songs.
  • the song list generating device further includes:
  • a feature acquiring module configured to determine the H features of the song to be analyzed.
  • the audio processing module is configured to perform Fourier transform on the song audio data of the song in the user favorite list and perform dimensionality reduction processing to obtain processed song audio data.
  • a score acquisition module configured to set a hidden layer number of the deep belief network to H, and input the reduced-dimensional processed song audio data to the deep belief network for unsupervised training learning, and obtain the H type of the song H song feature scores corresponding to the song features to be analyzed respectively.
  • the song screening module 64 includes:
  • a song selection module configured to randomly select a song from the preset song library, acquire the third song feature score matrix corresponding to a dimension 1 ⁇ H corresponding to the randomly selected song, and select the third song feature score The matrix is similarly matched to the second song feature score.
  • a song recording module configured to record the randomly selected song as a recommended song corresponding to the favorite type list, and return to perform the operation of randomly selecting a song from the preset song library if the similarity matching is successful Until the recommended song of the recommended number of songs corresponding to the favorite type list is recorded.
  • the song list generating device further includes:
  • the blacklist generating module is configured to extract, from the list of the played songs, the L songs whose user preference level is lower than or equal to the preset level, to obtain a blacklist of songs, where L is a positive integer.
  • the blacklist matching module is configured to perform dimension reduction on the fourth song feature score matrix of the L ⁇ H corresponding to the blacklist of the song, to obtain a fifth song feature score matrix corresponding to the dimension 1 ⁇ H.
  • a song culling module configured to perform, according to the fifth song feature score matrix, the similarity matching on a sixth song feature score matrix of the recommended song corresponding to the user end, and reject the successfully matched song in the recommended song .
  • the size of the sequence of the steps in the above embodiments does not mean that the order of execution is performed.
  • the order of execution of each process should be determined by its function and internal logic, and should not be construed as limiting the implementation process of the embodiments of the present application.
  • the terms "first”, “second”, etc. are used in the text to describe various elements in the embodiments of the present application, but these elements should not be limited by these terms. These terms are only used to distinguish one element from another.
  • the first contact can be named a second contact, and similarly, the second contact can be named the first contact without departing from the scope of the various described embodiments. Both the first contact and the second contact are contacts, but they are not the same contact.
  • FIG. 7 is a schematic diagram of a terminal device according to an embodiment of the present application.
  • the terminal device 7 of this embodiment includes a processor 70, a memory 71 in which computer readable instructions 72 executable on the processor 70 are stored.
  • the processor 70 executes the computer readable instructions 72, the steps in the foregoing embodiments of the respective song list generation methods are implemented, such as steps 101 to 104 shown in FIG.
  • the processor 70 when executing the computer readable instructions 72, implements the functions of the various modules/units in the various apparatus embodiments described above, such as the functions of the modules 61-64 shown in FIG.
  • the terminal device 7 may be a computing device such as a desktop computer, a notebook, a palmtop computer, and a cloud server.
  • the terminal device may include, but is not limited to, a processor 70 and a memory 71. It will be understood by those skilled in the art that FIG. 7 is only an example of the terminal device 7, and does not constitute a limitation of the terminal device 7, and may include more or less components than those illustrated, or combine some components or different components.
  • the terminal device may further include an input transmitting device, a network access device, a bus, and the like.
  • the so-called processor 70 can be a central processing unit (Central Processing Unit, CPU), can also be other general purpose processors, digital signal processors (DSP), application specific integrated circuits (Application Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, etc.
  • the general purpose processor may be a microprocessor or the processor or any conventional processor or the like.
  • the memory 71 may be an internal storage unit of the terminal device 7, such as a hard disk or a memory of the terminal device 7.
  • the memory 71 may also be an external storage device of the terminal device 7, for example, a plug-in hard disk equipped on the terminal device 7, and a smart memory card (Smart Media Card, SMC), Secure Digital (Secure Digital, SD) card, flash card (Flash Card) and so on. Further, the memory 71 may also include both an internal storage unit of the terminal device 7 and an external storage device.
  • the memory 71 is configured to store the computer readable instructions and other programs and data required by the terminal device.
  • the memory 71 can also be used to temporarily store data that has been sent or is about to be transmitted.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
  • the above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.
  • the integrated modules/units if implemented in the form of software functional units and sold or used as separate products, may be stored in a computer readable storage medium.
  • the present application implements all or part of the processes in the foregoing embodiments, and may also be implemented by computer readable instructions, which may be stored in a computer readable storage medium.
  • the computer readable instructions when executed by a processor, may implement the steps of the various method embodiments described above.
  • the computer readable instructions comprise computer readable instruction code, which may be in the form of source code, an object code form, an executable file or some intermediate form or the like.
  • the computer readable medium may include any entity or device capable of carrying the computer readable instruction code, a recording medium, a USB flash drive, a removable hard disk, a magnetic disk, an optical disk, a computer memory, a read only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), electrical carrier signals, telecommunications signals, and software distribution media. It should be noted that the content contained in the computer readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in a jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, computer readable media Does not include electrical carrier signals and telecommunication signals.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Multimedia (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种歌曲清单生成方法、装置、终端设备及介质，适用于数据处理技术领域，该方法包括：获取已播放歌曲清单中歌曲的用户喜爱等级并提取出N个用户喜爱清单；将每个用户喜爱清单分别拆分为M n个喜爱类型清单；将每个喜爱类型清单对应的第一歌曲特征分数矩阵进行降维得到每个喜爱类型清单分别对应单维的第二歌曲特征分数矩阵；获取每种歌曲类型的歌曲推荐数目，并基于每个第二歌曲特征分数矩阵对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个喜爱类型清单分别对应的歌曲推荐数目的匹配成功歌曲得到推荐歌曲清单。实现了对用户喜爱推荐歌曲清单生成的个性化生成，极大地提升了对用户推荐歌曲清单生成的精确度。

