Von: Marc Voigt via ak discourse
Datum: Sun, 29 Oct 2023
Betreff: [ak-discourse] Einladung zum Forschungskolloquium am 31.10.2023
Liebe Freunde der Audiokommunikation,
am kommenden Dienstag, 31.10.2023, 16:00 Uhr im Raum EN 324 präsentiert Yijun Zhao die Ergebnisse seiner Masterarbeit zum Thema
„Effect of Voice Activity Detection on Speech Enhancement“.
Dazu möchten wir Sie herzlich einladen. Eine Kurzzusammenfassung finden Sie am Ende dieser E-Mail.
Der Zoom-Link für auswärtige Gäste lautet:
https://tu-berlin.zoom.us/j/65603959784?pwd=ZEs1SkVrbFJzZ0pXeG9TSGp3YkI3UT09
Meeting-ID: 656 0395 9784
Kenncode: 20231009
Mit herzlichen Grüßen
Marc Voigt & Stefan Weinzierl
—
Yijun Zhao: Effect of Voice Activity Detection on Speech Enhancement
In speech enhancement (SE), effectively distinguishing between speech and noise is crucial for optimal performance. This thesis proposes an SE method guided by voice activity detection (VAD). A deep learning-based VAD model is used to predict speech probabilities for each audio frame. The computed speech probabilities, in conjunction with the noisy speech waveform, are then input into the waveform-to-waveform SE model, “Demucs”. This approach seeks to prioritize the speech segment within the SE model, leading to a more focused enhancement process. The evaluation of the enhanced audio files utilizes quantitative metrics, including DNSMOS and PESQ, and is complemented by subjective listening tests. Results indicate that, under specific model structures, integrating VAD probabilities into the inputs yields more stable model training and a slight improvement in enhancement outcomes.
—
Marc Voigt
IT-Administration
Technische Universität Berlin
Fakultät I – Geistes- und Bildungswissenschaften
Institut für Sprache und Kommunikation
Fachgebiet Audiokommunikation
Faculty I – Humanities and Educational Sciences
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