Achtung:

Sie haben Javascript deaktiviert!
Sie haben versucht eine Funktion zu nutzen, die nur mit Javascript möglich ist. Um sämtliche Funktionalitäten unserer Internetseite zu nutzen, aktivieren Sie bitte Javascript in Ihrem Browser.

Source separation and interference reduction for automatic speech recognition in dynamic acoustic systems

Spectogram of a noisy speech signal

This project is dedicated to a holistic approach for speech enhancement and recognition in an automatic house environment. Although, in recent years many advances in the field of automatic speech recognition (ASR) were made most cannot be applied in dynamic environments since they are burdened with high latencies and computational effort. Furthermore, most research on ASR is based on clean speech data without reverberation and multiple speakers which are essential parts of the acoustic signals in an automatic house

Spectral mask calculated by a deep recurrent neural network

As a result a preprocessing system for speaker separation and interference reduction utilizing recurrent deep neuronal networks with low latency and small computational complexity is considered. To achieve best possible speech quality for the ASR system in an assumed automatic house environment a multi channel approach to speech enhancement using spectral masks is combined with different approaches to speaker separation. The resulting speech signal is examined on its usability in speech recognition and other applications in an automatic house.

This project is a transfer project with the voice INTER connect GmbH financed by the “Deutsche Forschungsgemeinschaft” (DFG).

Ansprechpartner

Jens Heitkämper

Nachrichtentechnik (NT)

Forschung & Lehre

Jens Heitkämper
Telefon:
+49 5251 60-5288
Fax:
+49 5251 60-3627
Büro:
P7.2.06

Die Universität der Informationsgesellschaft