Description

一种歌曲清单生成方法、装置、终端设备及介质
本申请要求于2018年05月14日提交中国专利局、申请号为201810457101.X 、发明名称为“一种歌曲清单生成方法及终端设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请属于数据处理技术领域，尤其涉及一种歌曲清单生成方法、装置、终端设备及介质。
 背景技术
现有的音乐软件都是简单的根据用户手动选取的歌曲类型标签来进行推荐歌曲清单生成的，如用户选取国语歌/粤语歌/英文歌后推荐对应语言的歌曲清单，但实际上由于歌曲库庞大，每一类标签对应的歌曲数量都是极其庞大的，即每位用户对应的推荐歌曲数量也十分庞大。因此，现有技术中对用户推荐歌曲清单生成的精准度较低，无法实现对用户个性化的推荐歌曲清单生成。
 技术问题
有鉴于此，本申请实施例提供了一种歌曲清单生成方法及终端设备，以解决现有技术中对用户推荐歌曲清单生成的精准度较低，无法实现对用户个性化的推荐歌曲清单生成的问题。
 技术解决方案
本申请实施例的第一方面提供了一种歌曲清单生成方法，包括：
获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数；
将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]；
将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数；
获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
本申请实施例的第二方面提供了一种歌曲清单生成装置，包括：
清单提取模块，用于获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数；
清单差分模块，用于将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]；
矩阵降维模块，用于将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数；
歌曲筛选模块，用于获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
本申请实施例的第三方面提供了一种终端设备，包括存储器、处理器，所述存储器上存储有可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数；
将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]；
将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数；
获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
本申请实施例的第四方面提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，其特征在于，所述计算机可读指令被至少一个处理器执行时实现如下步骤：
获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数；
将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]；
将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数；
获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
 有益效果
通过对用户播放过的歌曲的喜爱等级划分，得到用户不同喜爱程度对应的歌曲清单，再对每个歌曲喜爱清单进行歌曲类型的划分，并对不同类型的歌曲分别进行歌曲特征分数合成，确定出每种类型歌曲的歌曲特点，以为后续的歌曲筛选准确率提供保障。在确定出不同类型歌曲的特点之后，再基于合成的歌曲特征分数矩阵从歌曲库中筛选出推荐数量的推荐歌曲，保证了对用户推荐歌曲时能充分贴合用户日常听歌的特点，在对用户爱听的歌曲多加推荐的同时，又交叉推荐不同类型不同口味用户喜爱的歌曲，以实现在满足用户喜爱歌曲推荐的同时，降低用户对推荐歌曲的厌烦感。通过对用户已播放歌曲清单进行先喜爱程度再喜爱类型的两级拆分，将用户对歌曲的爱好精确定位，再基于对喜爱歌曲的歌曲特征进行分析匹配，以实现对用户喜爱歌曲的精确匹配，因此本申请实施例实现了对用户喜爱推荐歌曲清单生成的个性化生成，极大地提升了对用户推荐歌曲清单生成的精确度。
 附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本发明实施例一提供的歌曲清单生成方法的实现流程示意图；
图2是本发明实施例二提供的歌曲清单生成方法的实现流程示意图；
图3是本发明实施例三提供的歌曲清单生成方法的实现流程示意图；
图4是本发明实施例四提供的歌曲清单生成方法的实现流程示意图；
图5是本发明实施例五提供的歌曲清单生成方法的实现流程示意图；
图6是本发明实施例六提供的歌曲清单生成装置的结构示意图；
图7是本发明实施例七提供的歌曲清单生成终端设备的示意图。
 本发明的实施方式
以下描述中，为了说明而不是为了限定，提出了诸如特定系统结构、技术之类的具体细节，以便透彻理解本申请实施例。然而，本领域的技术人员应当清楚，在没有这些具体细节的其它实施例中也可以实现本申请。在其它情况中，省略对众所周知的系统、装置、电路以及方法的详细说明，以免不必要的细节妨碍本申请的描述。
为了说明本申请所述的技术方案，下面通过具体实施例来进行说明。
图1示出了本申请实施例一提供的歌曲清单生成方法的实现流程图，详述如下：
S101，获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在已播放歌曲清单中提取出N个用户喜爱清单，每个用户喜爱清单对应一个大于预设等级的用户喜爱等级，其中，N为正整数。
其中，用户端即为用户播放歌曲的设备终端，如手机电脑等。已播放歌曲清单是指用户在用户端播放过的歌曲的清单，如在手机上播放过的歌曲清单，考虑到实际情况中一个应用程序可能不具有读取其他应用程序播放记录的权限，因此在结合实际应用时，该已播放歌曲清单可以是指用户端中单个或多个应用程序中播放过的歌曲清单。同时，考虑到实际情况中用户对歌曲的喜好口味可能会随着时间而改变，如当过久没有播放过时可能代表用户对这些歌曲失去了兴趣，因此，本申请实施例中的已播放歌曲清单既可以是用户端播放过的所有歌曲清单，也可以仅是某一段时间内播放过的歌曲清单，如最近一年内播放过的歌曲清单，具体可由技术人员根据需求进行设定。用户喜爱等级，即对用户对一首歌的喜爱程度量化后对应的等级，其既可以是由用户自行对歌曲进行评定得到，也可以是根据用户对歌曲的操作进行评定得到，如根据用户对歌曲播放的次数、是否评论过、是否收藏过以及是否标记过等操作数据进行评分，得到用户对一首歌的喜爱等级，具体可参考本申请实施例二的相关说明。
