1. Field of the Invention
The present invention relates generally to computational techniques including digital signal processing for automated processing of speech and, in particular, to techniques whereby a system or device may be programmed to automatically transform an input audio encoding of speech into an output encoding of song, rap or other expressive genre having meter or rhythm for audible rendering.
2. Description of the Related Art
The installed base of mobile phones and other handheld compute devices grows in sheer number and computational power each day. Hyper-ubiquitous and deeply entrenched in the lifestyles of people around the world, they transcend nearly every cultural and economic barrier. Computationally, the mobile phones of today offer speed and storage capabilities comparable to desktop computers from less than ten years ago, rendering them surprisingly suitable for real-time sound synthesis and other digital signal processing based transformations of audiovisual signals.
Indeed, modern mobile phones and handheld compute devices, including iOS™ devices such as the iPhone™, iPod Touch™ and iPad™ digital devices available from Apple Inc. as well as competitive devices that run the Android operating system, all tend to support audio and video playback and processing quite capably. These capabilities (including processor, memory and I/O facilities suitable for real-time digital signal processing, hardware and software CODECs, audiovisual APIs, etc.) have contributed to vibrant application and developer ecosystems. Examples in the music application space include the popular I Am T-Pain and Glee Karaoke social music apps available from Smule, Inc., which provide real-time continuous pitch correction of captured vocals, and the LaDiDa reverse karaoke app from Khush, Inc. which automatically composes music to accompany user vocals.
It has been discovered that captured vocals may be automatically transformed using advanced digital signal processing techniques that provide captivating applications, and even purpose-built devices, in which mere novice user-musicians may generate, audibly render and share musical performances. In some cases, the automated transformations allow spoken vocals to be segmented, arranged, temporally aligned with a target rhythm, meter or accompanying backing tracks and pitch corrected in accord with a score or note sequence. Speech-to-song music applications are one such example. In some cases, spoken vocals may be transformed in accord with musical genres such as rap using automated segmentation and temporal alignment techniques, often without pitch correction. Such applications, which may employ different signal processing and different automated transformations, may nonetheless be understood as speech-to-rap variations on the theme.
In speech-to-song and speech-to-rap applications (or purpose-built devices such as for toy or amusement markets), an automatic transformation of captured vocals is typically shaped by features (e.g., rhythm, meter, repeat/reprise organization) of a backing musical track with which the transformed vocals are eventually mixed for audible rendering. On the other hand, while mixing with a musical backing track is typical in many implementations of the invented techniques, in some cases, automated transforms of captured vocals may be adapted to provide expressive performances that are temporally aligned with a target rhythm or meter (such as a poem, iambic cycle, limerick, etc.) without musical accompaniment. These and other variations will be understood by persons of ordinary skill in the art who have access to the present disclosure and with reference to the claims that follow.
In some embodiments in accordance with the present invention, a computational method is implemented for transforming an input audio encoding of speech into an output that is rhythmically consistent with a target song. The method includes (i) segmenting the input audio encoding of the speech into plural segments, the segments corresponding to successive sequences of samples of the audio encoding and delimited by onsets identified therein; (ii) mapping individual ones of the plural segments to respective sub-phrase portions of a phrase template for the target song, the mapping establishing one or more phrase candidates; (iii) temporally aligning at least one of the phrase candidates with a rhythmic skeleton for the target song; and (iv) preparing a resultant audio encoding of the speech in correspondence with the temporally aligned phrase candidate-mapped from onset-delimited segments of the input audio encoding.
In some embodiments, the method further includes mixing the resultant audio encoding with an audio encoding of a backing track for the target song, and audibly rendering the mixed audio. In some embodiments, the method further includes capturing (e.g., from a microphone input of a portable handheld device) speech voiced by a user thereof as the input audio encoding, and retrieving (e.g., responsive to a selection of the target song by the user) a computer readable encoding of at least one of the phrase template and the rhythmic skeleton. In some cases, the retrieving responsive to user selection includes obtaining, from a remote store and via a communication interface of the portable handheld device, at least the phrase template.
