Master 2018 2019
Stages de la spécialité SAR
Energy Invariant Real-Time Score Following


Site : Antescofo SAS
Lieu : Antescofo SAS, 38 Rue Meslay, 75003 PARIS
Encadrant : Arshia Cont and Philippe Cuvillier
Dates :18/02/2019 au 16/08/2019
Rémunération :1000 €/mois
Mots-clés : Parcours ATIAM : Traitement du signal

Cliquer ici pour vous authentifier


Description

This Masters project inscribes itself in the context of Score Following (or Real-time Audio to Score Alignment), and within the Antescofo project, and related to both its usage context in musical creation and day-to-day music practice through the Metronaut App. In these contexts, a real-time system (i) employs user’s real-time audio signals to infer positions and musical parameters given a music score and regardless of instrument or acoustical situation (Machine Listening), and (ii) uses that information in order to anticipatively synchronize future events with heterogeneous temporal natures (Reactive Synchronous System). This project is concerned with improving the former component, the Machine Listening, in a general context.

The problematics at the core of this project is to liberate the Machine Listening from its direct or indirect dependency on the Input Signal Energy. The current design is based on a one-time Calibration mechanism that prepares the system’s observations to distinguish between musical and non-musical events. Despite its robustness, the absence of the calibration process can cause uncertainties in probabilistic observations that are not handled by the inference mechanism. The goal of this project is thus to propose models that overcome this outcoming.

There are potentially three directions to achieve this : (1) to study and propose alternative observation mechanisms, sufficiently discriminative in the usage context, without altering the main probabilistic model ; (2) propose novel probabilistic decodings that take into account incoherencies in input signal in terms of energy ; or (3) a combination of (1) and (2). During the course of this project, the candidate is expected to understand and get familiar with the Antescofo Core Machine Listening process, study and discover proximity literature for tackling the above challenge both from a signal processing and machine learning perspective, undertake experimental setups respecting the Antescofo workflow and solid evaluation procedure based on studied literature, and provide insights on possible approach to achieve energy invariant score following.

The project will kick start with an immersion in the Antescofo R&D code-base and experimental procedures both in C++ and Python and corresponding literature, followed by the establishment of a thorough bibliography for the first two approaches above, setting up experimental procedures for the validation of the literature on a small-scale setup, and propose solutions for the larger-scale system and validations thereof.

Bibliographie

Arshia Cont. A Coupled Duration-Focused Architecture for Real-Time Music-to-Score Alignment. IEEE Transactions on Pattern Analysis and Machine Intelligence, Institute of Electrical and Electronics Engineers, 2010, 32, pp.974-987.

Philippe Cuvillier. On temporal coherency of probabilistic models for audio-to-score alignment. Sound [cs.SD]. Université Pierre et Marie Curie - Paris VI, 2016. English.

Philippe Cuvillier, Arshia Cont. Coherent Time Modeling of semi-Markov Models with Application to Real-Time Audio-to-Score Alignment. Larsen, Jan and Guelton, Kevin. MLSP 2014 - IEEE International Workshop on Machine Learning for Signal Processing (2014), Sep 2014, Reims, France. IEEE, 2014.