考虑到实际情况中用户可能会播放自己并不喜欢的歌曲，如听一些在线音乐电台时可能会播放到用户不喜欢的歌曲，因此如果直接以所有已播放歌曲为对象来进行推荐歌曲的分析，可能会导致得到一些用户并不喜欢的歌曲并推送给用户。因此为了保证最终得到的推荐歌曲清单的准确有效，本申请实施例在得到已播放歌曲清单之后会根据歌曲的用户喜爱等级进行筛选，仅将其中高于预设等级的歌曲作为推荐歌曲分析的基础数据，并将基础数据中每个不同用户喜爱等级的歌曲分别提取出来得到与用户喜爱等级一一对应的用户喜爱清单，以此进行推荐歌曲的分析以及清单的生成。例如假设共有第一用户喜爱等级和第二用户喜爱等级两个大于预设等级的用户喜爱等级，此时会将所有第一用户喜爱等级的歌曲提取为一张用户喜爱清单，将所有第二用户喜爱等级的歌曲提取为另一张用户喜爱清单，共得到对应的两张用户喜爱清单。其中预设等级具体等级高低可由技术人员根据实际情况进行设定。
S102，将每个用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]。
考虑到用户可能同时喜爱多种不同类型的歌曲，同时对这些不同类型的歌曲喜爱程度可能会差不多，如用户对摇滚和爵士风格的歌曲喜爱程度差不多，此时若直接对用户喜爱清单进行推荐歌曲分析，难以实现对用户喜爱清单风格的统一，从而使得得到的推荐歌曲的准确度较低。因此，在得到N张用户喜爱清单之后，本申请实施例会继续对每张用户喜爱清单进行歌曲分类，并将其中不同类型的歌曲拆分为不同的喜爱类型清单，如假设用户喜爱清单A中包含着10首歌曲，歌曲分类后确定出其中第一类型的歌曲有3首，第二类型的歌曲有4首，第三类型的歌曲有3首，此时会将用户喜爱清单A拆分为分别包含3首第一类型歌曲、4首第二类型歌曲以及3首第三类型歌曲的三张喜爱类型清单。其中，对歌曲分类的方法包括但不限于如利用神经网络对歌曲进行自适应分类，或者基于技术人员预先对歌曲设置的类型标签来对歌曲进行分类，具体可由技术人员根据实际情况进行设定。
S103，将每个喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，歌曲特征分数是对待分析歌曲特征进行分析得到的对应分数，其中， P n 为喜爱类型清单包含的歌曲数目，H为待分析歌曲特征的种类数， P n 及H均为正整数。
在得到不同的喜爱类型清单之后，需要基于这些喜爱类型清单来进行歌曲筛选匹配，以得到所需的推荐歌曲。为了实现基于喜爱类型清单的推荐歌曲筛选匹配，在本申请实施例中会根据歌曲的歌曲特征分数来进行推荐歌曲的分析匹配，但考虑到喜爱类型清单中包含的歌曲数目可能大于1，此时一个喜爱类型清单对应着多组不同的歌曲特征分数，如当一张喜爱类型清单包含3首歌曲时该喜爱类型清单对应着三组不同的歌曲特征分数，而待匹配的歌曲库中每首歌曲仅对应着一组歌曲特征分数，因此，此时无法直接利用喜爱类型清单来进行推荐歌曲的匹配。因此，为了根据得到的喜爱类型清单来进行推荐歌曲的匹配，本申请实施例中会将喜爱类型清单中包含的多组歌曲特征分数进行降维，以实现对多组歌曲特征分数的合成，使得每张喜爱类型清单仅对应着一组歌曲特征分数，以保证后续推荐歌曲的正常匹配。
其中，本申请实施例具体所需使用到的待分析歌曲特征种类及数量，以及对应的评分方法，均可由技术人员自行设定，如可以直接将歌曲音频数据的短时能量、短时平均过零等时域特征作为本申请实施例的待分析歌曲特征，并利用一些常见的音频分析模型进行评分，也可以使用深度学习算法来对歌曲音频数据进行分析学习，自适应地确定出一些待分析歌曲特征并进行评分，具体可由技术人员根据实际需求进行设定。同时，对于喜爱类型清单的歌曲特征分数矩阵降维合成方法，此处亦不予限定，包括但不限于如将每首歌曲的同一待分析歌曲特征的分数求平均值，并将得到的H个歌曲特征分数平均值作为喜爱类型清单对应的一组歌曲特征分数等，如将喜爱类型清单中所有歌曲的短时能量的分数求平均值，得到喜爱类型清单对应的一个短时能量分数。
应当说明地，由于用户喜爱清单中包含的歌曲数目以及歌曲类型无法确定，因此，本申请实施例中拆分得到的喜爱类型清单的数量，以及每个喜爱类型清单中包含的歌曲数目也无法确定，因此在本申请实施例中 M n 与 P n 的具体值，需根据用户喜爱清单的实际情况，以及所设定的歌曲分类方法来进行确定。
作为本申请的一个优选实施例，本申请实施例中的歌曲特征分数本身可以是一个1×R的向量，其中R为大于1的正整数。考虑到歌曲具有一定的渐变性，即整首歌曲中，歌曲特征可能会随着时间变化而有所不同，如歌曲的节奏等特征，因此，此时若将一个歌曲特征仅通过一个分数值表示，可能难以较为精确地体现出该歌曲特征的实际特点，因此，为了提升对歌曲特征的精确分析，以为后续精确匹配提供保障，本申请实施例中会将整首歌曲按照时间的先后顺序划分为多个歌曲片段，并针对每个片段单独进行歌曲特征分析，从而使得歌曲特征对应着多个不同分数值，得到其对应的歌曲特征分数向量。其中，为了保证对时长不同的歌曲的统一化划分与歌曲特征分析，本申请实施例中可以以歌曲总时长固定比例的形式来将每首歌曲都统一划分为相同数量的歌曲片段，如以总时长的25%、25%、25%以及25%，将歌曲划分为4个片段，此时无论歌曲实际时长是多少都会统一划分为四个歌曲片段进行分析。应当说明地，对于一些与时间变化无关的歌曲特征，如歌曲的总过零点数等，其歌曲特征分数可以不采用向量的方式，而是仍直接使用一个具体的分数值来进行表示。
S104，获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个喜爱类型清单分别对应的歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
在得到每张喜爱类型清单的歌曲特征分数矩阵之后，我们已经确定出了用户所喜爱的歌曲类型数据，因此，只需要将预先设置的歌曲库中的歌曲特征分数矩阵与喜爱类型清单的歌曲特征分数矩阵进行匹配对比，即可判断出歌曲是否是用户所喜爱的类型，是否需要将该歌曲推荐给用户。因此，本申请实施例中会基于得到的喜爱类型清单的歌曲特征分数矩阵来对歌曲库进行歌曲匹配，以确定出所需的推荐歌曲。
由于实际情况中预设歌曲库所包含的歌曲数目一般较多，实际应用中，为了保证得到的推荐歌曲清单的有效性，不可能会无数量限制地进行歌曲推荐，因此本申请实施例中需要设置中歌曲类型对应的所需推荐歌曲的数目。其中，对每种歌曲类型歌曲推荐数目的设置方法可由技术人员自行设定，包括但不限于如直接对每种类型的歌曲设置一个固定的数目，或者对最终得到的推荐歌曲清单设置一个总的固定数目，并根据每张喜爱类型清单所包含的具体歌曲数目，来按比例分配歌曲推荐数目。
作为本申请的一个优选实施例，在生成推荐歌曲清单时，优选地按照歌曲对应的用户喜爱等级从高到低的顺序，将歌曲从上至下进行排序得到推荐歌曲清单，以保证用户喜爱程度高的歌曲的曝光度，提升推荐歌曲清单的有效性。