In some cases, the segmenting includes applying a spectral difference type (SDF-type) function to the audio encoding of the speech and picking temporally indexed peaks in a result thereof as onset candidates within the speech encoding; and agglomerating adjacent onset candidate-delimited sub-portions of the speech encoding into segments based, at least in part, on comparative strength of onset candidates. In some cases, the SDF-type function operates on a psychoacoustically-based representation of power spectrum for the speech encoding. In some cases, the agglomerating is performed, at least in part, based on a minimum segment length threshold. In some cases, the method includes iterating on the agglomerating to achieve a total number of segments within a target range.
In some cases, the mapping includes enumerating a set of onset-delimited, N-part, partitionings of the speech encoding based on groupings of adjacent ones of the segments, wherein N corresponds to the number of sub-phrase portions of the phrase template. The mapping also includes, for each of the partitionings, constructing a corresponding mapping of the speech encoding segment groupings to sub-phrase portions, the mappings providing plural of the phrase candidates.
In some cases, the mapping provides plural phrase candidates, wherein the temporal aligning is performed for each of the plural phrase candidates, and further includes selecting from amongst the plural phrase candidates based upon degree of rhythmic alignment with the rhythmic skeleton for the target song.
In some cases, the rhythmic skeleton corresponds to a pulse train encoding of tempo of the target song. In some cases, the target song includes plural constituent rhythms, and the pulse train encoding includes respective pulses scaled in accord with relative strengths of the constituent rhythms.
In some embodiments, the method further includes performing beat detection for a backing track of the target song to produce the rhythmic skeleton. In some embodiments, the method further includes pitch shifting the resultant audio encoding in accord with a note sequence for the target song. In some cases, the pitch shifting employs cross synthesis of a glottal pulse.
In some embodiments, the method further includes retrieving a computer readable encoding of the note sequence. In some cases, the retrieving is responsive to user selection at a user interface of a portable handheld device and obtains at least the phrase template and the note sequence for the target song from a remote store via a communication interface of the portable handheld device.
In some embodiments, the method further includes mapping onsets of notes for the target song to temporally-proximate, segment delimiting onsets in the speech encoding and, for respective portions of the speech encoding that correspond to the mapped note onsets, temporally stretching or compressing the respective portion to fill duration of the mapped note. In some embodiments, the method further includes characterizing frames of the speech encoding based, at least in part, on spectral roll-off, wherein generally greater roll-off of high frequency content is indicative of voiced vowels and dynamically varying magnitude of the temporal stretching applied to a respective portion of the speech encoding based on the characterized vowel-indicative spectral roll-off for the corresponding frame. In some cases, the dynamic varying employs a composition of a melodic density vector for the target song and a spectral roll-off vector for the speech encoding.
In some embodiments, the method is performed on a portable computing device selected from the group of a compute pad, a personal digital assistant or book reader, and a mobile phone or media player. In some embodiments, the method is performed using a purpose-built, toy or amusement device. In some embodiments, a computer program product encodes, in one or more media, instructions executable on a processor of a portable computing device to cause the portable computing device to perform the method. In some cases, the one or more media are readable by the portable computing device or readable incident to a computer program product conveying transmission to the portable computing device.
In some embodiments, in accordance with the present invention, an apparatus includes a portable computing device and machine readable code embodied in a non-transitory medium and executable on the portable computing device to transform an input audio encoding of speech into an output that is rhythmically consistent with a target song, the machine readable code including instructions executable to segment the input audio encoding of the speech into plural segments, the segments corresponding to successive sequences of samples of the audio encoding and delimited by onsets identified therein. The machine readable code is further executable to map individual ones of the plural segments to respective sub-phrase portions of a phrase template for the target song, the mapping establishing one or more phrase candidates. The machine readable code is further executable to temporally align at least one of the phrase candidates with a rhythmic skeleton for the target song. The machine readable code is further executable to prepare a resultant audio encoding of the speech in correspondence with the temporally aligned phrase candidate-mapped from onset-delimited segments of the input audio encoding. In some cases, the apparatus is embodied as one or more of a compute pad, a handheld mobile device, a mobile phone, a personal digital assistant, a smart phone, a media player and a book reader.
