Past talks

Speaker: Shayenne da Luz Moura
Date and time: Monday, April 23, 2018 - 17:00
Place: Auditório Antonio Gilioli - Bloco A

Humans can easily identify portions of singing voice in an audio with a mixture of sound sources. However, trying to identify such segments computationally is not a trivial task. This seminar will present the fundamentals of the problem of singing voice detection in polyphonic audio signals, a brief description of the techniques used to solve it, and its applications in other tasks of music information retrieval (MIR). Finally, some challenges will be highlighted regarding the performance improvement in the automatic detection of segments with singing voice.

(video presentation in portuguese)

Speaker: Emilios Cambouropoulos, Aristotle University of Thessaloniki (Greece)
Date and time: Friday, September 1, 2017 - 11:00
Place: Jacy Monteiro Auditorium, IME/USP

The emergence of musical patterns via repetition/similarity is paramount in making sense and understanding music. Yet, despite the efforts made towards its systematic description, musical similarity remains an elusive concept, resisting robust formalisation. Why does the introduction of well-established powerful pattern matching techniques (exact or approximate) in the musical domain, usually ends up with rather limited/partial/fragmentary results? Why is it so difficult to create a general model of musical similarity that may capture musically and cognitively plausible patterns? In this presentation, we will focus on three sources of difficulty in describing musical similarity. Firstly, it is not always easy, to get a musical sequence per se on which to apply pattern matching techniques; especially in non-monophonic music (i.e., most music), it is anything but trivial to derive cognitively meaningful auditory images/streams within which patterns may emerge. Secondly, it is most important to decide how a coherent sequence of musical entities may be represented; representation in music is complex due to the multi-dimensional and hierarchic nature of musical data. Thirdly, it is vital to define the nature of a certain similarity process, as special models may have to be devised (rather than use of standard off-the- shelf algorithms). In this presentation, examples and techniques from recent research on musical pattern discovery, in melodic, harmonic and rhythmic contexts, will be presented to highlight the importance of looking in detail at the musical and cognitive aspects of music pattern discovery tasks before attempting to use/develop specific pattern matching algorithms.

Emilios Cambouropoulos is Associate Professor in Musical Informatics at the School of Music Studies, Aristotle University of Thessaloniki. He studied Physics, Music, and Music Technology before obtaining his PhD in 1998 on Artificial Intelligence and Music at the University of Edinburgh. He worked as a research associate at King’s College London (1998-1999) on a musical data-retrieval project and was employed at the Austrian Research Institute for Artificial Intelligence (OeFAI) in Vienna on the project Artificial Intelligence Models of Musical Expression (1999-2001). Recently he was principal investigator for the EU FP7 project Concept Invention Theory COIVENT (2013-2016). His research interests cover topics in the domain on cognitive and computational musicology (CCM Group - and has published extensively in this field in scientific journals, books and conference proceedings. Homepage:

(video presentation in portuguese)

Speaker: Thais Fernandes Rodrigues Santos, UFMG
Date and time: Wednesday, August 2, 2017 - 16:00
Place: CCSL Auditorium

(Abstract available in Portuguese only)

Músicos manipulam, conscientemente, diferentes parâmetros acústicos, com o objetivo de expressar suas ideias musicais para os ouvintes e para outros músicos. Eles performam, ainda, movimentos físicos que, embora sejam essenciais para a produção de som no instrumento musical, possuem estreita relação com as intenções artísticas dos performers.

O termo ancillary gesture (gestos auxiliar) caracteriza os movimentos que fazem parte da performance, mas que não necessariamente produzem o som. Gestos aparecem ao longo da performance e, consequentemente, assumem a intenção de comunicar.

Nesta comunicação, apontarei para a complexa inter-relação entre os gestos executados pelos músicos e a organização das frases musicais interpretadas por eles, definidas a partir de análises de parâmetros acústicos manipulados, além de relatos de expressão de suas intenções musicais. Este trabalho pretende mostrar também que os gestos auxiliares não são aleatórios, uma vez que os mesmos são recorrentes e estáveis, em todas as performances estudadas.

(video presentation in portuguese)

Speaker: Roberto Piassi Passos Bodo
Date and time: Tuesday, March 28, 2017 - 13:00
Place: Auditório do CCSL

Despite the complexity of assigning genres to songs, automatic musical genre classification of large databases is of great importance for Music Information Retrieval tasks. Historically, several Machine Learning techniques were applied using audio features extracted from audio files, scores and even the lyrics of the songs.

In this talk we will present the results of experiments performed with the GTZAN dataset, using global audio features (extracted with the LibROSA library) and traditional classification algorithms (implemented by scikit-learn).

Source code:

(video presentation in portuguese)

Speaker: Antonio Goulart
Date and time: Tuesday, December 13, 2016 - 15:00
Place: Room B-7

(Abstract available in Portuguese only)

Neste seminário vamos abordar o projeto de efeitos de áudio baseados em decomposições que levam um sinal proveniente de um instrumento musical para a representação AM/FM. A decomposição AM/FM produz um par de sinais também no domínio do tempo, que representam o envelope (porção AM) e a frequência instantânea (porção FM) do sinal analisado. Estes dois sinais atuam em conjunto e podem recriar o sinal original caso utilizados para modular um oscilador senoidal em amplitude e em frequência. Por outro lado, a manipulação individual das porções AM e FM oferece novas possibilidades para processamento de sinais e implementação de efeitos musicais. Neste trabalho discutimos aspectos sobre diferentes técnicas para a decomposição, baseadas na Transformada de Hilbert em conjunto com o Sinal Analítico ou no Operador de Energia de Teager-Kaiser em conjunto com o Algoritmo de Separação de Energia. Apresentamos novos efeitos de áudio baseados em filtragem simples, mapeamento não-linear e manipulações das porções AM e FM inspiradas em efeitos clássicos, entre outros efeitos. Implementações para operação em tempo-real são apresentadas e discutidas. Também apresentaremos uma avaliação experimental da aplicação dos efeitos propostos, baseada na análise do comportamento de descritores de áudio relacionados com amplitude (variação dinâmica) e frequência (variação espectral) de sinais.

(video presentation in portuguese)


Subscribe to Upcoming talks Subscribe to Past talks