This article briefly summarises the author's research on automatic performance, started at CSC (Centro di Sonologia Computazionale, University of Padua) and continued at TMH-KTH (Speech, Music Hearing Department at the Royal Institute of Technology, Stockholm). The focus is on the evolution of the architecture of an artificial neural networks (ANNs) framework, from the first simple model, able to learn the KTH performance rules, to the final one, that accurately simulates the style of a real pianist performer, including time and loudness deviations. The task was to analyse and synthesise the performance process of a professional pianist, playing on a Disklavier. An automatic analysis extracts all performance parameters of the pianist, starting from the KTH rule system. The system possesses good generalisation properties: applying the same ANN, it is possible to perform different scores in the performing style used for the training of the networks. Brief descriptions of the program Melodia and of the two Java applets Japer and Jalisper are given in the Appendix. In Melodia, developed at the CSC, the user can run either rules or ANNs, and study their different effects. Japer and Jalisper, developed at TMH, implement in real time on the web the performance rules developed at TMH plus new features achieved by using ANNs.

In this investigation we use the term musical punctuation for the marking of melodic structure by commas inserted at the boundaries that separate small structural units. Two models are presented that automatically try to locate the positions of such commas. They both use the score as the input and operate with a short context of maximally five notes. The first model is based on a set of subrules. One group of subrules mark possible comma positions, each provided with a weight value. Another group alters or removes these weight values according to different conditions. The second model is an artificial neural network using a similar input as that used by the rule system. The commas proposed by either model are realized in terms of micropauses and of small lengthenings of interonset durations. The models are evaluated by using a set of 52 musical excerpts, which were marked with punctuations according to the preference of an expert performer.

Articulation strategies applied by pianists in expressive performances of the same core are analysed. Measurements of key overlap time and its relation to the inter-onset-interval are collected for notes marked legato and staccato in the first sixteen bars of the Andante movement of W.A. Mozart's Piano Sonata in G major, K 545. Five pianists played the piece nine times. First, they played in a wa that they considered "optimal". In the remaining eight performances they were asked to represent different expressive characters, as specified in terms of different adjectives. Legato,staccato, and repeated notes articulation applied by the right hand were examined by means of statistical analysis. Although the results varied considerably between pianists, some trends could be observed. The pianists generally used similar strategies in the rendering intended to represent different expressive characters. legato was played with a key overlap ratio that depended on the inter-onset-interval (IOI). Staccato tones had approximate duration of 40% of the IOI. Repeated notes were played with a duration of about 60% of the IOI. The results seem useful as a basis for articulation rules in grammars for automatic piano performance.