In some embodiments in accordance with the present invention, a computer program product is encoded in non-transitory media and including instructions executable to transform an input audio encoding of speech into an output that is rhythmically consistent with a target song. The computer program product encodes and includes instructions executable to segment the input audio encoding of the speech into plural segments, the segments corresponding to successive sequences of samples of the audio encoding and delimited by onsets identified therein. The computer program product further encodes and includes instructions executable to map individual ones of the plural segments to respective sub-phrase portions of a phrase template for the target song, the mapping establishing a one or more phrase candidates. The computer program product further encodes and includes instructions executable to temporally align at least one of the phrase candidates with a rhythmic skeleton for the target song. The computer program product further encodes and includes instructions executable to prepare a resultant audio encoding of the speech in correspondence with the temporally aligned phrase candidate-mapped from onset-delimited segments of the input audio encoding. In some cases, the media are readable by the portable computing device or readable incident to a computer program product conveying transmission to the portable computing device.
These and other embodiments, together with numerous variations thereon, will be appreciated by persons of ordinary skill in the art based on the description, claims and drawings that follow.
The present invention may be better understood, and its numerous objects, features, and advantages made apparent to those skilled in the art by referencing the accompanying drawings.
The use of the same reference symbols in different drawings indicates similar or identical items.
As described herein, automatic transformations of captured user vocals may provide captivating applications executable even on the handheld compute platforms that have become ubiquitous since the advent of iOS and Android-based phones, media devices and tablets. The automatic transformations may even be implemented in purpose-built devices, such as for the toy, gaming or amusement device markets.
Advanced digital signal processing techniques described herein allow implementations in which mere novice user-musicians may generate, audibly render and share musical performances. In some cases, the automated transformations allow spoken vocals to be segmented, arranged, temporally aligned with a target rhythm, meter or accompanying backing tracks and pitch corrected in accord with a score or note sequence. Speech-to-song music implementations are one such example and exemplary songification application is described below. In some cases, spoken vocals may be transformed in accord with musical genres such as rap using automated segmentation and temporal alignment techniques, often without pitch correction. Such applications, which may employ different signal processing and different automated transformations, may nonetheless be understood as speech-to-rap variations on the theme. Adaptations to provide an exemplary AutoRap application are also described herein.
In the interest of concreteness, processing and device capabilities, terminology, API frameworks and even form factors typical of a particular implementation environment, namely the iOS device space popularized by Apple, Inc. have been assumed. Notwithstanding descriptive reliance on any such examples or framework, persons of ordinary skill in the art having access to the present disclosure will appreciate deployments and suitable adaptations for other compute platforms and other concrete physical implementations.
Various illustrated functional blocks (e.g., audio signal segmentation 371, segment to phrase mapping 372, temporal alignment and stretch/compression 373 of segments, and pitch correction 374) will be understood, with reference to signal processing techniques detailed herein, to operate upon audio signal encodings derived from captured vocals and represented in memory or non-volatile storage on the compute platform.
When lyrics are set to a melody, it is often the case that certain phrases are repeated to reinforce musical structure. Our speech segmentation algorithm attempts to determine boundaries between words and phrases in the speech input so that phrases can be repeated or otherwise rearranged. Because words are typically not separated by silence, simple silence detection may, as a practical matter, be insufficient in many applications. Exemplary techniques for segmentation of the captured speech audio signal will be understood with reference to
The speech utterance is typically digitized as speech encoding 501 using a sample rate of 44100 Hz. A power spectrum is computed from the spectrogram. For each frame, an FFT is taken using a Hann window of size 1024 (with a 50% overlap). This returns a matrix, with rows representing frequency bins and columns representing time-steps. In order to take into account human loudness perception, the power spectrum is transformed into a sone-based representation. In some implementations, an initial step of this process involves a set of critical-band filters, or bark band filters 511, which model the auditory filters present in the inner ear. The filter width and response varies with frequency, transforming the linear frequency scale to a logarithmic one. Additionally, the resulting sone representation 502 takes into account the filtering qualities of the outer ear as well as modeling spectral masking. At the end of this process, a new matrix is returned with rows corresponding to critical bands and columns to time-steps.
One approach to segmentation involves finding onsets. New events, such as the striking of a note on a piano, lead to sudden increases in energy in various frequency bands. This can often be seen in the time-domain representation of the waveform as a local peak. A class of techniques for finding onsets involves computing (512) a spectral difference function (SDF). Given a spectrogram, the SDF is the first difference and is computed by summing the differences in amplitudes for each frequency bin at adjacent time-steps. For example:
SDF[i]=(Σ(B[i]−B[i−1]).25)4
Here we apply a similar procedure to the sone representation, yielding a type of SDF 513. The illustrated SDF 513 is a one-dimensional function, with peaks indicating likely onset candidates.
