Training Ircam's Score Follower

Abstract

This paper describes our attempt to make the Hidden Markov Model (HMM) score following system developed at Ircam sensible to past experiences in order to obtain better audio to score real-time alignment for musical applications. A new observation modeling based on Gaussian Mixture Models is developed which is trainable using a learning algorithm we would call automatic discriminative training. The novelty of this system lies in the fact that this method, unlike classical methods for HMM training, is not concerned with modeling the music signal but with correctly choosing the sequence of music events that was performed. Besides obtaining better alignment, new system's parameters are controllable in a physical manner and the training algorithm learns different styles of music performance as discussed.

Real-Time Audio-to-Score Alignment Using Particle Filter for Coplayer Music Robots

“…Most popular score following methods are based on either dynamic time warping (DTW) [14,15] or hidden Markov models (HMMs) [16,17]. Although the target of these systems is MIDI-based automatic accompaniment, the prediction of upcoming musical notes is not included in their score following model.…”

“…HMM-based methods [16][17][18] update the estimated score position for each frame of short-time Fourier transform. Although this approach can naturally assume the transients of each musical note, for example, the onset, sustain, and release, the estimation can be affected by some frames that contain unexpected signals, such as the remainder of previous musical notes or percussive sounds without a harmonic structure.…”

Aligning music audio with symbolic scores using a hybrid graphical model

“…Many researchers have treated on-line and off-line musical parsing including (Dannenberg, 1984;Vercoe, 1984;Baird, Blevins and Zahler, 1993;Puckette, 1995;Grubb & Dannenberg, 1997;Raphael, 1999;Loscos, Cano and Bonada, 1999;Cont, Schwarz and Schnell, 2004;Orio & Dechelle, 2001;Turetsky & Ellis, 2003;Soulez, Rodet and Schwarz, 2003), to name several. While many variations exist, the predominant approach seeks a best possible match by "warping" the score to fit the data using some form of dynamic programming.…”

Off-line refinement of audio-to-score alignment by observation template adaptation

“…Most previous studies rely on a prior training of this model, which can be instrumentspecific [10][11][12][13][14] or generic [16,17]. However, these methods requires relevant training data, which are not always available.…”

“…Some works use a prior learning of the observation model, with statistical [10][11][12] or template-based approaches [13,14]. In the latter approach, a template is built for each symbolic element, as the superposition of single-note templates.…”