Von: ak, sekretariat
Datum: Thu, 8 Oct 2026
Betreff: [ak-discourse] Re: Einladung zum Forschungskolloquium am 13.10.2026 / 17:00 Uhr
___
Prof. Dr. Stefan Weinzierl: Das Audio-Paper. Eine neues Format für die wissenschaftliche Kommunikation
Was ist ein „Audio Paper“ – und welche neuen Möglichkeiten eröffnet es für wissenschaftliches Arbeiten und Publizieren? Der Vortrag stellt das Konzept des Audio Papers vor und zeigt an Beispielen, wie wissenschaftliche Inhalte, Klang und gesprochene Sprache miteinander verbunden werden können, und welchen Mehrwert dieses Format gegenüber klassischen wissenschaftlichen Publikationen bieten kann.
Samuel Braun: Real-Time Chord Estimation Using Structured Compact Neural Networks and Chord Decomposition
Real-time chord estimation remains a challenging task due to the absence of future context and the need for temporally stable predictions. This work presents a causal chord recognition system based on structured compact neural networks derived from the FifthNet architecture. The proposed model combines musically motivated feature representations with pitch-shift-invariant processing and recurrent pitch-class modeling to enable low-latency inference. A structured chord decomposition is employed to represent harmonic content through interpretable components, allowing flexible vocabulary scaling and improved modeling of extended chords. To address temporal fragmentation, a boundary-aware auxiliary decision model is introduced to refine frame-level predictions using differentiable change penalties while maintaining real-time constraints. Experiments on a mixture of publicly available and proprietary datasets demonstrate that recurrent pitch-class modeling substantially improves performance compared to causal convolutional networks. While real-time operation leads to reduced accuracy compared to noncausal models, the proposed temporal smoothing approach significantly reduces oversegmentation and improves overall chord recognition performance. The results indicate that structured representations combined with causal neural architectures provide an effective framework for practical real-time chord estimation.
…
Liebe Freunde der Audiokommunikation,
am kommenden Dienstag, den 13.10.2026, um 17:00 Uhr im Raum EN 324 präsentiert Samuel Braun seine Arbeit zum Thema:
„Real-Time Chord Estimation Using Structured Compact Neural Networks and Chord Decomposition“
Dazu möchten wir Sie herzlich einladen. Die Kurzzusammenfassung finden Sie am Ende dieser E-Mail.
Der Zoom-Link für unsere auswärtigen Gäste lautet:
tu-berlin.zoom.us/j/65356423211
Meeting-ID: 653 5642 3211
Kenncode: 141962
Mit herzlichen Grüßen
Kirill Andreev & Stefan Weinzierl
___
Samuel Braun: Real-Time Chord Estimation Using Structured Compact Neural Networks and Chord Decomposition
Real-time chord estimation remains a challenging task due to the absence of future context and the need for temporally stable predictions. This work presents a causal chord recognition system based on structured compact neural networks derived from the FifthNet architecture. The proposed model combines musically motivated feature representations with pitch-shift-invariant processing and recurrent pitch-class modeling to enable low-latency inference. A structured chord decomposition is employed to represent harmonic content through interpretable components, allowing flexible vocabulary scaling and improved modeling of extended chords. To address temporal fragmentation, a boundary-aware auxiliary decision model is introduced to refine frame-level predictions using differentiable change penalties while maintaining real-time constraints. Experiments on a mixture of publicly available and proprietary datasets demonstrate that recurrent pitch-class modeling substantially improves performance compared to causal convolutional networks. While real-time operation leads to reduced accuracy compared to noncausal models, the proposed temporal smoothing approach significantly reduces oversegmentation and improves overall chord recognition performance. The results indicate that structured representations combined with causal neural architectures provide an effective framework for practical real-time chord estimation.
__
Kirill Andreev
Er/ihm – he/him
Anrede/Form of address: Herr – Mr.
Sekretariat / Office
Technische Universität Berlin
Fakultät I / Inst. f. Sprache u. Kommunikation
FG Audiokommunikation / Sekr. EN-8
Faculty I / Institute for Language & Communication
Audio Communication Group / Office No. EN-8
Einsteinufer 17, 10587 Berlin
GERMANY
sekretariat@ak.tu-berlin.de
+49 (0)30 314-22236
Erreichbarkeit: Di, Do 9-17 Uhr, Fr. n. V.