We next define onset candidates 503 to be the temporal location of local maxima (or peaks 513.1, 513.2, 513.3 . . . 513.99) that may be picked from the SDF (513). These locations indicate the possible times of the onsets. We additionally return a measure of onset strength that is determined by subtracting the level of the SDF curve at the local maximum from the median of the function over a small window centered at the maximum. Onsets that have an onset strength below a threshold value are typically discarded. Peak picking 514 produces a series of above-threshold-strength onset candidates 503.
We define a segment (e.g., segment 515.1) to be a chunk of audio between two adjacent onsets. In some cases, the onset detection algorithm described above can lead to many false positives leading to very small segments (e.g. much smaller than the duration of a typical word). To reduce the number of such segments, certain segments (see e.g., segment 515.2) are merged (515.2)) using an agglomeration algorithm. First, we determine whether there are segments that are shorter than a threshold value (here we start at 0.372 seconds threshold). If so, they are merged with a segment that temporally precedes or follows. In some cases, the direction of the merge is determined based on the strength of the neighboring onsets.
The result is segments that are based on a strong onset candidates and agglomeration of short neighboring segments to produce the segments (504) that define a segmented version of the speech encoding (501) that are used in subsequent steps. In the case of speech-to-song embodiments (see
One goal of previously described the phrase construction step is to create phrases by combining segments (e.g., segments 504 such as may be generated in accord with techniques illustrated and described above relative to
In some implementations of the techniques, it is useful to require the number of segments to be greater than the number of sub-phrases. Mapping of segments to sub-phrases can be framed as a partitioning problem. Let m be the number of sub-phrases in the target phrase. Then we require m-1 dividers in order to divide the vocal utterance into the correct number of phrases. In our process, we allow partitions only at onset locations. For example, in
Note that some embodiments, a user may select and reselect from a library of phrase templates for differing target songs, performances, artists, styles etc. In some embodiments, phrase templates may be transacted, made available or demand supplied (or computed) in accordance with a part of an in-app-purchase revenue model or may be earned, published or exchanged as part of a gaming, teaching and/or social-type user interaction supported.
Because the number of possible phrases increases combinatorially with the number of segments, in some practical implementations, we restrict the total segments to a maximum of 20. Of course, more generally and for any given application, search space may be increased or decreased in accord with processing resources and storage available. If the number of segments is greater than this maximum after the first pass of the onset detection algorithm, the process is repeated using a higher minimum duration for agglomerating the segments. For example, if the original minimum segment length was 0.372 seconds, this might be increased to 0.5 seconds, leading to fewer segments. The process of increasing the minimum threshold will continue until the number of target segments is less than the desired amount. On the other hand, if the number of segments is less than the number of sub-phrases, then it will generally not be possible to map segments to sub-phrases without mapping the same segment to more than one sub-phrase. To remedy this, the onset detection algorithm is reevaluated in some embodiments using a lower segment length threshold, which typically results in fewer onsets agglomerated into a larger number of segments. Accordingly, in some embodiments, we continue to reduce the length threshold value until the number of segments exceeds the maximum number of sub-phrases present in any of the phrase templates. We have a minimum sub-phrase length we have to meet, and this is lowered if necessary to allow partitions with shorter segments.
Based on the description herein, persons of ordinary skill in the art will recognize numerous opportunities for feeding back information from later stages of a computational process to earlier stages. Descriptive focus herein on the forward direction of process flows is for ease and continuity of description and is not intended to be limiting.
Each possible partition described above represents a candidate phrase for the currently considered phrase template. To summarize, we exclusively map one or more segments to a sub-phrase. The total phrase is then created by assembling the sub-phrases according to the phrase template. In the next stage, we wish to find the candidate phrase that can be most closely aligned to the rhythmic structure of the backing track. By this we mean we would like the phrase to sound as if it is on the beat. This can often be achieved by making sure accents in the speech tend to align with beats, or other metrically important positions.
