Score-Informed Audio Parameterization , by Sebastian Ewert, 10am (+Audiosig talks)

Tuesday, april the 10th, 9:15pm
Telecom ParisTech (Dareau)
Room: DA-006 (Vitrine de la recherche)

Author: Sebastian Ewert, Université de Bonn

Audiosig  also organize some talks about the research themes of the group. This is the program:

9:15 Cyril Joder, Conditional Random Fields for Audio-to-Score Alignment
10:00 Sebastian Ewert : Score-Informed Audio Parameterization
11:10 Thomas Fillon, Audio Segmentation + Beat/Downbeat Estimation
11:45 Rémi Foucard, Multi-Tag Audio Classification with Boosting
12:20 Mounira Maazaoui, Blind Source Separation for Robot Audition using Microphone Array

Abstract of  Sebastian Ewert talk:
In recent years, the processing of audio recordings by exploiting musical knowledge as specified by a musical score has turned out to be a promising research direction. Here, one assumes that, additionally to the audio recording to be analyzed, one is given a MIDI file (representing the score) of the same underlying piece of music. The note event information specified by the MIDI file can then be used to support audio analysis tasks such as source separation or instrument equalization. In this talk, we consider the problem of score-informed audio parameterization with the objective to successively adapt and enrich the note event information provided by the MIDI file to explain the given audio recording. More precisely, our goal is to parameterize the spectrogram of the audio recording by exploiting the score information. Our parameterization approach works iteratively proceeding in several steps. In the first step, we compute a temporal alignment between the MIDI file and the audio recording. Since the alignment accuracy is of major importance, we employ a refined synchronization method that exploits onset information. In the next steps, we  successively adapt model parameters referring to dynamics, timbre, and instrumentation such that the spectrogram described by the model reflects the audio spectrogram as accurately as possible.

Comments are closed.