在本申请实施例中，通过对用户播放过的歌曲的喜爱等级划分，得到用户不同喜爱程度对应的歌曲清单，再对每个歌曲喜爱清单进行歌曲类型的划分，并对不同类型的歌曲分别进行歌曲特征分数合成，确定出每种类型歌曲的歌曲特点，以为后续的歌曲筛选准确率提供保障。在确定出不同类型歌曲的特点之后，再基于合成的歌曲特征分数矩阵从歌曲库中筛选出推荐数量的推荐歌曲，保证了对用户推荐歌曲时能充分贴合用户日常听歌的特点，在对用户爱听的歌曲多加推荐的同时，又交叉推荐不同类型不同口味用户喜爱的歌曲，以实现在满足用户喜爱歌曲推荐的同时，降低用户对推荐歌曲的厌烦感。通过对用户已播放歌曲清单进行先喜爱程度再喜爱类型的两级拆分，将用户对歌曲的爱好精确定位，再基于对喜爱歌曲的歌曲特征进行分析匹配，以实现对用户喜爱歌曲的精确匹配，因此本申请实施例实现了对用户喜爱推荐歌曲清单生成的个性化生成，极大地提升了对用户推荐歌曲清单生成的精确度。
作为对歌曲进行用户喜爱等级评定的一种具体实现方式，如图2所示，本申请实施例二，包括：S201，获取用户端对已播放歌曲清单中每首歌曲的操作数据，并基于预设的操作评分表对操作数据进行评分处理，得到已播放歌曲清单中每首歌曲的用户喜爱分数。
S202，基于预设的多个喜爱等级分数阈值，对用户喜爱分数进行划分，确定出已播放歌曲清单中每首歌曲分别对应的用户喜爱等级。
其中，操作数据是指对用户对歌曲进行播放、评论、收藏以及标记等操作行为进行记录得到的数据。操作评分表是由技术人员预先对用户操作行为进行分析，并为每个操作行为设置对应的分数后得到的数据表。由于不同的操作行为可以反映出用户对该歌曲是否喜爱以及喜爱或者讨厌的程度，如用户若收藏过歌曲，说明其非常喜爱该首歌曲，因此，为了实现对用户歌曲喜爱程度的准确量化，本申请实施例中会由技术人员事先对用户可能的操作行为进行分析量化并设置对应的分数值，再基于得到的操作评分表来对每首歌的操作数据进行分析评估，得到每首歌曲对应的用户喜爱分数，以实现对用户歌曲喜爱程度的有效量化。在得到用户喜爱分数后，根据设定的喜爱分数阈值，对歌曲进行用户喜爱等级的划分，以确定出没首歌曲具体的用户喜爱等级。
以实例进行说明，假设操作数据中记录着用户对歌曲的播放次数、播放时长、是否评论过、是否收藏过以及是否标记喜欢过等数据，同时设定操作评分表如下表1：
表1
单曲循环过 +3 评论过 +4 标记喜欢 +7
播放总时长>7h +5 收藏过 +5 标记不喜欢 -7
分享过 +4 频繁跳过 -4 加入过歌单 +5
此时只需根据歌曲的操作数据进行分析，并依据表1进行各个操作行为的分数查询，即可确定出对应的用户喜爱分数以及对应的用户喜爱等级。
作为对歌曲进行分类的一种具体实现方式，在进行歌曲分类时，包括：对歌曲音频数据进行待分析歌曲特征的提取，并利用神经网络对得到的待分析歌曲特征进行处理，以对歌曲进行自适应分类。
在本申请实施例中，可以使用先进行待分析歌曲特征提取，再基于神经网络对得到的待分析歌曲特征进行处理，以对歌曲进行自适应分类的方法，此时得到的歌曲类型就不是简单的“摇滚”或者“爵士”这种人为划分的类型了，而是更加贴合于待分析歌曲特征的特点的歌曲类型，由于所划分的类型更加贴近歌曲本身的特征特点，因此可以实现歌曲更为精确的分类，从而为推荐歌曲的精确匹配提供基础，保证最终得到的推荐歌曲的精确性。例如，在将歌曲音频数据的短时能量、短时平均过零等时域特征作为待分析歌曲特征，并进行特征数据的提取时，此时会根据得到的这些时域特征的特征数据来对歌曲进行自适应分类，得到与不同时域特征相对应的不同歌曲类型，如自适应分析后将歌曲分为短时能量高且短时平均过零多的作为第一类型歌曲，将短时能量高且短时平均过零少的作为第二类型歌曲，将短时能量低且短时平均过零多的作为第三类型歌曲，并将短时能量低且短时平均过零少作为第四类型歌曲，从而将用户喜爱清单中的歌曲划分为四种不同的类型，并进行喜爱类型清单的拆分。
作为本申请实施例三，如图3所示，在对喜爱类型清单的歌曲特征分数矩阵进行降维合成前，还包括：
S301，确定出H种待分析歌曲特征。其中H具体值需由技术人员根据实际需求设定。
S302，对用户喜爱清单中歌曲的歌曲音频数据进行傅里叶变换并进行降维处理，得到处理后的歌曲音频数据。其中可以利用主要成分分析算法来完成对歌曲音频数据的降维处理。
S303，设置深度信念网络的隐含层数量为H，并将降维处理后的歌曲音频数据输入至深度信念网络进行无监督训练学习，得到歌曲的H种待分析歌曲特征分别对应的H个歌曲特征分数。
在深度信念网络中，网络的每一个隐含层都代表着输入数据在某一特征空间上的标识，这些特征与人工选择的特征不同，中间层特征都是由训练数据得到的，在隐含层处理完成后再利用softmax分类器进行分类，从而使得利用深度信念网络来处理歌曲音频数据更能反映歌曲音频数据本质，网络学习得到的特征具有更好地分类和预测效果，因此，在本申请实施例中，通过使用深度信念网络自适应地对歌曲音频数据进行特征提取评估分类，极大地提升了对歌曲分类的有效性，为后续歌曲的分类匹配准确性提供了保障。
作为歌曲匹配筛选的一种具体实现方式，可以将预设歌曲库中所有歌曲的第三歌曲特征分数矩阵逐个与喜爱类型清单的第二歌曲特征分数进行匹配，并将匹配相似度最高的前歌曲推荐数目位歌曲作为推荐歌曲，由于每首歌曲都进行了匹配，从而保证了最终得到的推荐歌曲能够最大限度地满足用户的实际歌曲喜好需求。
作为歌曲匹配筛选的另一种具体实现方式，如图4所示，本申请实施例四，包括：
S401，从预设歌曲库中随机选取歌曲，获取随机选取歌曲对应的维度为1×H的第三歌曲特征分数矩阵，并将第三歌曲特征分数矩阵与第二歌曲特征分数进行相似度匹配。
S402，若相似度匹配成功，将随机选取歌曲记录为喜爱类型清单对应的推荐歌曲，并返回执行从预设歌曲库中随机选取歌曲的操作，直至记录得到喜爱类型清单对应的歌曲推荐数目的推荐歌曲为止。考虑到实际情况中预设数据库所包含的歌曲数目较多，若每个都进行匹配需要耗费大量的处理资源以及时间，使得处理的成本急剧上升，因此在本申请实施例中，以随机匹配的方式来进行推荐歌曲的挑选，只要歌曲的歌曲特征分数矩阵与喜爱类型清单的歌曲特征分数矩阵相似度达到预设的阈值，即可认为该歌曲属于用户喜爱的歌曲，可以作为推荐歌曲进行记录，直至记录的推荐歌曲数目达到要求即可完成匹配的过程，这样可以极大地减小推荐歌曲匹配筛选的成本，提高匹配效率。
作为本申请实施例五，如图5所示，在获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级之后，还包括：
S501，将已播放歌曲清单中用户喜爱等级低于或等于预设等级的L首歌曲提取得到歌曲黑名单，L为正整数。当用户喜爱等级非常低时，说明用户并不喜欢这些歌曲，可能只是不小心播放到了而已，因此，为了保证推荐歌曲的准确性，本申请实施例中会将这些歌曲单独提取为一张歌曲黑名单，以供后续对推荐歌曲的筛选处理。
S502，将歌曲黑名单对应的维度为L×H的第四歌曲特征分数矩阵进行降维，得到对应维度为1×H的第五歌曲特征分数矩阵。
本步骤与本申请实施例一的处理方式相同，此处不予赘述。
S503，基于第五歌曲特征分数矩阵，对用户端对应的推荐歌曲的第六歌曲特征分数矩阵进行相似度匹配，并剔除推荐歌曲中匹配成功的歌曲。