To provide this rhythmic alignment, we introduce a rhythmic skeleton (RS) 603 as illustrated in
We measure the degree of rhythmic alignment (RA), between the rhythmic skeleton and the phrase, by taking the cross correlation of the RS with the spectral difference function (SDF), calculated using the sone representation. Recall that the SDF represents sudden changes in signal that correspond to onsets. In the music information retrieval literature we refer to this continuous curve that underlies onset detection algorithms as a detection function. The detection function is an effective method for representing the accent or mid-level event structure of the audio signal. The cross correlation function measures the degree of correspondence for various lags, by performing a point-wise multiplication between the RS and the SDF and summing, assuming different starting positions within the SDF buffer. Thus for each lag the cross correlation returns a score. The peak of the cross correlation function indicates the lag with the greatest alignment. The height of the peak is taken as a score of this fit, and its location gives the lag in seconds.
The alignment score A is then given by
This process is repeated for all phrases and the phrase with the highest score is used. The lag is used to rotate the phrase so that it starts from that point. This is done in a circular manner. It is worth noting that the best fit can be found across phrases generated by all phrase templates or just a given phrase template. We choose to optimize across all phrase templates, giving a better rhythmic fit and naturally introducing variety to the phrase structure.
When a partition mapping requires a sub-phrase to repeat (as in a rhythmic pattern such as specified by the phrase template {A A B C}), the repeated sub-phrase was found to sound more rhythmic when the repetition was padded to occur on the next beat. Likewise, the entire resultant partitioned phrase is padded to the length of a measure before repeating with the backing track.
Accordingly, at the end of the phrase construction (613) and rhythmic alignment (614) procedure, we have a complete phrase constructed from segments of the original vocal utterance that has been aligned to the backing track. If the backing track or vocal input is changed, the process is re-run. This concludes the first part of an illustrative “songification” process. A second part, which we now describe, transforms the speech into a melody.
To further synchronize the onsets of the voice with the onsets of the notes in the desired melody line, we use a procedure to stretch voice segments to match the length of the melody. For each note in the melody, the segment onset (calculated by our segmentation procedure described above) that occurs nearest in time to the note onset while still within a given time window is mapped to this note onset. The notes are iterated through (typically exhaustively and typically in a generally random order to remove bias and to introduce variability in the stretching from run to run) until all notes with a possible matching segment are mapped. The note-to-segment map then is given to the sequencer which then stretches each segment the appropriate amount such that it fills the note to which it is mapped. Since each segment is mapped to a note that is nearby, the cumulative stretch factor over the entire utterance should be more or less unity, however if a global stretch amount is desired (e.g. slow down the result utterance by 2), this is achieved by mapping the segments to a sped-up version of the melody: the output stretch amounts are then scaled to match the original speed of the melody, resulting in an overall tendency to stretch by the inverse of the speed factor.
Although the alignment and note-to-segment stretching processes synchronize the onsets of the voice with the notes of the melody, the musical structure of the backing track can be further emphasized by stretching the syllables to fill the length of the notes. To achieve this without losing intelligibility, we use dynamic time stretching to stretch the vowel sounds in the speech, while leaving the consonants as they are. Since consonant sounds are usually characterized by their high frequency content, we used spectral roll-off up to 95% of the total energy as the distinguishing feature between vowels and consonants. Spectral roll-off is defined as follows. If we let |X[k]| be the magnitude of the k-th Fourier coefficient, then the roll-off for a threshold of 95% is defined to be k_roll=Σk=0k
The spectral roll-off of the voice segments are calculated for each analysis frame of 1024 samples and 50% overlap. Along with this the melodic density of the associated melody (MIDI symbols) is calculated over a moving window, normalized across the entire melody and then interpolated to give a smooth curve. The dot product of the spectral roll-off and the normalized melodic density provides a matrix, which is then treated as the input to the standard dynamic programming problem of finding the path through the matrix with the minimum associated cost. Each step in the matrix is associated with a corresponding cost that can be tweaked to adjust the path taken through the matrix. This procedure yields the amount of stretching required for each frame in the segment to fill the corresponding notes in the melody.
Although fundamental frequency, or pitch, of speech varies continuously, it is does not generally sound like a musical melody. The variations are typically too small, too rapid, or too infrequent to sound like a musical melody. Pitch variations occur for a variety of reasons including the mechanics of voice production, the emotional state of the speaker, to indicate phrase endings or questions, and an inherent part of tone languages.
In some embodiments, the audio encoding of speech segments (aligned/stretched/compressed to a rhythmic skeleton or grid as described above) is pitch corrected in accord with a note sequence or melody score. As before, the note sequence or melody score may be precomputed and downloaded for, or in connection with, a backing track.
