Patent Application
20200042879
A number of publications, identified as References [1] to [39], are listed in a section entitled “REFERENCES” located at the end of the DETAILED DESCRIPTION herein. Those References will be referred to throughout this application.
The field of Music Information Retrieval (MIR) concerns itself, among other things, with the analysis of music in its many facets, such as melody, timbre or rhythm (see, e.g., publication [20]). Among those aspects, popular western commercial music (i.e., “pop” music) is arguably characterized by emphasizing mainly the melody and accompaniment aspects of music. For purposes of simplicity, the melody, or main musical melodic line, also referred to herein as a “foreground”, and the accompaniment also is referred to herein a “background” (see, e.g., Reference [27]). Typically, in pop music the melody is sung, whereas the accompaniment is performed by one or more instrumentalists. Often, a singer delivers the lyrics, and the backing musicians provide harmony as well as genre and style cues (see, e.g., Reference [29]).
The task of automatic singing voice separation consists of estimating what the sung melody and accompaniment would sound like in isolation. A clean vocal signal is helpful for other related MIR tasks, such as singer identification (see, e.g., Reference [18]) and lyric transcription (see, e.g., Reference [17]). As for commercial applications, it is evident that the karaoke industry, estimated to be worth billions of dollars globally (see, e.g., Reference [4]), would directly benefit from such technology.
Several techniques have been proposed for blind source separation of musical audio. Successful results have been achieved with non-negative matrix factorization (see, e.g., References [26, 30, 32]), Bayesian methods (see, e.g., Reference [21]), and the analysis of repeating structures (See, e.g., Reference [23]).
Deep learning models have recently emerged as powerful alternatives to traditional methods. Notable examples include Reference [25] where a deep feed-forward network learns to estimate an ideal binary spectrogram mask that represents the spectrogram bins in which the vocal is more prominent than the accompaniment. In Reference [9], the authors employ a deep recurrent architecture to predict soft masks that are multiplied with the original signal to obtain the desired isolated source.
Convolutional encoder-decoder architectures have been explored in the context of singing voice separation in References [6] and [8]. In both of these works, spectrograms are compressed through a bottleneck layer and re-expanded to the size of the target spectrogram. While this “hourglass” architecture is undoubtedly successful in discovering global patterns, it is unclear how much local detail is lost during contraction.
One potential weakness shared by the publications cited above is the lack of large training datasets. Existing models are usually trained on hundreds of tracks of lower-than-commercial quality, and may therefore suffer from poor generalization.
Over the last few years, considerable improvements have occurred in the family of machine learning algorithms known as image-to-image translation (see, e.g., Reference [11]) pixel-level classification (see, e.g., Reference [2]), automatic colorization (see, e.g., Reference [33]), and image segmentation (see, e.g., Reference [1])—largely driven by advances in the design of novel neural network architectures.
It is with respect to these and other general considerations that embodiments have been described. Also, although relatively specific problems have been discussed, it should be understood that the embodiments should not be limited to solving the specific problems identified in the background.
The foregoing and other limitations are overcome by a system, method and computer product for training a neural network system. In one example embodiment herein, the method comprises applying an audio signal to the neural network system, the audio signal including a vocal component and a non-vocal component. The method also comprises comparing an output of the neural network system to a target signal, and adjusting at least one parameter of the neural network system to reduce a result of the comparing, for training the neural network system to estimate one of the vocal component and the non-vocal component. According to an example aspect herein, the neural network system includes a U-Net.
Also in one example embodiment herein, the audio signal and the target signal represent different versions of a same musical song, the audio signal includes mixed vocal and non-vocal (e.g., instrumental) content (i.e., the audio signal is therefore also referred to herein as a “mixed signal”, or an “input (mixed) signal”), and the target signal includes either vocal content or non-vocal content. Also in one example embodiment herein, the non-vocal component is an instrumental component, and the target signal represents an instrumental signal or a vocal signal. According to another example embodiment herein, the method further comprises obtaining the target signal by removing an instrumental component from a signal that includes vocal and instrumental components.
Additionally, in still another example embodiment herein, the method further comprises identifying the audio signal and the target signal as a pair, wherein the identifying includes determining at least one of:
that the audio signal and the target signal relate to a same artist,
that a title associated with at least one of the audio signal and the target signal does not include predetermined information, and
that durations of the audio signal and the target signal differ by no more than a predetermined length of time.
In one example embodiment herein, the method further comprises converting the audio signal to an image in the neural network system, and the U-Net comprises a convolution path for encoding the image, and a deconvolution path for decoding the image encoded by the convolution path.
The U-Net, in one example embodiment herein, additionally comprises concatenations between the paths (e.g., encoder and decoder paths). Moreover, in one example embodiment herein, the method further comprises applying an output of the deconvolution path as a mask to the image.
A system, method and computer product also are provided herein for estimating a component of a provided audio signal, according to an example aspect herein. The method comprises converting the provided audio signal to an image, and applying the image to a U-Net trained to estimate one of vocal content and instrumental content. The method of this aspect of the present application also comprises converting an output of the U-Net to an output audio signal. The output audio signal represents an estimate of either a vocal component of the provided audio signal or an instrumental component of the provided audio signal, depending on whether the U-Net is trained to estimate the vocal content or the instrumental content, respectively.
According to another example aspect herein, another system, method, and computer product are provided for estimating a component of a provided audio signal. In one example embodiment, the method comprises converting the provided audio signal to an image, and inputting the image to a U-Net trained to estimate different types of content. The different types of content include one or more of vocal content and musical instrument content. In response to the input image, the U-Net outputs signals, each representing a corresponding one of the different types of content. The method also comprises converting each of the signals output by the U-Net to an audio signal.
The U-Net according to this example embodiment preferably comprises a convolution path for encoding the image, and a plurality of deconvolution paths for decoding the image encoded by the convolution path, each of the deconvolution paths corresponding to one of the different types of content.
According to one example embodiment herein, the method further comprises applying an output of at least one of the deconvolution paths as a mask to the image.
The musical instrument content, in one example embodiment, includes different types of musical instrument content, and each of the signals output by the U-Net corresponds to one of the different types of musical instrument content, in a case where the U-Net is trained to estimate the one of the different types of content.
In one example embodiment herein, the different types of musical instrument content include at least some of guitar content, bass content, drum content, piano content, and/or any other type(s) of musical instrument content.
According to still another example aspect herein, a method, system, and computer program product are provided for training a neural network system. In one example embodiment the method comprises inputting an audio signal to the neural network system to generate a plurality of outputs, the audio signal including different types of content. The different types of content includes one or more of vocal content and musical instrument content, and each output corresponds to a respective one of the different types of content. The method also comprises comparing individual ones of the outputs of the neural network system to corresponding target signals, and, for each individual output compared in the comparing, adjusting at least one parameter of the neural network system to reduce a result of the comparing performed for the individual output, to train the neural network system to estimate the different types of content.
The neural network system includes a U-Net, in one example embodiment herein, and the U-Net comprises a convolution path and a plurality of de-convolution paths.
In one example embodiment herein, the method further comprises obtaining the target signal by removing an instrumental component from a signal that includes vocal and instrumental components.
Also in one example embodiment herein, the musical instrument content includes different types of musical instrument content, such as at least some of guitar content, bass content, drum content, piano content, and/or any other type(s) of musical instrument content.