在确定出用户不喜欢歌曲的歌曲黑名单的歌曲特征分数矩阵后，为了保证对用户歌曲推荐的精准有效，提升推荐歌曲的质量，本申请实施例中会利用得到的歌曲黑名单的歌曲特征分数矩阵来对生成的推荐歌曲进行匹配，并剔除匹配成功的用户不喜爱的歌曲。其中，应当特别说明地，由于实际情况中推荐歌曲的来源很多，不仅仅只是上述本申请实施例一中计算处理得到，有可能是其他用户自行整理的推荐歌曲清单并推送给用户的，或者也有可能是根据一些音乐排行榜生成的推荐歌曲清单，因此，本申请实施例中所指的推荐歌曲是指所有对用户的推荐歌曲，而非仅本申请实施例一中的推荐歌曲。
对应于上文实施例的方法，图6示出了本申请实施例提供的歌曲清单生成装置的结构框图，为了便于说明，仅示出了与本申请实施例相关的部分。图6示例的歌曲清单生成装置可以是前述实施例一提供的歌曲清单生成方法的执行主体。
参照图6，该歌曲清单生成装置包括：
清单提取模块61，用于获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数。
清单差分模块62，用于将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]。
矩阵降维模块63，用于将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数。
歌曲筛选模块64，用于获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
进一步地，所述清单提取模块61，包括：
歌曲评分模块，用于获取所述用户端对所述已播放歌曲清单中每首歌曲的操作数据，并基于预设的操作评分表对所述操作数据进行评分处理，得到所述已播放歌曲清单中每首歌曲的用户喜爱分数。
等级评定模块，用于基于预设的多个喜爱等级分数阈值，对所述用户喜爱分数进行划分，确定出所述已播放歌曲清单中每首歌曲分别对应的所述用户喜爱等级。
进一步地，该歌曲清单生成装置，还包括：
特征获取模块，用于确定出H种所述待分析歌曲特征。
音频处理模块，用于对所述用户喜爱清单中歌曲的歌曲音频数据进行傅里叶变换并进行降维处理，得到处理后的歌曲音频数据。
分数获取模块，用于设置深度信念网络的隐含层数量为H，并将所述降维处理后的歌曲音频数据输入至所述深度信念网络进行无监督训练学习，得到歌曲的H种所述待分析歌曲特征分别对应的H个所述歌曲特征分数。
进一步地，所述歌曲筛选模块64，包括：
歌曲选取模块，用于从所述预设歌曲库中随机选取歌曲，获取所述随机选取歌曲对应的维度为1×H的所述第三歌曲特征分数矩阵，并将所述第三歌曲特征分数矩阵与所述第二歌曲特征分数进行相似度匹配。
歌曲记录模块，用于若所述相似度匹配成功，将所述随机选取歌曲记录为所述喜爱类型清单对应的推荐歌曲，并返回执行所述从所述预设歌曲库中随机选取歌曲的操作，直至记录得到所述喜爱类型清单对应的所述歌曲推荐数目的推荐歌曲为止。
进一步地，该歌曲清单生成装置，还包括：
黑名单生成模块，用于将所述已播放歌曲清单中所述用户喜爱等级低于或等于所述预设等级的L首歌曲提取得到歌曲黑名单，L为正整数。
黑名单匹配模块，用于将所述歌曲黑名单对应的维度为L×H的第四歌曲特征分数矩阵进行降维，得到对应维度为1×H的第五歌曲特征分数矩阵。
歌曲剔除模块，用于基于所述第五歌曲特征分数矩阵，对所述用户端对应的推荐歌曲的第六歌曲特征分数矩阵进行所述相似度匹配，并剔除所述推荐歌曲中匹配成功的歌曲。
本申请实施例提供的歌曲清单生成装置中各模块实现各自功能的过程，具体可参考前述图1所示实施例一的描述，此处不再赘述。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。还应理解的，虽然术语“第一”、“第二”等在文本中在一些本申请实施例中用来描述各种元素，但是这些元素不应该受到这些术语的限制。这些术语只是用来将一个元素与另一元素区分开。例如，第一接触可以被命名为第二接触，并且类似地，第二接触可以被命名为第一接触，而不背离各种所描述的实施例的范围。第一接触和第二接触都是接触，但是它们不是同一接触。
图7是本申请一实施例提供的终端设备的示意图。如图7所示，该实施例的终端设备7包括：处理器70、存储器71，所述存储器71中存储有可在所述处理器70上运行的计算机可读指令72。所述处理器70执行所述计算机可读指令72时实现上述各个歌曲清单生成方法实施例中的步骤，例如图1所示的步骤101至104。或者，所述处理器70执行所述计算机可读指令72时实现上述各装置实施例中各模块/单元的功能，例如图6所示模块61至64的功能。所述终端设备7可以是桌上型计算机、笔记本、掌上电脑及云端服务器等计算设备。所述终端设备可包括，但不仅限于，处理器70、存储器71。本领域技术人员可以理解，图7仅仅是终端设备7的示例，并不构成对终端设备7的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如所述终端设备还可以包括输入发送设备、网络接入设备、总线等。
所称处理器70可以是中央处理单元（Central Processing Unit，CPU），还可以是其他通用处理器、数字信号处理器（Digital Signal Processor，DSP）、专用集成电路（Application Specific Integrated Circuit，ASIC）、现成可编程门阵列（Field-Programmable Gate Array，FPGA）或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。所述存储器71可以是所述终端设备7的内部存储单元，例如终端设备7的硬盘或内存。所述存储器71也可以是所述终端设备7的外部存储设备，例如所述终端设备7上配备的插接式硬盘，智能存储卡（Smart Media Card，SMC），安全数字（Secure Digital，SD）卡，闪存卡（Flash Card）等。进一步地，所述存储器71还可以既包括所述终端设备7的内部存储单元也包括外部存储设备。所述存储器71用于存储所述计算机可读指令以及所述终端设备所需的其他程序和数据。所述存储器71还可以用于暂时地存储已经发送或者将要发送的数据。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
所述集成的模块/单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实现上述实施例方法中的全部或部分流程，也可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一计算机可读存储介质中，该计算机可读指令在被处理器执行时，可实现上述各个方法实施例的步骤。其中，所述计算机可读指令包括计算机可读指令代码，所述计算机可读指令代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述计算机可读介质可以包括：能够携带所述计算机可读指令代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器（Read-Only Memory，ROM）、随机存取存储器（Random Access Memory，RAM）、电载波信号、电信信号以及软件分发介质等。需要说明的是，所述计算机可读介质包含的内容可以根据司法管辖区内立法和专利实践的要求进行适当的增减，例如在某些司法管辖区，根据立法和专利实践，计算机可读介质不包括电载波信号和电信信号。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使对应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种歌曲清单生成方法，其特征在于，包括：