For some embodiments, a desirable attribute of an implemented speech-to-melody (S2M) transformation is that the speech should remain intelligible while sounding clearly like a musical melody. Although persons of ordinary skill in the art will appreciate a variety of possible techniques that may be employed, our approach is based on cross-synthesis of a glottal pulse, which emulates the periodic excitation of the voice, with the speaker's voice. This leads to a clearly pitched signal that retains the timbral characteristics of the voice, allowing the speech content to be clearly understood in a wide variety of situations.
The input speech 703 is sampled at 44.1 kHz and its spectrogram is calculated (704) using a 1024 sample Hann window (23 ms) overlapped by 75 samples. The glottal pulse (705) was based on the Rosenberg model which is shown in
Parameters of the Rosenberg glottal pulse include the relative open duration (tf-t0/Tp) and the relative closed duration ((Tp-tf)/Tp). By varying these ratios the timbral characteristics can be varied. In addition to this, the basic shape was modified to give the pulse a more natural quality. In particular, the mathematically defined shape was traced by hand (i.e. using a mouse with a paint program), leading to slight irregularities. The “dirtied waveform was then low-passed filtered using a 20-point finite impulse response (FIR) filter to remove sudden discontinuities introduced by the quantization of the mouse coordinates.
The pitch of the above glottal pulse is given by Tp. In our case, we wished to be able to flexibly use the same glottal pulse shape for different pitches, and to be able to control this continuously. This was accomplished by resampling the glottal pulse according to the desired pitch, thus changing the amount by which to hop in the waveform. Linear interpolation was used to determine the value of the glottal pulse at each hop.
The spectrogram of the glottal waveform was taken using a 1024 sample Hann window overlapped by 75%. The cross synthesis (702) between the periodic glottal pulse waveform and the speech was accomplished by multiplying (706) the magnitude spectrum (707) of each frame of the speech by the complex spectrum of the glottal pulse, effectively rescaling the magnitude of the complex amplitudes according to the glottal pulse spectrum. In some cases or embodiments, rather than using the magnitude spectrum directly, the energy in each bark band is used after pre-emphasizing (spectral whitening) the spectrum. In this way, the harmonic structure of the glottal pulse spectrum is undisturbed while the formant structure of the speech is imprinted upon it. We have found this to be an effective technique for the speech to music transform.
One issue that arises with the above approach is that un-voiced sounds such as some consonant phonemes, which are inherently noisy, are not modeled well by the above approach. This can lead to a “ringing sound” when they are present in the speech and to a loss of percussive quality. To better preserve these sections, we introduce a controlled amount of high passed white noise (708). Unvoiced sounds tend to have a broadband spectrum, and spectral roll-off is again used as an indicative audio feature. Specifically, frames that are not characterized by significant roll-off of high frequency content are candidates for a somewhat compensatory addition of high passed white noise. The amount of noise introduced is controlled by the spectral roll-off of the frame, such that unvoiced sounds that have a broadband spectrum, but which are otherwise not well modeled using the glottal pulse techniques described above, are mixed with an amount of high passed white noise that is controlled by this indicative audio feature. We have found that this leads to output which is much more intelligible and natural.
Some implementations of the speech to music songification process described above employ a pitch control signal which determines the pitch of the glottal pulse. As will be appreciated, the control signal can be generated in any number of ways. For example, it might be generated randomly, or according to statistical model. In some cases or embodiments, a pitch control signal (e.g., 711) is based on a melody (701) that has been composed using symbolic notation, or sung. In the former case, a symbolic notation, such as MIDI is processed using a Python script to generate an audio rate control signal consisting of a vector of target pitch values. In the case of a sung melody, a pitch detection algorithm can be used to generate the control signal. Depending on the granularity of the pitch estimate, linear interpolation is used to generate the audio rate control signal.
A further step in creating a song is mixing the aligned and synthesis transformed speech (output 710) with a backing track, which is in the form of a digital audio file. It should be noted that as described above, it is not known in advance how long the final melody will be. The rhythmic alignment step may choose a short or long pattern. To account for this, the backing track is typically composed so that it can be seamlessly looped to accommodate longer patterns. If the final melody is shorter than the loop, then no action is taken and there will be a portion of song with no vocals.