Additionally according to an example embodiment herein, the method further comprises converting the audio signal to an image in the neural network system, and applying an output of at least one of the deconvolution paths as a mask to the image.
In the following detailed description, references are made to the accompanying drawings that form a part hereof, and in which are shown by way of illustrations specific embodiments or examples. These aspects may be combined, other aspects may be utilized, and structural changes may be made without departing from the present disclosure. Embodiments may be practiced as methods, systems or devices. Accordingly, embodiments may take the form of a hardware implementation, an entirely software implementation, or an implementation combining software and hardware aspects. The following detailed description is therefore not to be taken in a limiting sense, and the scope of the present disclosure is defined by the appended claims and their equivalents.
Example aspects described herein address a voice (or other content) separation task, whose domain is often considered from a time-frequency perspective, as the translation of a mixed spectrogram into vocal and instrumental spectrograms. By using this framework, the technology exploits to advantage some advances in image-to-image translation—especially in regard to the reproduction of fine-grained details—for use in blind source separation for music.
The decomposition of a music audio signal into its vocal and backing track components, and/or into individual ones of a plurality of instrument components, is analogous to image-to-image translation, where a mixed spectrogram is transformed into its constituent sources. According to an example aspect herein, a U-Net architecture—initially developed for medical imaging—is employed for the task of source separation, given its proven capacity for recreating the fine, low-level detail required for high-quality audio reproduction. At least some example embodiments herein, through both quantitative evaluation and subjective assessment, demonstrate that they achieve state-of-the-art performance.
An example aspect described herein adapts a U-Net architecture to the task of content separation. That architecture was introduced originally in biomedical imaging, to improve precision and localization of microscopic images of neuronal structures. The architecture builds upon a fully convolutional network (see, e.g., Reference [14]) and, in one example, may be similar to the deconvolutional network (see, e.g., Reference [19]). In a deconvolutional network, a stack of convolutional layers—where each layer halves the size of an image but doubles the number of channels—encodes the image into a small and deep representation. That encoding is then decoded to the original size of the image by a stack of upsampling layers.
In the reproduction of a natural image, displacements by just one pixel are usually not perceived as major distortions. In the frequency domain, however, even a minor linear shift in a spectrogram may have significant effects on perception. This is particularly relevant in music signals, because of the logarithmic perception of frequency. Moreover, a shift in the time dimension can become audible as jitter and other artifacts. Therefore, it can be useful that a reproduction preserves a high level of detail. According to an example aspect herein, the U-Net architecture herein adds additional skip connections between layers at the same hierarchical level in the encoder and decoder. This enables low-level information to flow directly from the high-resolution input to the high-resolution output.
At least one neural network architecture described herein, according to one example embodiment, can predict vocal and instrumental components of an input signal indirectly. In one example embodiment herein, an output of a final decoder layer is a soft mask that is multiplied element-wise with a mixed spectrogram to obtain a final estimate. Also in one example embodiment herein, separate models are trained for the extraction of instrumental and vocal components, respectively, of a signal, to allow for more divergent training schemes for the models in the future. Also, in one example embodiment herein, models are trained for the extraction of different types of vocal components and instrumental components (e.g., bass, drums, guitar, etc.), to allow for more divergent training schemes for the models in the future. In one example embodiment herein, the neural network model operates exclusively on the magnitude of audio spectrograms. The audio signal for an individual (vocal/instrumental) component is rendered by constructing a spectrogram, wherein the output magnitude is given by applying a mask predicted by the U-Net to the magnitude of the original spectrum, while the output phase is that of the original spectrum, unaltered. Experimental results presented below indicate that such a simple methodology proves effective.
Dataset
According to an example aspect herein, the model architecture can employ training data available in the form of a triplet (original signal, vocal component, instrumental component), and the instrumental component may include one or more types of instrumental components (e.g., guitar content, bass content, drum content, piano content, and/or any other type of non-vocal content). However, in the event that vast amounts of unmixed multi-track recordings are not available, an alternative strategy according to an example aspect herein can be employed to mine for matching or candidate pairs of tracks, to obtain training data. For example, it is not uncommon for artists to release instrumental versions of tracks along with the original mix. An instrumental track may include one or more of various types of instrumental content, such as, by example and without limitation, guitar content, bass content, drum content, piano content, and/or any other type of non-vocal content (although in other example embodiments, an instrumental track may include, at least in part, vocal content processed through an instrument or synthesizer). In accordance with one example aspect herein, pairs of (original, instrumental) tracks from a large commercial music database are retrieved. Candidates are found by examining associated metadata for tracks with, in one example embodiment, matching duration and artist information, where the track title (fuzzily) matches except for the string “Instrumental” occurring in exactly one title in the pair. The pool of tracks is pruned by excluding exact content matches. In one example, such procedures are performed according to the technique described in Reference [10], which is incorporated by reference herein in its entirety, as if set forth fully herein. The approach enables a large source of X (mixed) and Yi (instrumental) magnitude spectrogram pairs to be provided. A vocal magnitude spectrogram Yv is obtained from their half-wave rectified difference. In one example, a final dataset included approximately 20,000 track pairs, resulting in almost two months-worth of continuous audio, which is perhaps the largest training data set ever applied to musical source separation. Table A below shows a relative distribution of frequent genres in the dataset, obtained from catalog metadata.
Selection of Matching Recordings
The manner in which candidate recording pairs are formed using a method according to an example embodiment herein will now be described, with reference to the flow diagram of
According to one example embodiment herein, the fuzzy matching is performed on track titles by first formatting them to a standardized format, by, for example, latinizing non-ASCII characters, removing parenthesized text, and then converting the result to lower-case text. In one example, this process yields about 164 k instrumental tracks, although this example is non-limiting. Also in one example embodiment herein, the method may provide a 1:n, n:n, or many-to-many mapping, in that an original song version may match to several different instrumentals or types of instrumental (e.g., one of a guitar track, a bass track, drum track, piano track, or the like) in step 206, and vice versa. Thus, although described herein in terms of an example case where tracks A and B can be matched, the invention is not so limited, and it is within the scope of the invention for more than two tracks to be matched together in step 206, and for more than two or a series of tracks to be matched in step 206. For example, multiple pairs or multiples series of tracks can be matched in that step.
In step 208, matching versions of a track, such as a pair of tracks (A, B) that were matched in step 206, are marked or otherwise designated (e.g., in a memory) as being either “instrumental” or “original”, based on whether or not the term “instrumental” (or an indication of instrument type) appears in the metadata associated with those tracks. In the present example wherein the metadata of track A does not indicate that it is an instrumental (and does not indicate an instrument type), and where the metadata of track B does indicate that track B is an instrumental (and/or indicates an instrument type), then the matching tracks (A, B) are marked as “(original, instrumental)”. In another example, in step 208, matching versions of a track, such as a pair of tracks (A, B) that were matched in step 206, are marked or otherwise designated (e.g., in a memory) as being either a specific type of instrumental (e.g., “guitar”, “bass”, “drums”, “piano” or the like) or “original”, based on whether or not a term representing the specific type of instrumental appears in the metadata associated with those tracks. In the present example wherein the metadata of track A does not indicate that it is an instrumental, and where the metadata of track B does indicate that track B is a type of instrumental, then the matching tracks (A, B) are marked as “(original, instrumental type)”, wherein “instrumental type” may indicate, in one example embodiment, “guitar instrumental”, “bass instrumental”, “drum instrumental”, “piano instrumental”, or any other type of instrument type, depending on the instrumental type indicated by the metadata of track B.