    获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数；
    将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]；
    将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数；
    获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
  2. 如权利要求1所述的歌曲清单生成方法，其特征在于，所述获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，包括：
    获取所述用户端对所述已播放歌曲清单中每首歌曲的操作数据，并基于预设的操作评分表对所述操作数据进行评分处理，得到所述已播放歌曲清单中每首歌曲的用户喜爱分数；
    基于预设的多个喜爱等级分数阈值，对所述用户喜爱分数进行划分，确定出所述已播放歌曲清单中每首歌曲分别对应的所述用户喜爱等级。
  3. 如权利要求1所述的歌曲清单生成方法，其特征在于，在所述将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维之前，还包括：
    确定出H种所述待分析歌曲特征；
    对所述用户喜爱清单中歌曲的歌曲音频数据进行傅里叶变换并进行降维处理，得到处理后的歌曲音频数据；
    设置深度信念网络的隐含层数量为H，并将所述降维处理后的歌曲音频数据输入至所述深度信念网络进行无监督训练学习，得到歌曲的H种所述待分析歌曲特征分别对应的H个所述歌曲特征分数。
  4. 如权利要求1所述的歌曲清单生成方法，其特征在于，所述基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的推荐歌曲，包括：
    从所述预设歌曲库中随机选取歌曲，获取所述随机选取歌曲对应的维度为1×H的所述第三歌曲特征分数矩阵，并将所述第三歌曲特征分数矩阵与所述第二歌曲特征分数进行相似度匹配；
    若所述相似度匹配成功，将所述随机选取歌曲记录为所述喜爱类型清单对应的推荐歌曲，并返回执行所述从所述预设歌曲库中随机选取歌曲的操作，直至记录得到所述喜爱类型清单对应的所述歌曲推荐数目的推荐歌曲为止。
  5. 如权利要求1所述的歌曲清单生成方法，其特征在于，在所述获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级之后，还包括：
    将所述已播放歌曲清单中所述用户喜爱等级低于或等于所述预设等级的L首歌曲提取得到歌曲黑名单，L为正整数；
    将所述歌曲黑名单对应的维度为L×H的第四歌曲特征分数矩阵进行降维，得到对应维度为1×H的第五歌曲特征分数矩阵；
    基于所述第五歌曲特征分数矩阵，对所述用户端对应的推荐歌曲的第六歌曲特征分数矩阵进行所述相似度匹配，并剔除所述推荐歌曲中匹配成功的歌曲。
  6. 一种歌曲清单生成装置，其特征在于，包括：
    清单提取模块，用于获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数；
    清单差分模块，用于将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]；
    矩阵降维模块，用于将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数；
    歌曲筛选模块，用于获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
  7. 如权利要求6所述的歌曲清单生成装置，其特征在于，所述清单提取模块，包括：
    歌曲评分模块，用于获取所述用户端对所述已播放歌曲清单中每首歌曲的操作数据，并基于预设的操作评分表对所述操作数据进行评分处理，得到所述已播放歌曲清单中每首歌曲的用户喜爱分数；
    等级评定模块，用于基于预设的多个喜爱等级分数阈值，对所述用户喜爱分数进行划分，确定出所述已播放歌曲清单中每首歌曲分别对应的所述用户喜爱等级。
  8. 如权利要求6所述的歌曲清单生成装置，其特征在于，还包括：
    特征获取模块，用于确定出H种所述待分析歌曲特征；
    音频处理模块，用于对所述用户喜爱清单中歌曲的歌曲音频数据进行傅里叶变换并进行降维处理，得到处理后的歌曲音频数据；
    分数获取模块，用于设置深度信念网络的隐含层数量为H，并将所述降维处理后的歌曲音频数据输入至所述深度信念网络进行无监督训练学习，得到歌曲的H种所述待分析歌曲特征分别对应的H个所述歌曲特征分数。
  9. 如权利要求6所述的歌曲清单生成装置，其特征在于，所述歌曲筛选模块，还包括：
    歌曲选取模块，用于从所述预设歌曲库中随机选取歌曲，获取所述随机选取歌曲对应的维度为1×H的所述第三歌曲特征分数矩阵，并将所述第三歌曲特征分数矩阵与所述第二歌曲特征分数进行相似度匹配；
    歌曲记录模块，用于若所述相似度匹配成功，将所述随机选取歌曲记录为所述喜爱类型清单对应的推荐歌曲，并返回执行所述从所述预设歌曲库中随机选取歌曲的操作，直至记录得到所述喜爱类型清单对应的所述歌曲推荐数目的推荐歌曲为止。
  10. 如权利要求6所述的歌曲清单生成装置，其特征在于，还包括：
    黑名单生成模块，用于将所述已播放歌曲清单中所述用户喜爱等级低于或等于所述预设等级的L首歌曲提取得到歌曲黑名单，L为正整数；
    黑名单匹配模块，用于将所述歌曲黑名单对应的维度为L×H的第四歌曲特征分数矩阵进行降维，得到对应维度为1×H的第五歌曲特征分数矩阵；
    歌曲剔除模块，用于基于所述第五歌曲特征分数矩阵，对所述用户端对应的推荐歌曲的第六歌曲特征分数矩阵进行所述相似度匹配，并剔除所述推荐歌曲中匹配成功的歌曲。
  11. 一种终端设备，其特征在于，所述终端设备包括存储器、处理器，所述存储器上存储有可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
    获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数；
    将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]；
    将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数；
