“Audio, speech, and music processing and content analysis by non-negative spectrogram decomposition”

Abstract:
Many everyday sounds, for example speech in noisy environments or music, are mixtures
of multiple sources. While powerful tools have been developed for the analysis and pro-
cessing of isolated sounds, their capability is rather limited in the case of mixtures of
sounds. Recently, spectrogram factorization methods have been widely applied in many
audio analysis and processing tasks. This presentation rst review basic concepts of spec-
trogram factorization. Both unsupervised and supervised cases are dicussed, as well as
more advanced models than the simple matrix factorization. We discuss how the above
methods can be used for speech reparation, robust speech recognition, singing separation
and analysis, music instrument recognition, and everyday sound event detection. Audio
and video demonstrations will be given.

Biography:
Tuomas Virtanen is a research fellow and an adjuct professor at the Department of Signal
Processing, Multimedia Research Group, Tampere University of Technology , where he is
steering the Audio Research Team. His research area is computational analysis of audio,
especially sound source separation, which has several applications in the analysis, editing
and manipulation of audio signals. These include for example structured audio coding,
automatic transcription of music, and noise-robust automatic speech recognition.