In one example embodiment herein, at least some of the results of step 206 can be evaluated manually (or automatically) to check for quality in step 210, since it may occur that some tracks were matched that should not have been matched. In general, such undesired matching can be a result of one or more errors, such as, for example, instrumental tracks appearing on multiple albums (such as compilations or movie soundtracks, where the explicit description of the track as “instrumental” may be warranted by the context). Pairs that are suspected of being incorrectly matched can be identified using a procedure according to an example aspect herein. For example an audio fingerprinting algorithm can be used to remove suspect pairs from the candidate set. In one example embodiment, that step is performed using an open-source fingerprinting algorithm, and the procedure described in Reference [34], can be employed although in other embodiments other types of algorithms can be employed. Reference [34] is hereby incorporated by reference in its entirety, as if set forth fully herein.
In one example embodiment, step 210 is performed according to procedure 300 illustrated in
If, on the other hand, the comparison performed in step 304 results in a determination that the code sequences do not overlap within the predetermined range of acceptability (“No” in step 306), then in step 310 the tracks are determined to be matched incorrectly, and thus at least one of them is removed from the results (step 312), and only those that remain are deemed to be correctly matched (step 308). A determination of “No” in step 306 may be a result of, for example, the codes not overlapping enough (e.g., owing to an erroneous fuzzy metadata match), or the codes overlapping too much (i.e., beyond the predetermined range of acceptability), which may occur in cases where, for example, the tracks are identical (e.g., the tracks are both instrumental, or are both of a same instrument type, or are both vocal or of the same vocal type).
The performance of step 312 may result in the removal of both tracks A and B, in certain situations. However, in the case for a 1:n, n:n, or many-to-many matching in earlier step 206, then only those tracks B which were determined to be matched with track A incorrectly are removed in step 312. In one example embodiment herein, step 312 is performed so that each original track is linked to only one non-redundant, instrumental track. The result of the performance of step 312 in that embodiment is that only pair(s) of tracks A, B deemed to match within the predetermined range of acceptability remain (step 308).
In one sample case where 10 million commercially available tracks are evaluated using the procedures 200 and 300, the processes yielded roughly 24,000 tracks, or 12,000 original-instrumental pairs, totaling about 1500 hours of audio track durations. 24,000 strongly labeled tracks were obtained for use as a training dataset.
Estimation of Vocal Activity
Before describing how matches tracks A, B are employed for training according to an example aspect herein, the manner in which vocal or non-vocal activity can be separated from a track and/or predicted, according to an example aspect herein, will first be described.
At step 404, the pair of TFRs (XO, XI) obtained in step 402 undergoes a conversion (by polar coordinate converter 604) to polar coordinates including magnitude and phase components, representing a frequency intensity at different points in time. The conversion produces corresponding spectrogram components (ZO, ZI), wherein the components (ZO, ZI) are a version of the pair of TFRs (XO, XI) that has been converted in step 404 into a magnitude and phase representation of the pair of TFRs, and define intensity of frequency at different points in time. The magnitude is the absolute value of a complex number, and the phase is the angle of the complex number. In step 405, patches are extracted from the spectrogram components (ZO, ZI) using patch extractor 606. In one example embodiment herein, step 405 results in slices of the spectrograms from step 404 (by way of polar coordinate converter 604) being obtained along a time axis, wherein the slices are fixed sized images (such as, e.g., 512 bins and 128 frames), according to one non-limiting and non-exclusive example embodiment herein. Patches can be obtained based on the magnitude of components (ZO, ZI) (wherein such patches also are hereinafter referred to as “magnitude patches (MPO,MPI)” or ““magnitude spectrogram patches (MPO,MPI)”)). In one example, step 405 is performed according to the procedures described in the Reference [38], which is incorporated by reference herein in its entirety, as if set forth fully herein.
In a next step 406, the magnitude patch)(MPO) (e.g., the original mix spectrogram magnitude) obtained in step 405 is applied to a pre-trained network architecture 500, wherein, according to one example aspect herein, the network architecture is a U-Net architecture. For purposes of this description, reference numeral “500” is used in conjunction with the terms “architecture”, “network architecture”, “U-Net”, “U-Net architecture”, and the like, each of which terms is used interchangeably herein to refer to the same element. Also, for purposes of the present description of
Also in one example embodiment herein, each layer of path 502 includes a strided 2D convolution of stride 2 and kernel size 5×5, batch normalization, and leaky rectified linear units (ReLU) with leakiness 0.2. The layers of path 504 employ strided deconvolution (also referred to as “transposed convolution”) with stride 2 and kernel size 5×5, batch normalization, plain ReLU, and a 50% dropout (in the first three layers). In at least the final layer (e.g., layer 504n), a sigmoid activation function can be employed, in one example embodiment herein.
Each downsampling layer 502b to 502n reduces in half the number of bins and frames, while increasing the number of feature channels. For example, where the input image of layer 502a is a 512×128×1 image slice (where 512 represents the number of bins, 128 represents the number of frames, and 1 represents the number of channels), application of that image slice to layer 502b results in a 256×64×16 image slice. Application of that 256×64×16 image slice to layer 502c results in a 128×32×32 image slice, and application of the 128×32×32 image slice to subsequent layer 502d results in a 64×16×64 image slice. Similarly, application of the 64×16×64 image slice to subsequent layer 502e results in a 32×8×128 image slice, and application of the 32×8×128 image slice to layer 502n results in a 16×4×256 image slice. Similarly, application of the 64×4×256 image slice to layer 506 results in a 8×2×512 image slice. Of course, the foregoing values are examples only, and the scope of the invention is not limited thereto.
Each layer in the expansive path 504 upsamples the (feature map) input received thereby followed by a 2×2 convolution (“up-convolution”) that doubles the number of bins and frames, while reducing the number of channels. Also, a concatenation with the correspondingly cropped feature map from the contracting path is provided, and two 3×3 convolutions, each followed by a ReLU.
In an example aspect herein, concatenations are provided by connections between corresponding layers of the paths 502 and 504, to concatenate post-convoluted channels to the layers in path 504. This feature is because, in at least some cases, when an image slice is provided through the path 504, at least some details of the image may be lost. As such, predetermined features 510 (such as, e.g., features which preferably are relatively unaffected by non-linear transforms) from each post-convolution image slice in the path 502 are provided to the corresponding layer of path 504, where the predetermined features are employed along with the image slice received from a previous layer in the path 504 to generate the corresponding expanded image slice for the applicable layer, and the predetermined features also are referred to herein as “concatenation features 510”. More particularly, in the illustrated embodiment, the 8×2×512 image slice obtained from layer 506, and concatenation features 510 from layer 502n, are applied to the layer 504a, resulting in a 16×4×256 image slice being provided, which is then applied along with concatenation features 510 from layer 502e to layer 504b, resulting in a 32×8×128 image slice being provided. Application of that 32×8×128 image slice, along with concatenation features 510 from layer 502d, to layer 504c results in a 64×16×64 image slice, which is then applied along with concatenation features 510 from layer 502c to layer 504d, resulting in a 128×32×32 image slice being provided. That latter image slice is then applied, along with concatenation features 510 from layer 502b, to layer 504e, resulting in a 256×16×16 image slice being provided, which, after being applied to layer 504n, results in a 512×128×1 image slice being provided. In one example embodiment herein, cropping may be performed to compensate for any loss of border pixels in every convolution.