    获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
  12. 如权利要求11所述的终端设备，其特征在于，所述获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，包括：
    获取所述用户端对所述已播放歌曲清单中每首歌曲的操作数据，并基于预设的操作评分表对所述操作数据进行评分处理，得到所述已播放歌曲清单中每首歌曲的用户喜爱分数；
    基于预设的多个喜爱等级分数阈值，对所述用户喜爱分数进行划分，确定出所述已播放歌曲清单中每首歌曲分别对应的所述用户喜爱等级。
  13. 如权利要求11所述的终端设备，其特征在于，在所述将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维之前，还包括：
    确定出H种所述待分析歌曲特征；
    对所述用户喜爱清单中歌曲的歌曲音频数据进行傅里叶变换并进行降维处理，得到处理后的歌曲音频数据；
    设置深度信念网络的隐含层数量为H，并将所述降维处理后的歌曲音频数据输入至所述深度信念网络进行无监督训练学习，得到歌曲的H种所述待分析歌曲特征分别对应的H个所述歌曲特征分数。
  14. 如权利要求11所述的终端设备，其特征在于，所述基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的推荐歌曲，包括：
    从所述预设歌曲库中随机选取歌曲，获取所述随机选取歌曲对应的维度为1×H的所述第三歌曲特征分数矩阵，并将所述第三歌曲特征分数矩阵与所述第二歌曲特征分数进行相似度匹配；
    若所述相似度匹配成功，将所述随机选取歌曲记录为所述喜爱类型清单对应的推荐歌曲，并返回执行所述从所述预设歌曲库中随机选取歌曲的操作，直至记录得到所述喜爱类型清单对应的所述歌曲推荐数目的推荐歌曲为止。
  15. 如权利要求11所述的终端设备，其特征在于，在所述获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级之后，还包括：
    将所述已播放歌曲清单中所述用户喜爱等级低于或等于所述预设等级的L首歌曲提取得到歌曲黑名单，L为正整数；
    将所述歌曲黑名单对应的维度为L×H的第四歌曲特征分数矩阵进行降维，得到对应维度为1×H的第五歌曲特征分数矩阵；
    基于所述第五歌曲特征分数矩阵，对所述用户端对应的推荐歌曲的第六歌曲特征分数矩阵进行所述相似度匹配，并剔除所述推荐歌曲中匹配成功的歌曲。
  16. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，其特征在于，所述计算机可读指令被至少一个处理器执行时实现如下步骤：
    获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，并在所述已播放歌曲清单中提取出N个用户喜爱清单，每个所述用户喜爱清单对应一个大于预设等级的所述用户喜爱等级，其中，N为正整数；
    将每个所述用户喜爱清单分别拆分为 M n 个喜爱类型清单，每个所述喜爱类型清单对应一种歌曲类型， M n 为正整数，n∈[1，N]；
    将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维，得到每个所述喜爱类型清单分别对应的维度为1×H的第二歌曲特征分数矩阵，所述歌曲特征分数矩阵用于记录待分析歌曲特征的歌曲特征分数，所述待分析歌曲特征为歌曲音频数据的时域特征、频域特征以及倒谱特征中的一种或多种特征，所述歌曲特征分数是对所述待分析歌曲特征进行分析得到的对应分数，其中， P n 为所述喜爱类型清单包含的歌曲数目，H为所述待分析歌曲特征的种类数， P n 及H均为正整数；
    获取每种歌曲类型分别对应的歌曲推荐数目，并基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的匹配成功歌曲，得到推荐歌曲清单。
  17. 根据权利要求16所述的计算机可读存储介质，其特征在于，所述获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级，包括：
    获取所述用户端对所述已播放歌曲清单中每首歌曲的操作数据，并基于预设的操作评分表对所述操作数据进行评分处理，得到所述已播放歌曲清单中每首歌曲的用户喜爱分数；
    基于预设的多个喜爱等级分数阈值，对所述用户喜爱分数进行划分，确定出所述已播放歌曲清单中每首歌曲分别对应的所述用户喜爱等级。
  18. 根据权利要求16所述的计算机可读存储介质，其特征在于，在所述将每个所述喜爱类型清单对应的维度为 P n ×H的第一歌曲特征分数矩阵进行降维之前，还包括：
    确定出H种所述待分析歌曲特征；
    对所述用户喜爱清单中歌曲的歌曲音频数据进行傅里叶变换并进行降维处理，得到处理后的歌曲音频数据；
    设置深度信念网络的隐含层数量为H，并将所述降维处理后的歌曲音频数据输入至所述深度信念网络进行无监督训练学习，得到歌曲的H种所述待分析歌曲特征分别对应的H个所述歌曲特征分数。
  19. 根据权利要求16所述的计算机可读存储介质，其特征在于，所述基于每个所述第二歌曲特征分数矩阵，分别对预设歌曲库中歌曲的第三歌曲特征分数矩阵进行相似度匹配，确定出每个所述喜爱类型清单分别对应的所述歌曲推荐数目的推荐歌曲，包括：
    从所述预设歌曲库中随机选取歌曲，获取所述随机选取歌曲对应的维度为1×H的所述第三歌曲特征分数矩阵，并将所述第三歌曲特征分数矩阵与所述第二歌曲特征分数进行相似度匹配；
    若所述相似度匹配成功，将所述随机选取歌曲记录为所述喜爱类型清单对应的推荐歌曲，并返回执行所述从所述预设歌曲库中随机选取歌曲的操作，直至记录得到所述喜爱类型清单对应的所述歌曲推荐数目的推荐歌曲为止。
  20. 根据权利要求16所述的计算机可读存储介质，其特征在于，在所述获取用户端的已播放歌曲清单中每首歌曲分别对应的用户喜爱等级之后，还包括：
    将所述已播放歌曲清单中所述用户喜爱等级低于或等于所述预设等级的L首歌曲提取得到歌曲黑名单，L为正整数；
    将所述歌曲黑名单对应的维度为L×H的第四歌曲特征分数矩阵进行降维，得到对应维度为1×H的第五歌曲特征分数矩阵；
    基于所述第五歌曲特征分数矩阵，对所述用户端对应的推荐歌曲的第六歌曲特征分数矩阵进行所述相似度匹配，并剔除所述推荐歌曲中匹配成功的歌曲。
PCT/CN2018/097112 2018-05-14 2018-07-25 一种歌曲清单生成方法、装置、终端设备及介质 WO2019218462A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810457101.XA CN108629047B (zh) 2018-05-14 2018-05-14 一种歌曲清单生成方法及终端设备
CN201810457101.X 2018-05-14

