Rhythmic Transcription of MIDI Signals Carmine Casciato MUMT

Rhythmic Transcription of MIDI Signals Carmine Casciato MUMT 611 Thursday, February 10, 2005

Uses of Rhythmic Transcription l Automatic scoring ¡ Improvisation l Score following ¡ Triggering of audio/visual components l Performance l Audio classification and retrieval ¡ Genre classification ¡ Ethnomusicology considerations ¡ Sample database management 2

MIDI Signals Unidirectional message stream at 3. 125 KHz l System Real Time Messages provide Timing Tick message l A simplification of acoustic signals l ¡ No noise, masking effects Easily retrieve note onsets, offsets, velocities, pitches l However, no knowledge of acoustic properties of sound l 3

Difficulties in Rhythmic Transcription Expressive performance vs mechanical performance l Inexact performance of notes l ¡ ¡ ¡ l Syncopations Silences Grace notes Robustness of beat tracker ¡ Can the tracker recover from incorrect beat induction? Real time implementation l (Dixon 2001) l 4

Human Limits of Rhythmic Perception Two note onsets are deemed synchronous when played within 40 ms of each other, 70 ms for > two notes l Piano and orchestral performances exhibit note onset asynchronicity of 30 -50 ms l Note onset differences of 50 ms to 2 s give rhythmic information l (Dixon 2001) l 5

Evaluation Criteria for Beat Trackers Informally - click track of reported beats added to signal l Visually marking the reporting beats l Comparing reported vs known, correct beats l (Dixon 2001) l 6

Definitions l l l Beat - “perceived pulses which are approximately equally spaced and define the rate at which notes in a piece are played” meterical, score , performance level tempo - beats per minute Inter-onset Intervals (IOI) - time intervals between note onsets (Dixon 2001) 7

Approaches - Probabilistic Frameworks Cemgil et al (2000) - Bayesian framework, using a tempogram (wavelet) and a 10 th order Kalman Filter to estimate tempo, which is a hidden state variable l Takeda et al (2002) - Hidden Markov models for fluctuating note lengths and note sequences, estimating both rhythms and tempo l Raphael (2002) - tempo and rhythm l 8

Approaches - Oscillators Period and phase that adjusts itself to synchronize to IOI input l Dannenberg and Allen (1990) - weighted IOIs and credibility evaluation based on past input l Meudic (2002) - real time implementation of Dixon l ¡ l Induce several beats and attempt to propagate them through the signal (agents), then choose the best Pardo (2004) - Oscillator, compared to Cemgil using same corpus 9

Pardo 2004 - Oscillatory Design Is a Kalman Filter (Cemgil) or oscillator better for online tempo tracking? l Performance as time series of weights, W, over T time steps l Weight of time step with no note onsets = 0, increased proportional to # of note onsets l 100 ms is minimum IOI allowed, minimum beat period l 10

Pardo 2004 Uses weighted average of last 20 beat periods, with one parameter varying degrees of smoothing l A correction parameter varies how far the period and phase of the next predicted beat is changed according to known information l A window size parameter affects how many periods may affect the current prediction l Chose 5000 random values of these three parameters, ran each triplet on 99 performances of Cemgil corpora l 11

Cemgil MIDI/Piano Corpora Four pro jazz, four pro classical, three amateur piano players l Yesterday and Michelle, fast, slow and normal, captured on a Yamaha Diskclavier l Available at www. nici. kun. nl/mmm/ l 12

Pardo 2004 - Error Measurement (Pardo 2004) • After finding best parameters values for Michelle corpus, applied same values to analysis of Yesterday corpus • Compared to Cemgil using that paper’s defined error metric, which takes into account both phase and period errors, to come up with a score 13

Comparison of Approaches (Pardo 2004) • Oscillator somewhat better than tempogram alone, • Somewhat worse than tempogram plus Kalman, yet fall within standard deviation (bracketed numbers) of Kalman scores 14

Other Considerations l Stylistic information ¡ Training l Musical of tracker importance of note ¡ Duration ¡ Pitch ¡ Velocity 15

Bibliography l l l Allen, P. , and R. Dannenberg. 1990. Tracking musical beats in real time. In Proceedings of the International Computer Music Conference 1990: 140– 3. Dixon, S. 2001. Automatic extraction of tempo and beat from expressive performances. Journal of New Music Research 30 (1): 39– 58. Meudic, B. 2002. A causal algorithm for beat-tracking. In Proceedings of Conference on Understanding and Creating Music. Pardo, B. 2004. Tempo tracking with a single oscillator. In Proceedings of the International Conference on Music Information Retrieval 2004. Raphael, C. 2002. A hybrid graphical model for rhythmic parsing. Artificial Intelligence 137: 217– 38. Takeda, H. , T. Nishimoto, and S. Sagayama. 2002. Automatic rhythm transcription from multiphonic MIDI signals. In Proceedings of the International Conference on Music Information Retrieval 2003. 16