Having described the architecture 500 of
The result of step 408 is then applied in step 410 to an inverse Short Time Fourier Transform (ISTFT) component 610 to transform (by way of a ISTFT) the result of step 408 from the frequency domain, into an audio signal in the time domain (step 410). In a present example where it is assumed that the architecture 500 is trained to learn/predict instrumental components of input signals (i.e., the mixed original signal, represented by the component MPO applied in step 406), the audio signal resulting from step 410 is an estimated instrumental audio signal. For example, the estimated instrumental audio signal represents an estimate of the instrumental portion of the mixed original signal first applied to the system 600 in step 402. In the foregoing manner, the instrumental component of a mixed original signal that includes both vocal and instrumental components can be obtained/predicted/isolated.
To obtain the vocal component of the mixed original signal, a method according to the foregoing procedure 400 is performed using system 600, but for a case where the architecture 500 is trained (e.g., in a manner as will be described later) for learn/predict vocal components of mixed signals. For example, the procedure for obtaining the vocal component includes performing steps 402 to 410 in the manner described above, except that, in one example embodiment, the U-Net architecture 500 employed in step 406 has been trained for estimating a vocal component of mixed original signals applied to the system 600. As a result of the performance of procedure 400 for such a case, the spectrogram 508 obtained in step 408 is an estimated vocal magnitude spectrum, and the audio signal obtained in step 410 is an estimated vocal audio signal, which represents an estimate of the vocal component of the mixed original signal applied to system 600 in step 402 (and an estimate of the component MPO applied to the architecture 500 in step 406).
Dataset
In one example embodiment herein, the model architecture assumes that training data is available in the form of a triplet (mixed original signal, vocal component, instrumental component), as would be the case in which, for example, access is available to vast amounts of unmixed multi-track recordings. In other example embodiments herein, an alternative strategy is provided to provide data for training a model. For example, one example solution exploits a specific but large set of commercially available recordings in order to “construct” training data: instrumental versions of recordings. Indeed, in one example embodiment, the training data is obtained in the manner described above in connection with
Training
In one example embodiment herein, the model herein can be trained using an ADAM optimizer. One example of an ADAM optimizer that can be employed is described in Reference [12], which is incorporated by reference herein in its entirety, as if set forth fully herein, although this example is non-limiting and non-exclusive. Given the heavy computational requirements of training such a model, in one example embodiment herein, input audio is downsampled to 8192 Hz in order to speed up processing. Then, a Short Time Fourier Transform is computed with a window size of 1024 and a hop length of 768 frames, and patches of, e.g., 128 frames (roughly 11 seconds) are extracted, which then are fed as input and targets to the architecture 500. Also in this example embodiment, the magnitude spectrograms are normalized to the range [0, 1]. Of course, these examples are non-exclusive and non-limiting.
The manner in which training is performed, according to an example embodiment herein, will now be described in greater detail, with reference to
In one example embodiment herein, in step 702 the system 600 of
L(X,Y;Θ)=∥f(X,Θ)⊗X−Y∥ (F1)
where X denotes the magnitude of the spectrogram of the original, mixed signal (e.g., including both vocal and instrumental components), Y denotes the magnitude of the spectrogram of the target instrumental (or vocal, where a vocal signal is used instead) audio (wherein Y may be further represented by either Yv for a vocal component or Yi for an instrumental component of the input signal), f(X, Θ) represents an output of mask combiner 608, and Θ represents the U-Net (or parameters thereof). For the case where the U-Net is trained to predict instrumental spectrograms, denotation Θ may be further represented by Θi (whereas for the case where the U-Net is trained to predict vocal spectrograms, denotation Θ may be further represented by Θv). In the above formula F1, the expression f(X, Θ)⊗X represents masking of the magnitude X (by mask combiner 608) using the version of the magnitude X after being applied to the U-Net 500.
A result of formula F1 is provided from loss calculator 612 to parameter adaptor 614, which, based on the result, varies one or more parameters of the architecture 500, if needed, to reduce the loss value (represented by L(X, Y; Θ)) (step 706). Procedure 700 can be performed again in as many iterations as needed to substantially reduce or minimize the loss value, in which case the architecture 500 is deemed trained. For example, in step 708 it is determined whether the loss value is sufficiently minimized. If “yes” in step 708, then the method ends at step 710 and the architecture is deemed trained. If “no” in step 708, then control passes back to step 702 where the procedure 700 is performed again as many times as needed until the loss value is deemed sufficiently minimized.
The manner in which the parameter adapter 614 varies the parameters of the architecture 500 in step 706 can be in accordance with any suitable technique, such as, by example and without limitation, that disclosed in Reference [36], which is incorporated by reference herein in its entirety, as if set forth fully herein. In one example embodiment, step 706 may involve altering one or more weights, kernels, and/or other applicable parameter values of the architecture 500, and can include performing a stochastic gradient descent algorithm.
A case where it is desired to train the architecture 500 to predict a vocal component of a mixed original signal will now be described. In this example embodiment, the procedure 700 is performed in the same manner as described above, except that the signal Y provided to the loss calculator 612 is a target vocal signal corresponding to the mixed original signal(s) φ (track(s) A) input to the system 650 (i.e., the target vocal signal and mixed original signal are deemed to be a match). The target vocal signal may be obtained from a database of such signals, if available (and a magnitude of the spectrogram thereof can be employed). In other example embodiments, and referring to the procedure 800 of
The above example embodiments are described in the context of separating/isolating and/or predicting a vocal or instrumental component from a mixed, original signal (mixed music track), In another example embodiment herein, instead of separating/isolating and/or predicting a vocal or instrumental component from a mixed, original signal, in another example embodiment herein, the above method(s), architecture 500, and systems (
In a case where it is desired to train the architecture 500 (and system 600) to estimate various types of vocal components, the procedure 700 is performed using system 600, but using target vocal signals (as signal Y) corresponding to the different types of vocal components (in one example embodiment herein, the target signals may be obtained from a database of such signals, or can be obtained in another manner as described above). As but one non-limiting example, the procedure 700 can be performed as described above but 1) for a case where a first target vocal signal Y is a vocal component from singer A, 2) for another case where a second target vocal signal Y is a vocal component from singer B, 3) for another case where a third target vocal signal Y is a vocal component from singer C, and/or the like. As a result, the architecture 500 can become trained to estimate/predict/isolate different types of vocal components from different singers, in one non-limited example herein.
Estimation of (Multiple) Instrument and/or Vocal Content
The above example embodiments are described in the context of separating/isolating and/or predicting a vocal or instrumental component from a mixed original signal (mixed music track), In one example embodiment herein, the instrumental component may include instrumental content, such as, by example and without limitation, content of one or more types of instruments such as guitar content, bass content, drum content, piano content, or the like. In another, present example aspect herein, respective different types of instrument content, and/or vocal content, can be separated/isolated and/or predicted from a mixed original signal, according to another example embodiment herein. This example aspect will now be described, with reference to
Also in one example embodiment herein, each layer of path 1502 includes a strided 2D convolution of stride 2 and kernel size 5×5, batch normalization, and leaky rectified linear units (ReLU) with leakiness 0.2. The layers of each path 1504-1 to 1504-x employ strided deconvolution (also referred to as “transposed convolution”) with stride 2 and kernel size 5×5, batch normalization, plain ReLU, and a 50% dropout (in the first three layers). In at least the final layer (e.g., layer 1504n), a sigmoid activation function can be employed, in one example embodiment herein.