Publications (1)

Publication Number Publication Date
WO2019218462A1 true WO2019218462A1 (zh) 2019-11-21

Family

ID=63692957

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/097112 WO2019218462A1 (zh) 2018-05-14 2018-07-25 一种歌曲清单生成方法、装置、终端设备及介质

Country Status (2)

Country Link
CN (1) CN108629047B (zh)
WO (1) WO2019218462A1 (zh)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109582820B (zh) * 2018-11-23 2021-05-25 维沃移动通信有限公司 一种歌曲播放方法、终端设备及服务器
CN109885719A (zh) * 2019-01-24 2019-06-14 广州小鹏汽车科技有限公司 一种歌曲推荐方法、系统、终端及存储介质
CN111753124B (zh) * 2019-03-29 2024-06-18 Tcl科技集团股份有限公司 一种音乐推荐方法及服务器
CN109992694A (zh) * 2019-04-17 2019-07-09 重庆誉存大数据科技有限公司 一种音乐智能推荐方法及系统
CN112163116A (zh) * 2020-09-28 2021-01-01 广州酷狗计算机科技有限公司 歌曲归类方法、装置及计算机可读存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102207972A (zh) * 2011-06-15 2011-10-05 中山大学 一种数字电视的影视节目推荐方法及其装置
CN105138696A (zh) * 2015-09-24 2015-12-09 深圳市冠旭电子有限公司 一种音乐推送方法及装置
CN106202073A (zh) * 2015-04-30 2016-12-07 中国电信股份有限公司 音乐推荐方法及系统
CN106528766A (zh) * 2016-11-04 2017-03-22 北京云知声信息技术有限公司 相似歌曲推荐方法及装置
US20170109642A1 (en) * 2015-10-16 2017-04-20 Adobe Systems Incorporated Particle Thompson Sampling for Online Matrix Factorization Recommendation
CN107908701A (zh) * 2017-11-06 2018-04-13 广东欧珀移动通信有限公司 音乐推荐的方法、装置、存储介质及终端设备
CN108021568A (zh) * 2016-10-31 2018-05-11 北京酷我科技有限公司 一种歌单推荐方法及装置

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101180619B (zh) * 2005-02-28 2012-05-23 雅虎公司 媒体管理系统和方法
CN105868325A (zh) * 2016-03-25 2016-08-17 苏凤芝 一种音乐播放方法和系统

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102207972A (zh) * 2011-06-15 2011-10-05 中山大学 一种数字电视的影视节目推荐方法及其装置
CN106202073A (zh) * 2015-04-30 2016-12-07 中国电信股份有限公司 音乐推荐方法及系统
CN105138696A (zh) * 2015-09-24 2015-12-09 深圳市冠旭电子有限公司 一种音乐推送方法及装置
US20170109642A1 (en) * 2015-10-16 2017-04-20 Adobe Systems Incorporated Particle Thompson Sampling for Online Matrix Factorization Recommendation
CN108021568A (zh) * 2016-10-31 2018-05-11 北京酷我科技有限公司 一种歌单推荐方法及装置
CN106528766A (zh) * 2016-11-04 2017-03-22 北京云知声信息技术有限公司 相似歌曲推荐方法及装置
CN107908701A (zh) * 2017-11-06 2018-04-13 广东欧珀移动通信有限公司 音乐推荐的方法、装置、存储介质及终端设备

Also Published As

Publication number Publication date
CN108629047B (zh) 2023-04-14
CN108629047A (zh) 2018-10-09

Similar Documents

Publication Publication Date Title
WO2019218462A1 (zh) 一种歌曲清单生成方法、装置、终端设备及介质
US11645301B2 (en) Cross media recommendation
Moscato et al. An emotional recommender system for music
Vall et al. Feature-combination hybrid recommender systems for automated music playlist continuation
CN110909205B (zh) 一种视频封面确定方法、装置、电子设备及可读存储介质
Kotsakis et al. Investigation of broadcast-audio semantic analysis scenarios employing radio-programme-adaptive pattern classification
CN113903346A (zh) 一种基于深度学习的音域平衡方法、装置及系统
CN109325146B (zh) 一种视频推荐方法、装置、存储介质和服务器
CA2887359A1 (en) Multi-input playlist selection
CN107967280B (zh) 一种标签推荐歌曲的方法及系统
CN111259192A (zh) 音频推荐方法和装置
McFee et al. Learning Similarity from Collaborative Filters.
Horsburgh et al. Learning pseudo-tags to augment sparse tagging in hybrid music recommender systems
CN107885745A (zh) 一种歌曲推荐方法及装置
CN106339507A (zh) 流媒体消息推送方法和装置
CN104090880A (zh) 音频文件均衡器参数配置的方法和装置
Wang et al. Finetuning convolutional neural networks for visual aesthetics
Panagakis et al. Music classification by low-rank semantic mappings
Chen et al. Music recommendation based on multiple contextual similarity information
Kostek Listening to live music: life beyond music recommendation systems
CN112446206A (zh) 一种菜谱标题的生成方法及装置
Tsipas et al. Content-based music structure analysis using vector quantization
US12013864B2 (en) Method for automatically generating news events of a certain topic and electronic device applying the same
Ospitia-Medina et al. Music Recommender Systems: A Review Centered on Biases
Yeh et al. Improving music auto-tagging by intra-song instance bagging

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18918711

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18918711

Country of ref document: EP

Kind code of ref document: A1