Variations for Output Consistent with Other Genres
We now describe further methods that are more suitable for transforming speech into “rap”, that is, speech that has been rhythmically aligned to a beat. We call this procedure “AutoRap” and persons of ordinary skill in the art will appreciate a broad range of implementations based on the description herein. In particular, aspects of a larger computational flow (e.g., as summarized in
As before, segmentation (here segmentation 911) employs a detection function is calculated using the spectral difference function based on a bark band representation. However, here we emphasize a sub-band from approximately 700 Hz to 1500 Hz, when computing the detection function. It was found that a band-limited or emphasized DF more closely corresponds to the syllable nuclei, which perceptually are points of stress in the speech.
More specifically, it has been found that while a mid-band limitation provides good detection performance, even better detection performance can be achieved in some cases by weighting the mid-bands but still considering spectrum outside the emphasized mid-band. This is because percussive onsets, which are characterized by broadband features, are captured in addition to vowel onsets, which are primarily detected using mid-bands. In some embodiments, a desirable weighting is based on taking the log of the power in each bark band and multiplying by 10, for he mid-bands, while not applying the log or rescaling to other bands.
When the spectral difference is computed, this approach tends to give greater weight to the mid-bands since the range of values is greater. However, because the L-norm is used with a value of 0.25 when computing the distance in the spectral distance function, small changes that occur across many bands will also register as a large change, such as if a difference of a greater magnitude had been observed in one, or a few, bands. If a Euclidean distance had been used, this effect would not have been observed. Of course, other mid-band emphasis techniques may be utilized in other embodiments.
Aside from the mid-band emphasis just described, detection function computation is analogous to the spectral difference (SDF) techniques described above for speech-to-song implementations (recall
Next, a rhythmic pattern (e.g., rhythmic skeleton or grid 903) is defined, generated or retrieved. Note that some embodiments, a user may select and reselect from a library of rhythmic skeletons for differing target raps, performances, artists, styles etc. As with phrase templates, rhythmic skeletons or grids may be transacted, made available or demand supplied (or computed) in accordance with a part of an in-app-purchase revenue model or may be earned, published or exchanged as part of a gaming, teaching and/or social-type user interaction supported.
In some embodiments, a rhythmic pattern is represented as a series of impulses at particular time locations. For example, this might simply be an equally spaced grid of impulses, where the inter-pulse width is related to the tempo of the current song. If the song has a tempo of 120 BPM, and thus an inter-beat period of 0.5 s, then the inter-pulse would typically be an integer fraction of this (e.g. 0.5, 0.25, etc.). In musical terms, this is equivalent to an impulse every quarter note, or every eighth note, etc. More complex patterns can also be defined. For example, we might specify a repeating pattern of two quarter notes followed by four eighth notes, making a four beat pattern. At a tempo of 120 BPM the pulses would be at the following time locations (in seconds): 0, 0.5. 1.5, 1.75, 2.0, 2.25, 3.0, 3.5, 4.0, 4.25, 4.5, 4.75.
After segmentation (911) and grid construction, alignment is (912) performed.
Two additional strategies are employed to minimize excessive stretching or compression. First, rather than only starting the mapping from S1, we consider all mapping starting from every possible segment and wrapping around when the end is reached. Thus, if we start at S5 the mapping will be segment S5 to pulse P1, S6 to P2 etc. For each starting point, we measure the total amount of stretching/compression, which we call rhythmic distortion. In some embodiments, a rhythmic distortion score is computed as the reciprocal of stretch ratios less than one. This procedure is repeated for each rhythmic pattern. The rhythmic pattern (e.g., rhythmic skeleton or grid 903) and starting point which minimize the rhythmic distortion score are taken to be the best mapping and used for synthesis.
In some cases or embodiments, an alternate rhythmic distortion score, that we found often worked better, was computed by counting the number of outliers in the distribution of the speed scores. Specifically, the data were divided into deciles and the number of segments whose speed scores were in the bottom and top deciles were added to give the score. A higher score indicates more outliers and thus a greater degree of rhythmic distortion.