Each downsampling layer 1502b to 1502n reduces in half the number of bins and frames, while increasing the number of feature channels. For example, where the input image of layer 1502a is a 512×128×1 image slice (where 512 represents the number of bins, 128 represents the number of frames, and 1 represents the number of channels), application of that image slice to layer 1502b results in a 256×64×16 image slice. Application of that 256×64×16 image slice to layer 1502c results in a 128×32×32 image slice, and application of the 128×32×32 image slice to subsequent layer 1502d results in a 64×16×64 image slice. Similarly, application of the 64×16×64 image slice to subsequent layer 1502e results in a 32×8×128 image slice, and application of the 32×8×128 image slice to layer 1502n results in a 16×4×256 image slice. Of course, the foregoing values are examples only, and the scope of the invention is not limited thereto.
Each expansive path 1504-1 to 1504-x is pre-trained to learn/estimate/predict/isolate, and thus corresponds to, a particular type of component, in one example embodiment herein. As but one non-limiting example, and merely for purposes of illustration, the expansive paths 1504-1 to 1504-x are pre-trained to learn/estimate/predict/isolate vocals, guitar, bass, drums, and other (e.g., piano) instrument (content) components, respectively, although this example is not limiting, and the paths 1504-1 to 1504-x may correspond to other types of instrument components instead.
Each layer 1504a to 1504n in respective ones of the expansive paths 1504-1 to 1504-x upsamples the (feature map) input received thereby followed by a 2×2 convolution (“up-convolution”) that doubles the number of bins and frames, while reducing the number of channels. Also, a concatenation with the correspondingly cropped feature map from a corresponding layer of the contracting path can be provided, and two 3×3 convolutions, each followed by a ReLU.
In an example aspect herein, concatenations are provided by connections between corresponding layers of the path 1502 and paths 1504-1 to 1504-x, to concatenate post-convoluted channels to the layers in respective paths 1504-1 to 1504-x (as described above for
In a present example embodiment herein, the U-Net architecture that is included in the system 600 of
Reference will now be made to
Generally, the procedure 1600 according to the present example aspect of the present application includes computing a Time-Frequency Representation (TFR) for the tracks TO and TI, using TFR obtainer 602, to yield corresponding TFRs XO and XI, respectively, in the frequency domain (step 1602), wherein the TFRs XO and X each are a spectrogram of 2D coefficients, having frequency and phase content, and then performing steps 1604 to 1610 as will be described below. It should be noted that, although described herein in the context of steps 1602 to 1605 being performed together for both types of tracks TO and TI (i.e., an “original” track and an “instrumental” track), the scope of the invention herein is not so limited, and in other example embodiments herein, those steps 1602 to 1605 may be performed separately for each separate type of track. In other example embodiments, steps 1602 to 1605 are performed for the “original” (“mixed”) track, such as, for example, in a case where it is desired to predict or isolate various instrumental and/or vocal components of the track, and steps 1602 to 1605 are performed separately for at least one instrumental track, for use in training (to be described below) to enable the prediction/isolation to occur. Thus, during real-time application, procedure 1600 need not be performed for the track TI. In one example, step 1602 is performed according to the procedures described in Reference [39], which is incorporated by reference herein in its entirety, as if set forth fully herein.
At step 1604, the pair of TFRs (XO, XI) obtained in step 1602 undergoes a conversion (by polar coordinate converter 604) to polar coordinates including magnitude and phase components, representing a frequency intensity at different points in time. The conversion produces corresponding spectrogram components (ZO, ZI), wherein the components (ZO, ZI) are a version of the pair of TFRs (XO, XI) that has been converted in step 1604 into a magnitude and phase representation of the pair of TFRs, and define intensity of frequency at different points in time. In step 1605, patches are extracted from the spectrogram components (ZO, ZI) using patch extractor 606. In one example embodiment herein, step 1605 results in slices of the spectrograms from step 1604 (by way of polar coordinate converter 604) being obtained along a time axis, wherein the slices are fixed sized images (such as, e.g., 512 bins and 128 frames), according to one non-limiting and non-exclusive example embodiment herein. Patches can be obtained based on the magnitude of components (ZO, ZI) (wherein such patches also are hereinafter referred to as “magnitude patches (MPO,MPI)” or ““magnitude spectrogram patches (MPO,MPI)”)). In one example, step 1605 is performed according to the procedures described in the Reference [38], which is incorporated by reference herein in its entirety, as if set forth fully herein.
In a next step 1606, the magnitude patch)(MPO) (e.g., the original mix spectrogram magnitude) obtained in step 1605 is applied to a pre-trained network architecture, wherein, according to one example aspect herein, the network architecture is the U-Net architecture 1500. For purposes of the present description of
In step 1608, the output of layer 1504n of each of the expansive paths 1504-1 to 1504-x is employed as a mask for being applied by mask combiner 608 to the input image of layer 1502a, to provide a corresponding estimated magnitude spectrogram 1508 (also referred to as “layer 1508”). That step 1608 is performed to combine the image (e.g., preferably a magnitude component) from layer 1504n of each respective expansive path 1504-1 to 1504-x with the phase component from the mixed original spectrogram (of input layer 1502a) to provide a corresponding complex value spectrogram having both phase and magnitude components (i.e., to render independent of the amplitude of the original spectrogram). Step 1608 may be performed in accordance with any suitable technique.
In the example scenario in which expansive path 1504-1 is pre-trained to learn/estimate/predict/isolate a vocal component, the corresponding estimated magnitude spectrogram 1508 obtained as a result of applying the output of layer 1504 of that path as a mask to the input image of layer 1502a in step 1608 is an estimated vocal magnitude spectrogram. In the example scenario in which expansive path 1504-2 is pre-trained to learn/estimate/predict/isolate a guitar component, the corresponding estimated magnitude spectrogram 1508 obtained as a result of applying the output of layer 1504 of that path as a mask to the input image of layer 1502a in step 1608 is an estimated guitar magnitude spectrogram. In the example scenario in which expansive path 1504-3 is pre-trained to learn/estimate/predict/isolate a bass component, the corresponding estimated magnitude spectrogram 1508 obtained as a result of applying the output of layer 1504 of that path as a mask to the input image of layer 1502a in step 1608 is an estimated bass magnitude spectrogram. In the example scenario in which expansive path 1504-3 is pre-trained to learn/estimate/predict/isolate a drum component, the corresponding estimated magnitude spectrogram 1508 obtained as a result of applying the output of layer 1504 of that path as a mask to the input image of layer 1502a in step 1608 is an estimated drum magnitude spectrogram. In the example scenario in which expansive path 1504-x is pre-trained to learn/estimate/predict/isolate a piano component, the corresponding estimated magnitude spectrogram 1508 obtained as a result of applying the output of layer 1504 of that path as a mask to the input image of layer 1502a in step 1608 is an estimated piano magnitude spectrogram.
Each result of step 1608 is then applied in step 1610 to an inverse Short Time Fourier Transform (ISTFT) component 610 to transform (by way of a ISTFT) the result of step 1608 from the frequency domain, into a corresponding audio signal in the time domain (step 1610).
