Modeling and Synthesis of Expressive Performance

TitleModeling and Synthesis of Expressive Performance
Publication TypeConference Paper
Year of Publication2007
Conference NameDigital Music Research Network Workshop
AuthorsKersten, S., Maestre E., & Ramirez R.
AbstractIn this poster we present a systematic approach to applying expressive performance models to non-expressive score transcriptions and synthesizing the results by means of concatenative synthesis. We first build models of expressive performance by automatically transcribing a collection of audio recordings of real Jazz saxophone performances and obtaining a symbolic representation of the musician’s expressive performance. This representation is used to derive a model consisting of inductive logic rules, that transforms a non-expressive, symbolic input score into a score annotated with expressivity hints. The enriched, expressive score is used as input to a concatenative synthesizer. A large database of individual saxophone notes extracted from complete expressively performed phrases is searched with an enhanced Viterbi algorithm to find the closest match for a given input score note in terms of pitch, duration, timbre and musical context. The phrase segments corresponding to notes are transformed in time and pitch by means of Spectrals Modeling Synthesis (SMS) and concatenated in order to yield a high fidelity rendering of the expressive input score. In order to cut down the cost of the sample search per input score note, the sample database is divided into groups of similar samples by offline clustering, based on intra-note (e.g. spectral) features. From this clustering an additional model is created that can be applied to each input score note and annotates the note with a cluster number, which in turn determines the candidate list during sample search. During the poster session we intend to demonstrate the expressive performance system software.
intranet