Second, phase vocoder 913 is used for stretching/compression at a variable rate. This is done in real-time, that is, without access to the entire source audio. Time stretch and compression necessarily result in input and output of different lengths—this is used to control the degree of stretching/compression. In some cases or embodiments, phase vocoder 913 operates with four times overlap, adding its output to an accumulating FIFO buffer. As output is requested, data is copied from this buffer. When the end of the valid portion of this buffer is reached, the core routine generates the next hop of data at the current time step. For each hop, new input data is retrieved by a callback, provided during initialization, which allows an external object to control the amount of time-stretching/compression by providing a certain number of audio samples. To calculate the output for one time step, two overlapping windows of length 1024 (nfft), offset by nfft/4, are compared, along with the complex output from the previous time step. To allow for this in a real-time context where the full input signal may not be available, phase vocoder 913 maintains a FIFO buffer of the input signal, of length 5/4 nfft; thus these two overlapping windows are available at any time step. The window with the most recent data is referred to as the “front” window; the other (“back”) window is used to get delta phase.
First, the previous complex output is normalized by its magnitude, to get a vector of unit-magnitude complex numbers, representing the phase component. Then the FFT is taken of both front and back windows. The normalized previous output is multiplied by the complex conjugate of the back window, resulting in a complex vector with the magnitude of the back window, and phase equal to the difference between the back window and the previous output.
We attempt to preserve phase coherence between adjacent frequency bins by replacing each complex amplitude of a given frequency bin with the average over its immediate neighbors. If a clear sinusoid is present in one bin, with low-level noise in adjacent bins, then its magnitude will be greater than its neighbors and their phases will be replaced by that of the true sinusoid. We find that this significantly improves resynthesis quality.
The resulting vector is then normalized by its magnitude; a tiny offset is added before normalization to ensure that even zero-magnitude bins will normalize to unit magnitude. This vector is multiplied with the Fourier transform of the front window; the resulting vector has the magnitude of the front window, but the phase will be the phase of the previous output plus the difference between the front and back windows. If output is requested at the same rate that input is provided by the callback, then this would be equivalent to reconstruction if the phase coherence step were excluded.
Some embodiments in accordance with the present invention(s) may take the form of, and/or be provided as, purpose-built devices such as for the toy or amusement markets.
While the invention(s) is (are) described with reference to various embodiments, it will be understood that these embodiments are illustrative and that the scope of the invention(s) is not limited to them. Many variations, modifications, additions, and improvements are possible. For example, while embodiments have been described in which vocal speech is captured and automatically transformed and aligned for mix with a backing track, it will be appreciated that automated transforms of captured vocals described herein may also be employed to provide expressive performances that are temporally aligned with a target rhythm or meter (such as may be characteristic of a poem, iambic cycle, limerick, etc.) and without musical accompaniment.
Furthermore, while certain illustrative signal processing techniques have been described in the context of certain illustrative applications, persons of ordinary skill in the art will recognize that it is straightforward to modify the described techniques to accommodate other suitable signal processing techniques and effects.
Some embodiments in accordance with the present invention(s) may take the form of, and/or be provided as, a computer program product encoded in a machine-readable medium as instruction sequences and other functional constructs of software tangibly embodied in non-transient media, which may in turn be executed in a computational system (such as a iPhone handheld, mobile device or portable computing device) to perform methods described herein. In general, a machine readable medium can include tangible articles that encode information in a form (e.g., as applications, source or object code, functionally descriptive information, etc.) readable by a machine (e.g., a computer, computational facilities of a mobile device or portable computing device, etc.) as well as tangible, non-transient storage incident to transmission of the information. A machine-readable medium may include, but is not limited to, magnetic storage medium (e.g., disks and/or tape storage); optical storage medium (e.g., CD-ROM, DVD, etc.); magneto-optical storage medium; read only memory (ROM); random access memory (RAM); erasable programmable memory (e.g., EPROM and EEPROM); flash memory; or other types of medium suitable for storing electronic instructions, operation sequences, functionally descriptive information encodings, etc.
In general, plural instances may be provided for components, operations or structures described herein as a single instance. Boundaries between various components, operations and data stores are somewhat arbitrary, and particular operations are illustrated in the context of specific illustrative configurations. Other allocations of functionality are envisioned and may fall within the scope of the invention(s). In general, structures and functionality presented as separate components in the exemplary configurations may be implemented as a combined structure or component. Similarly, structures and functionality presented as a single component may be implemented as separate components. These and other variations, modifications, additions, and improvements may fall within the scope of the invention(s).
The present application claims priority of Provisional Application No. 61/617,643, filed Mar. 29, 2012, the entirety of which is incorporated herein by reference.
Number | Date | Country | |
---|---|---|---|
61617643 | Mar 2012 | US |