For example, application of the estimated vocal magnitude spectrogram obtained in step 1608 to the ISTFT component 610 in step 1610 results in an estimated vocal audio signal, application of the estimated guitar magnitude spectrogram obtained in step 1608 to the ISTFT component 610 in step 1610 results in an estimated guitar audio signal, and application of the estimated bass magnitude spectrogram obtained in step 1608 to the ISTFT component 610 in step 1610 results in an estimated bass audio signal. Similarly by example, application of the estimated drum magnitude spectrogram obtained in step 1608 to the ISTFT component 610 in step 1610 results in an estimated drum audio signal, application of the estimated piano magnitude spectrogram obtained in step 1608 to the ISTFT component 610 in step 1610 results in an estimated piano audio signal, etc.
The estimated vocal audio signal(s) represent estimate(s) of the vocal portion(s) of the mixed original signal first applied to the system 600 in step 1602, and the estimated type(s) of instrumental audio signal(s) represent estimate(s) of the corresponding type(s) of instrumental portion(s) of the mixed original signal first applied to the system 600 in step 1602.
In the foregoing manner, the vocal and/or instrumental component(s) of a mixed original signal that includes both vocal and instrumental type components can be estimated/obtained/predicted/isolated.
It should be noted that, although for convenience the method 1600 is described as being performed simultaneously for each expansive path 1504-1 to 1504-x, the scope of the invention is not necessarily limited thereto, and, it is within the scope of the invention to perform method 1600 separately for each such path 1504-1 to 1504-x, or for more or less than the number of expansive paths 1504-1 to 1504-x and type(s) of vocal or instrumental components shown/described, depending on, for example, the application of interest, and the type(s) of component(s) the architecture 1500 is trained to learn/predict.
Also, although described above in the context of estimating a vocal component and various types of instrumental components, it is within the scope of the present invention to estimate various types of vocal components as well. For example, the mixed original signal may include more than a single vocal component (e.g., and vocal components from different singers having differing vocal/audio characteristics). To estimate the various types of vocal components, the foregoing procedure 1600 is performed using system 600, but for a case where the architecture 1500 is pre-trained (e.g., as described below in connection with
Training in Multiple Instrument and/or Vocal Content Embodiment
In one example of the present embodiment herein, a model of the present embodiment can be trained using an ADAM optimizer. One example of an ADAM optimizer that can be employed is described in Reference [12], which is incorporated by reference herein in its entirety, as if set forth fully herein, although this example is non-limiting and non-exclusive. Given the heavy computational requirements of training such a model, in one example embodiment herein, input audio is downsampled to 8192 Hz in order to speed up processing. Then, a Short Time Fourier Transform is computed with a window size of 1024 and a hop length of 768 frames, and patches of, e.g., 128 frames (roughly 11 seconds) are extracted, which then are fed as input and targets to the architecture 1500. Also in this example embodiment, the magnitude spectrograms are normalized to the range [0, 1]. Of course, these examples are non-exclusive and non-limiting.
The manner in which training is performed, according to an example of the present embodiment herein, will now be described in greater detail, with reference to
In one example embodiment herein, in step 1702 the system 650 of
For each instance of the signal φ applied to the system 600 of
In step 1704 the loss calculator 612 employs a loss function to determine how much difference there is between corresponding ones of the outputs f (X, Θ) (of mask combiner 608) and respective target signals Y. For example, in one example embodiment the loss calculator 612 can employ: a loss function to determine how much difference there is between the output f (X, Θ) associated with path 1504-1 and the target “vocal” signal Y; a loss function to determine how much difference there is between the output f(X, Θ) associated with path 1504-2 and the target “guitar” signal Y; a loss function to determine how much difference there is between the output f(X, Θ) associated with path 1504-3 and the target “bass” signal Y; etc. As described above, in one example embodiment herein, each loss function is the L1,1 norm (e.g., wherein the norm of a matrix is the sum of the absolute values of its elements) of a difference between a target spectrogram and the corresponding masked input spectrogram, as represented by formula (F1):
L(X,Y;Θ)=∥f(X,Θ)⊗X−Y∥ (F1)
where X denotes the magnitude of the spectrogram of the original, mixed signal (e.g., including both vocal and instrumental components), Y denotes the magnitude of the spectrogram of the corresponding target signal (wherein Y may be further represented by either Yv for a vocal component or Yi for an instrumental component of the input signal), f(X, Θ) represents a corresponding output of mask combiner 608, and Θ represents the U-Net (or parameters thereof). In the above formula F1, the expression f(X, Θ)⊗X represents masking of the corresponding magnitude X (by mask combiner 608) using the corresponding version of the magnitude X after being applied to the U-Net architecture 1500.
In one example embodiment herein, a result of the performance of each formula F1 is provided from loss calculator 612 to parameter adaptor 614, which, based on the result, varies one or more parameters of the architecture 1500, if needed for each case, to reduce the corresponding loss value(s) (represented by L(X, Y; Θ)) (step 1706). Procedure 1700 can be performed again in as many iterations as needed to substantially reduce or minimize one or more of the respective loss value(s), as needed, in which case (of substantial reduction or minimization) the architecture 1500 is deemed trained to learn the specific type of component type(s) corresponding to the corresponding target audio signal(s) Y used in respective performance(s) of the formula(s). For example, in step 1708 it is determined whether each or respective ones of the loss values is/are sufficiently minimized. If “yes” in step 1708, then control passes to step 1709 to be described below, and the architecture 1500 is deemed trained to learn the particular type(s) of component type(s) corresponding to the target signals Y. If “no” in step 1708, then control passes back to step 1702 where the procedure 1700 is performed again as many times, for one or more respective loss value(s), as needed until the corresponding loss value(s) are deemed sufficiently minimized (in step 708).
An indication of “yes” in step 1708 also indicates that one or more predetermined paths 1504-1 to 1504-x of the architecture 1500 (and system 600) are trained to predict/estimate/isolate predetermined type(s) of component(s) (e.g., a component of the same type as signal(s) Y applied to loss calculator 612). By example, based on training as described in the above examples, the result is that the architecture 1500 (and systems 600 and 650) is trained such that path 1504-1 can predict/estimate/isolate vocal content of mixed original signals φ, path 1504-2 can predict/estimate/isolate guitar content of mixed original signals φ, path 1504-3 can predict/estimate/isolate bass content of mixed original signals φ, path 1504-4 can predict/estimate/isolate drum content of mixed original signals φ, and path 1504-x can predict/estimate/isolate other instrument content (e.g., piano) of mixed original signals φ. In step 1709 it is determined whether an additional one of the expansive paths 1504-1 to 1504-x will be trained. If “yes” in step 1709, then the procedure 1700 returns back to step 1702 where it begins again, but also for the additional one of the expansive paths 1504-1 to 1504-x. If “no” in step 1709, then the procedure ends in step 1710.
It should be noted that the manner in which the parameter adapter 614 varies the parameters of the architecture 1500 in step 1706 can be in accordance with any suitable technique, such as, by example and without limitation, that disclosed in Reference [36], which is incorporated by reference herein in its entirety, as if set forth fully herein. In one example embodiment, step 1706 may involve altering one or more weights, kernels, and/or other applicable parameter values of the architecture 1500, and can include performing a stochastic gradient descent algorithm.
