Multiple pitch transcription and melody harmonization...
Transcript of Multiple pitch transcription and melody harmonization...
![Page 1: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/1.jpg)
Multiple pitch transcription and melody harmonization with probabilistic musicological
models
Stanisław A. Raczyński, Emmanuel Vincent
![Page 2: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/2.jpg)
Introduction
● Musical quantities can be thought of as interconnected variables.
● Each variable holds information about itself and about others as well.
● For example, certain chords and their progressions (e.g., iim7 |V7 |Imaj7) will suggest that the genre is jazz, which will also implicate the use of certain instruments (e.g., saxophone, piano, double bass)
![Page 3: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/3.jpg)
Bayesian networks● Relations between variables can be represented
in a form of a Bayesian network [1]:
![Page 4: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/4.jpg)
Relation to language models● In processing natural language (e.g., continuous
speech recognition), probabilistic models of language are used and they are called linguistic models or language models.
● In music information retrieval, their equivalents are referred to as musicological models or music models.
![Page 5: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/5.jpg)
Multiple pitch estimation
Estimating note pitches, onsets and durations given an audio recording
![Page 6: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/6.jpg)
Current approaches
● The most popular approaches are based on Nonnegative Matrix Factorization (NMF).
● A spectrogram (typically obtained using a constantQ or ERB filter bank) X of the recording is factorized to obtain the dictionary matrix A and the salience matrix S:
● The salience matrix is then analyzed to find the positions of notes
X = AS
![Page 7: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/7.jpg)
Current approaches
● NMF is a midlevel representation of the audio.● Typically, the salience values are analyzes
individually, e.g., thresholded.
● Better results can be obtained if relations between the underlying binary note variables and more aspects of the music are modeled jointly.
Nt,f
St,f
Latent note variable Observed salience variable
![Page 8: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/8.jpg)
Music pitch model
Salienceobtained with acoustic model(NMF)
Notes
Chords
In our experiments we have used a Dynamic Bayesian Network to model relations between the latent and observed variables:
![Page 9: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/9.jpg)
Harmonization
● Guessing the underlying chord sequence given a melody
● Used for automatic music composition, automatic accompaniment, etc.
![Page 10: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/10.jpg)
Current approaches
● A typical approach for harmonization uses Hidden Markov Models (HMMs) to model relations between the latent chords and the melody:
● This approach is used in such commercial applications as MySong [2] or Bandinabox [3].
![Page 11: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/11.jpg)
Music melody modelIn our experiments we have used a Dynamic Bayesian Network to model relations between the latent and observed variables:
Melody
Chords
Tonality
![Page 12: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/12.jpg)
Model complexity
2K × 24 × 2K = 2.3∙1054 parameters for K = 88
● Jointly modeling multiple variables causes the number of parameters to explode
![Page 13: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/13.jpg)
Model interpolation
● Complexity can be reduced by approximating the joint model with a combination of simpler models – model interpolation.
● Model interpolation has been successfully used in natural language processing by Klakow [4].
● This technique is also used to reduce overfitting: models of different order are combined (model smoothing).
![Page 14: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/14.jpg)
Model interpolation: linear
Submodels Pi use only a small subset of the conditioning
variable set, e.g.:
![Page 15: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/15.jpg)
Model interpolation: log-linear
![Page 16: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/16.jpg)
Models and submodels used
and their trained parameter values
![Page 17: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/17.jpg)
Pitch submodels
Harmony
Duration
Voice
Polyphony
Neighbor
![Page 18: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/18.jpg)
Harmony submodel● Independent of octave, depends only on the
chord mode and the interval from chord's root:
![Page 19: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/19.jpg)
Duration submodel● Simple binary bigram model:
![Page 20: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/20.jpg)
Voice submodel● Pitch activity depends only on the distance to
the closest active pitch in the previous frame:
Perfect fifth Octave
![Page 21: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/21.jpg)
Polyphony submodel● Pitch activity depends only on the number of
active notes below the current pitch:
![Page 22: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/22.jpg)
Neighbor submodel● A binary trigram model in the frequency domain:
11 sequences are less likely than 01
![Page 23: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/23.jpg)
Chord model● Modeled with a multinomial distribution.● 24chord dictionary.● Statetying is used because we do not model the tonality.
