Von: Voigt, Marc via ak-discourse
Datum: Mon, 31 Aug 2020
Betreff: [ak-discourse] Einladung zum Forschungskolloquium am 01.09.2020 – Videokonferenz
Liebe Kollegen und Studierende, liebe Interessenten an Veranstaltungen
am Fachgebiet Audiokommunikation,
am kommenden Dienstag 1.09.2020, 16:00 Uhr (Zoom) wird Roman Gebhardt seine Forschung zum Thema „Visual and Auditive Explanations of Deep Music Emotion Recognition via Gradient-based Localization“ in deutscher Sprache vorstellen.
Im Anschluß wird Maximilian Weber seine Masterarbeit zum Thema „A Framework for Audio-Tactile Signal Translation“ in englischer Sprache vortragen.
Dazu möchten wir Sie sehr herzlich einladen. Eine Kurzzusammenfassung darüber finden Sie, wie immer, am Ende dieser E-Mail.
Ich kann im Moment leider kein Zoom-Meeting anlegen, die Zugangsdaten werden daher morgen kurzfristig versendet.
Aufgrund der aktuellen Gesundheitslage findet dieses Forschungskolloquium auch als Videokonferenz statt, hierfür nutzen wir die Software Zoom in einer für die TU Berlin angepassten Version, installieren sie dazu bitte die Client-Software für ihr System, die ausführlichen Zugangsdaten werden morgen versendet.
Roman Gebhardt: Visual and Auditive Explanations of Deep Music Emotion Recognition via Gradient-based Localization
Our explorative study aims at analysing and interpreting patterns in audio signals learned by a convolutional recurrent neural network (CRNN) on a task of music emotion recognition (MER). With the gradient-weighted class activation mapping (Grad-CAM) technique we highlighted those regions in the input which are most relevant for the network’s prediction. We show that the network, which was trained to predict mutually exclusive emotion classes represented by the four quadrants of the valence-arousal plane, focuses on different frequency regions for the prediction of each class. We demonstrate that for different class predictions, the network puts significantly higher attention to certain instruments compared to other such as drums for the second quadrant and voice for the fourth one. To further corroborate our findings we present a method to sonify highlighted regions in the spectrogram and return the resulting audio files. Finally, we show how the predictions are affected by removing different instruments from the mixed signal. Our work shows that the learned features appear to be intuitively interpretable.
Maximilian Weber: A Framework for Audio-Tactile Signal Translation
To enable a transition away from primitive, buzzing vibrations towards an new generation of wideband vibrotactile display systems requires strategies, standards and tools for designing, storing and transmitting tactile stimuli signals and patterns. Complicating the matter is a variety of different tactile display technologies and resulting display system variance due to missing industry standards.
Due to the early sensory integration of both the auditory and vibrotactile modalities, and the resulting perceptual similarities, it appears to be feasible to translate auditory to vibrotactile stimulus signals and use audio material as a starting point for the design of tactile stimuli. On this basis, this work aimed to validate a novel audio-tactile signal translation method by measuring the coherence of both unimodal and bimodal vibrotactile stimuli towards their (non-musical) auditory sources, while expanding on signal processing methods discussed in previous works.
Viele Grüße
Marc Voigt
—
Marc Voigt
IT-Administration
Technische Universität Berlin
Fakultät I – Geistes- und Bildungswissenschaften
Institut für Sprache und Kommunikation
Fachgebiet Audiokommunikation
Faculty I Humanities
Institute of Speech and Communication
Audio Communication Group
Einsteinufer 17c, 10587 Berlin
GERMANY
Telefon: +49 (0)30 314-25557
Telefax: +49 (0)30 314-21143
marc.voigt@tu-berlin.de
—
Mailinglist AK discourse / TU Berlin – Fachgebiet Audiokommunikation
https://lists.tu-berlin.de/mailman/listinfo/ak-discourse