It should be noted that each respective type of target signal Y may be obtained from a database of such signals, if available (and a magnitude of the spectrogram thereof can be employed). In other example embodiments, each target signal Y may be obtained as described above in conjunction with procedure 800 of
In a case where it is desired to train the architecture 1500 (and system 600) to estimate various types of vocal components, the above-described procedure 1700 is performed using system of
Quantitative Evaluation
To provide a quantitative evaluation, an example embodiment herein is compared to the Chimera model (see, e.g., Reference[15]) that produced the highest evaluation scores in a 2016 MIREX Source Separation campaign. A web interface can be used to process audio clips. It should be noted that the Chimera web server runs an improved version of the algorithm that participated in MIREX, using a hybrid “multiple heads” architecture that combines deep clustering with a conventional neural network (see, e.g., Reference [16]).
For evaluation purposes an additional baseline model was built, resembling the U-Net model but without skip connections, essentially creating a convolutional encoder-decoder, similar to the “Deconvnet” (see, e.g., Reference[19]).
The three models were evaluated on the standard iKala (see, e.g., Reference [5]) and MedleyDB dataset (see, e.g., Reference [3]). The iKala dataset has been used as a standardized evaluation for the annual MIREX campaign for several years, so there are many existing results that can be used for comparison. MedleyDB on the other hand was recently proposed as a higher-quality, commercial-grade set of multi-track stems.
Isolated instrumental and vocal tracks were generated by weighting sums of instrumental/vocal stems by their respective mixing coefficients as supplied by a MedleyDB Python API. The evaluation is limited to clips that are known to contain vocals, using the melody transcriptions provided in both iKala and MedleyDB.
The following functions are used to measure performance: Signal-To-Distortion Ratio (SDR), Signal-to-Interference Ratio (SIR), and Signal-to-Artifact Ratio (SAR) (see, e.g., Reference [31]). Normalized SDR (NSDR) is defined as
NSDR(Se,Sr,Sm)=SDR(Se,Sr)−SDR(Sm,Sr) (F2)
where Se is the estimated isolated signal, Sr is the reference isolated signal, and Sm is the mixed signal. Performance measures are computed using the mireval toolkit (see, e.g., Reference [22]).
Table 2 and Table 3 show that the U-Net significantly outperforms both the baseline model and Chimera on all three performance measures for both datasets.
Assuming that the distribution of tracks in the iKala hold-out set used for MIREX evaluations matches those in the public iKala set, results of an example embodiment herein are compared to the participants in the 2016 MIREX Singing Voice Separation task. Table 4 and Table 5 show NSDR scores for the example models herein compared to the best performing algorithms of the 2016 MIREX campaign.
In order to assess the effect of the U-Net's skip connections, masks generated by the U-Net and baseline models can be visualized. From
Subjective Evaluation
Emiya et al. introduced a protocol for the subjective evaluation of source separation algorithms (see, e.g., Reference [7]). They suggest asking human subjects four questions that broadly correspond to the SDR/SIR/SAR measures, plus an additional question regarding the overall sound quality.
These four questions were asked to subjects without music training, and the subjects found them ambiguous, e.g., they had problems discerning between the absence of artifacts and general sound quality. For better clarity, the survey was distilled into the following two questions in the vocal extraction case:
For instrumental extraction similar questions were asked:
Data was collected using CrowdFlower, an online platform where humans carry out micro-tasks, such as image classification, simple web searches, etc., in return for small per-task payments.
In the survey, CrowdFlower users were asked to listen to three clips of isolated audio, generated by U-Net, the baseline model, and Chimera. The order of the three clips was randomized. Each question asked one of the Quality and Interference questions. In an Interference question a reference clip was included. The answers were given according to a 7 step Likert scale (see, e.g., Reference [13]), ranging from “Poor” to “Perfect”.
To ensure the quality of the collected responses, the survey was interspersed with “control questions” that the user had to answer correctly according to a predefined set of acceptable answers on the Likert scale. Users of the platform were unaware of which questions are control questions. If questions were answered incorrectly, the user was disqualified from the task. A music expert external to the research group was asked to provide acceptable answers to a number of random clips that were designated as control questions.
For the survey 25 clips from the iKala dataset and 42 clips from MedleyDB were used. There were 44 respondents and 724 total responses for the instrumental test, and 55 respondents supplied 779 responses for the voice test.
The example embodiments herein take advantage of a U-Net architecture in the context of singing voice separation, and, as can be seen, provide clear improvements over existing systems. The benefits of low-level skip connections were demonstrated by comparison to plain convolutional encoder-decoders.
The example embodiments herein also relate to an approach to mining strongly labeled data from web-scale music collections for detecting vocal activity in music audio. This is achieved by automatically pairing original recordings, containing vocals, with their instrumental counterparts, and using such information to train the U-Net architecture to estimate vocal or instrumental components of a mixed signal.
Acoustic attribute computation system also referred to herein as a “computer”) 1100 may include without limitation a processor device 1110, a main memory 1125, and an interconnect bus 1105. The processor device 1110 (410) may include without limitation a single microprocessor, or may include a plurality of microprocessors for configuring the system 1100 as a multi-processor acoustic attribute computation system. The main memory 1125 stores, among other things, instructions and/or data for execution by the processor device 1110. The main memory 1125 may include banks of dynamic random access memory (DRAM), as well as cache memory.
The system 1100 may further include a mass storage device 1130, peripheral device(s) 1140, portable non-transitory storage medium device(s) 1150, input control device(s) 1180, a graphics subsystem 1160, and/or an output display interface (also referred to herein as an “output display”) 1170. A digital signal processor (DSP) 1182 may also be included to perform audio signal processing. For explanatory purposes, all components in the system 1100 are shown in
Mass storage device 1130 additionally stores a neural network system engine (such as, e.g., a U-Net network engine) 1188 that is trainable to predict an estimate vocal and/or instrumental component(s) of a mixed original signal, a comparing engine 1190 for comparing an output of the neural network system engine 1188 to a target instrumental or vocal signal to determine a loss, and a parameter adjustment engine 1194 for adapting one or more parameters of the neural network system engine 1188 to minimize the loss. A machine learning engine 1195 provides training data, and an attenuator/volume controller 1196 enables control of the volume of one or more tracks, including inverse proportional control of simultaneously played tracks.
The portable storage medium device 1150 operates in conjunction with a nonvolatile portable storage medium, such as, for example, a solid state drive (SSD), to input and output data and code to and from the system 1100. In some embodiments, the software for storing information may be stored on a portable storage medium, and may be inputted into the system 1100 via the portable storage medium device 1150. The peripheral device(s) 1140 may include any type of computer support device, such as, for example, an input/output (I/O) interface configured to add additional functionality to the system 1100. For example, the peripheral device(s) 1140 may include a network interface card for interfacing the system 1100 with a network 1120.
The input control device(s) 1180 provide a portion of the user interface for a user of the computer 1100. The input control device(s) 1180 may include a keypad and/or a cursor control device. The keypad may be configured for inputting alphanumeric characters and/or other key information. The cursor control device may include, for example, a handheld controller or mouse, a trackball, a stylus, and/or cursor direction keys. In order to display textual and graphical information, the system 1100 may include the graphics subsystem 1160 and the output display 1170. The output display 1170 may include a display such as a CSTN (Color Super Twisted Nematic), TFT (Thin Film Transistor), TFD (Thin Film Diode), OLED (Organic Light-Emitting Diode), AMOLED display (Activematrix Organic Light-emitting Diode), and/or liquid crystal display (LCD)-type displays. The displays can also be touchscreen displays, such as capacitive and resistive-type touchscreen displays. The graphics subsystem 1160 receives textual and graphical information, and processes the information for output to the output display 1170.