![Page 24: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/24.jpg)
Harmonization submodels
Melody
Tonality
Chord bigram
Note: it is a discriminative model
![Page 25: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/25.jpg)
Melody submodel● Mt is a set of active notes at time frame t.
● State tying: note patterns with the same content relative to the chord root were given identical probabilities, e.g., the unordered note combination (C,G) in the chord of Cmajor is equally probable as the note combination (D ,A ) in the chord of D major♯ ♯ ♯
![Page 26: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/26.jpg)
Chord bigram submodel● A binary trigram model in the frequency domain.● Chord labelled by one of 13 root pitch classes:
C, C , D, D , E, F, F , G, G , A, A , B or “none” for nonchords♯ ♯ ♯ ♯ ♯
and one of 27 chord types:major, minor, dominant, diminished, halfdiminished, augmented, power, suspendedsecond, suspendedfourth, majorsixth, minorsixth, majorseventh, minorseventh, dominantseventh, diminishedseventh, augmentedseventh, majorninth, minorninth, dominantninth, augmentedninth, minoreleventh, dominanteleventh, majorminor, minormajor, majorthirteenth, dominantthirteenth or “none” for nonchords
● N = 351 distinct chord labels
![Page 27: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/27.jpg)
Chord bigram submodel
1 frame = 1 beat 1 frame = 16 beats
Ct1
= Gmaj
![Page 28: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/28.jpg)
Melody submodel
Mt = (C) M
t = (C,E,G)
1 frame = 1 beat
![Page 29: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/29.jpg)
Tonality submodel● Tonality encoded as one of 24 different key labels resulting from the
combination of 12 tonics (C, C , D, D , E, F, F , G, G , A, A , B) and 2 ♯ ♯ ♯ ♯ ♯modes (major or minor)
● State tying: chords corresponding to the same scale degree in different keys are tied together.
1 frame = 1 beat
![Page 30: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/30.jpg)
Smoothing● To avoid overfitting in the submodels, they are
interpolated with simpler chord models (additive smoothing): chord unigram and zerogram:
Subset of variables Full set of variables
![Page 31: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/31.jpg)
Smoothing● In case of loglinear interpolation, each submodel is
smoothed separately:
for all i
![Page 32: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/32.jpg)
Chord unigram submodel
1 frame = 1 beat
![Page 33: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/33.jpg)
Evaluation
![Page 34: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/34.jpg)
Multiple pitch analysis data
● Mutopia dataset was used:● ~1300 files for training model parameters● 100 fles for validation● 100 files for testing● 1 frame = 1/6 of a beat
● RWC files annotated with harmony was used to train the harmony submodel and the chord models
![Page 35: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/35.jpg)
Harmonization data
● For training, we have used a collection of around 2000 lead sheets from the Wikifonia web page:● melodies annotated with keys and absolute chord
labels,● mostly popular (e.g., pop, rock) songs from the
twentieth and the twentyfirst centuries,● the songs were first screened for improper chord
labels and wrong keys.
![Page 36: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/36.jpg)
Training
● Model parameters were trained by counting occurrences (maximizing the likelihood) on the training dataset.
● The smoothing parameters were optimized by maximizing the average crossentropy of individual submodels on the validation dataset.
● Interpolation coefficients and smoothing for linearcombined harmonization model were optimized by maximizing crossentropy of the validation dataset
![Page 37: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/37.jpg)
Reference pitch modelBPB
![Page 38: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/38.jpg)
Cross-entropy
● Common metric for measuring modeling power of language [7] and music [5,6] models.
● Multipitch estimation:
● Harmonization:
![Page 39: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/39.jpg)
Contextual cross-entropy
● For multipitch analysis, the crossentropy value is dominated by the silence (97% notes are inactive on average).