The user interface 1400 also includes forward control 1406 and reverse control 1404 for scrolling through a track in either respective direction, temporally. According to an example aspect herein, the user interface 1400 further includes a volume control bar 1408 having a volume control 1409 (also referred to herein as a “karaoke slider”) that is operable by a user for attenuating the volume of at least one track. By example, assume that the play button 1402 is selected to playback a song called “Night”. According to one non-limiting example aspect herein, when the play button 1402 is selected, the “mixed” original track of the song, and the corresponding instrumental track of the same song (i.e., wherein the tracks may be identified as being a pair according to procedures described above), are retrieved from the mass storage device 1130, wherein, in one example, the instrumental version is obtained according to one or more procedures described above, such as that shown in
Also according to one example embodiment herein, the more the volume control 1409 is moved in a rightward direction along the bar 1408, the greater is the volume of the instrumental track and the lesser is the volume of the “mixed” original track. By example, when the volume control 1409 is positioned precisely in the middle between the center position 1410 and rightmost end 1414 of the bar 1408, then the volume of the “mixed” original track is played back at 25% of the predetermined maximum volume, and the instrumental track is played back at 75% of the predetermined maximum volume. When the volume control 1409 is positioned all the way to the right along the bar 1408, at the rightmost end 1414, then the volume of the “mixed” original track is played back at 0% of the predetermined maximum volume, and the instrumental track is played back at 100% of the predetermined maximum volume.
In the above manner, a user can control the proportion of the volume levels between the “mixed” original track and the corresponding instrumental track.
Of course, the above example is non-limiting. By example, according to another example embodiment herein, when the play button 1402 is selected, the “mixed” original track of the song, as well as the vocal track of the same song (i.e., wherein the tracks may be identified as being a pair according to procedures described above), can be retrieved from the mass storage device 1130, wherein, in one example, the vocal track is obtained according to one or more procedures described above, such as that shown in
In still another example embodiment herein, when the play button 1402 is selected to play back a song, the instrumental track of the song, as well as the vocal track of the same song (wherein the tracks are recognized to be a pair) are retrieved from the mass storage device 1130, wherein, in one example, the tracks are each obtained according to one or more procedures described above, such as that shown in
Of course, the above-described directionalities of the volume control 1409 are merely representative in nature, and, in other example embodiments herein, movement of the volume control 1409 in a particular direction can control the volumes of the above-described tracks in an opposite manner than those described above, and/or the percentages described above may be different that those described above, in other example embodiments. Also, in one example embodiment herein, which particular type of combination of tracks (i.e., a mixed original signal paired with either a vocal or instrumental track, or paired vocal and instrumental tracks) is employed in the volume control technique described above can be predetermined according to pre-programming in the system 1100, or can be specified by the user by operating the user interface 1400.
Referring again to
Input control devices 1180 can control the operation and various functions of system 1100.
Input control devices 1180 can include any components, circuitry, or logic operative to drive the functionality of system 1100. For example, input control device(s) 1180 can include one or more processors acting under the control of an application.
Each component of system 1100 may represent a broad category of a computer component of a general and/or special purpose computer. Components of the system 1100 (400) are not limited to the specific implementations provided herein.
Software embodiments of the examples presented herein may be provided as a computer program product, or software, that may include an article of manufacture on a machine-accessible or machine-readable medium having instructions. The instructions on the non-transitory machine-accessible machine-readable or computer-readable medium may be used to program a computer system or other electronic device. The machine- or computer-readable medium may include, but is not limited to, floppy diskettes, optical disks, and magneto-optical disks or other types of media/machine-readable medium suitable for storing or transmitting electronic instructions. The techniques described herein are not limited to any particular software configuration. They may find applicability in any computing or processing environment. The terms “computer-readable”, “machine-accessible medium” or “machine-readable medium” used herein shall include any medium that is capable of storing, encoding, or transmitting a sequence of instructions for execution by the machine and that causes the machine to perform any one of the methods described herein. Furthermore, it is common in the art to speak of software, in one form or another (e.g., program, procedure, process, application, module, unit, logic, and so on), as taking an action or causing a result. Such expressions are merely a shorthand way of stating that the execution of the software by a processing system causes the processor to perform an action to produce a result.
Some embodiments may also be implemented by the preparation of application-specific integrated circuits, field-programmable gate arrays, or by interconnecting an appropriate network of conventional component circuits.
Some embodiments include a computer program product. The computer program product may be a storage medium or media having instructions stored thereon or therein which can be used to control, or cause, a computer to perform any of the procedures of the example embodiments of the invention. The storage medium may include without limitation an optical disc, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory, a flash card, a magnetic card, an optical card, nanosystems, a molecular memory integrated circuit, a RAID, remote data storage/archive/warehousing, and/or any other type of device suitable for storing instructions and/or data.
Stored on any one of the computer-readable medium or media, some implementations include software for controlling both the hardware of the system and for enabling the system or microprocessor to interact with a human user or other mechanism utilizing the results of the example embodiments of the invention. Such software may include without limitation device drivers, operating systems, and user applications. Ultimately, such computer-readable media further include software for performing example aspects of the invention, as described above.
Included in the programming and/or software of the system are software modules for implementing the procedures described herein.
While various example embodiments of the present invention have been described above, it should be understood that they have been presented by way of example, and not limitation. It will be apparent to persons skilled in the relevant art(s) that various changes in form and detail can be made therein. Thus, the present invention should not be limited by any of the above described example embodiments, but should be defined only in accordance with the following claims and their equivalents.
In addition, it should be understood that the
Further, the purpose of the foregoing Abstract is to enable the U.S. Patent and Trademark Office and the public generally, and especially the scientists, engineers and practitioners in the art who are not familiar with patent or legal terms or phraseology, to determine quickly from a cursory inspection the nature and essence of the technical disclosure of the application. The Abstract is not intended to be limiting as to the scope of the example embodiments presented herein in any way. It is also to be understood that the procedures recited in the claims need not be performed in the order presented.
This application is a continuation-in-part of each of (1) U.S. application Ser. No. 16/055,870, filed Aug. 6, 2018, entitled “SINGING VOICE SEPARATION WITH DEEP U-NET CONVOLUTIONAL NETWORKS”, (2) U.S. application Ser. No. 16/242,525, filed Jan. 8, 2019, entitled “SINGING VOICE SEPARATION WITH DEEP U-NET CONVOLUTIONAL NETWORKS”, and (3) U.S. application Ser. No. 16/165,498, filed Oct. 19, 2018, entitled “SINGING VOICE SEPARATION WITH DEEP U-NET CONVOLUTIONAL NETWORKS”, each of which applications (1) to (3) is hereby incorporated by reference herein in its entirety, as if set forth fully herein.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16055870 | Aug 2018 | US |
| Child | 16521756 | US | |
| Parent | 16242525 | Jan 2019 | US |
| Child | 16055870 | US | |
| Parent | 16165498 | Oct 2018 | US |
| Child | 16242525 | US |