● We would like to know how well do the models model the note activity, i.e., note onsets, note offsets and notes – contextual crossentropy.
![Page 40: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/40.jpg)
Pitch cross-entropy
● Regular crossentropy (in milibits):
● Contextual crossentropy (in milibits):
![Page 41: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/41.jpg)
Pitch cross-entropy
![Page 42: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/42.jpg)
Harmonization cross-entropy
M = melody submodel, T = tonality submodel, B = chord bigram submodel
![Page 43: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/43.jpg)
Harmonization cross-entropy
![Page 44: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/44.jpg)
Accuracy
● Multipitch estimation:● Precision, Recall and Fmeasure● Reference musicological model: Bernoulli (equivalent to
thresholded NMF) and pitchdependent Bernoulli (eq. to pitchdependent threshold)
● Harmonization:● Root note estimation accuracy (compared to leadsheets)
and triad accuracy (root note + first chord interval)● Reference musicological model: Harmonic Analyzer by
Temperley & Sleator [34]
![Page 45: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/45.jpg)
Pitch estimation accuracyPrecision P, Recall R and Fmeasure F
P, R and F relative to the full model
![Page 46: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/46.jpg)
Harmonization accuracy
● Root note estimation accuracies
simple weighter
![Page 47: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/47.jpg)
Harmonization accuracy
● Triad accuracies
WR = weighted root note accuracy
![Page 48: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/48.jpg)
Conclusion
● Multiple musical variables can be jointly modeled to improve their estimates
● Model interpolation is efficient in dealing with joint model complexity
● Linear interpolation seems to work slightly worse than the loglinear one
![Page 49: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/49.jpg)
Possible future work
● A larger number of more complex submodels could be investigated for further improvement in terms of crossentropy and accuracy.
● Proposed method could be tested on a larger populations of songs that would include more diverse musical genres.
● Subjective listening tests could also be used to analyze the quality of the harmonizations in more detail.
● Model interpolation could be applied to other MIR tasks that would potentially benefit from modeling several musical aspects simultaneously.
![Page 50: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/50.jpg)
Thank you!
![Page 51: Multiple pitch transcription and melody harmonization …repmus.ircam.fr/_media/dyci2/mpt-and-mh-with-pmm.pdf · harmonization with probabilistic musicological models ... double bass)](https://reader031.fdocuments.us/reader031/viewer/2022020205/5b8b444b7f8b9a9e508c476c/html5/thumbnails/51.jpg)
References[1] E. Vincent, S. Raczyński, N. Ono, and S. Sagayama, “A roadmap towards versatile
MIR,” in Proc. 11th International Conference on Music Information Retrieval (ISMIR), 2010, pp. 662–664.
[2] Simon, I., Morris, D., & Basu, S. (2008). “MySong: automatic accompaniment generation for vocal melodies.” In Proc. 26th SIGCHI Conference on Human Factors in Computing Systems (pp. 725–734).
[3] PG Music Inc. (2012, August). “Bandinabox.” http://www.pgmusic.com/.[4] Klakow, D. (1998). “Loglinear interpolation of language models.” In Proc. 5th
International Conference on Spoken Language Processing (pp. 1695–1698).[5] Allan, M., & Williams, C. (2005). “Harmonising chorales by probabilistic inference.”
Advances in Neural Information Processing Systems, 17 , 25–32.[6] Paiement, J., Eck, D., & Bengio, S. (2006). “Probabilistic melodic harmonization.” In Proc.
19th Canadian Conf. on Artificial Intelligence (pp. 218–229).[7] Kneser, R.; Ney, H., "Improved backingoff for Mgram language modeling," Acoustics,
Speech, and Signal Processing, 1995. ICASSP95., 1995 International Conference on , vol.1, no., pp.181,184 vol.1, 912 May 1995
[34] Temperley, D., & Sleator, D. (2012, August). “Harmonic Analyzer.” http://www.cs.cmu.edu/~sleator/harmonicanalysis/.