Jörn Anemüller
and Birger Kollmeier, `Adaptive separation of acoustic sources for anechoic
conditions: A constrained frequency domain approach', Speech Communication
39(1-2), pp. 79-95, 2003.
G. Jongen, J. Anemüller,
D. Bolle, A. C. C. Coolen, C. Perez-Vicente, `Coupled dynamics of fast spins
and slow exchange interactions in the xy spin glass', Journal of Physics
A: Mathematical and General, 34(19), pp. 3957-3984, 2001.
Jörn Anemüller,
`Across-frequency processing in convolutive blind source separation', Ph.D.
thesis, Dept. of Physics, University of Oldenburg, Oldenburg, Germany, 2001.
Jörn Anemüller
and Birger Kollmeier, `Convolutive blind source separation of speech signals
based on amplitude modulation decorrelation', Journal of the Acoustical Society
of America, 108, p. 2630, 2000.
Jörn Anemüller
and Birger Kollmeier, `Amplitude modulation decorrelation for convolutive
blind source separation', in: P. Pajunen and J. Karhunen (Eds.), `Proceedings
of the second international workshop on independent component analysis and
blind signal separation', June 19-22, 2000, Helsinki, Finland, pp. 215-220..
Abstract:
The problem of blind
separation of a convolutive mixture of speech signals is considered. Signal
separation is performed in the frequency domain.
Based on observations
from amplitude spectrograms of speech signals, the notion of amplitude modulation
correlation (`AMCor') across different frequency channels is introduced.
From the corresponding principle of amplitude modulation decorrelation, a
novel cost--function and an algorithm for convolutive blind source separation
are derived. The algorithms' main features are discussed. Successful separation
of synthetic data and of real--room recordings of speech is performed. The
results of the latter are compared to the performance of previous algorithms
on the same data.
Audio examples are
available from the authors' web page.
Jörn Anemüller,
Michael Kleinschmidt und Birger Kollmeier, `Blinde Quellentrennung als Vorverarbeitung
zur robusten Spracherkennung', in: `Fortschritte der Akustik: DAGA 2000',
Oldenburg, Germany, March 20-23, 2000, in press.
Zusammenfassung:
In diesem Beitrag evaluieren
wir den Nutzen blinder Quellentrennung als Vorverarbeitungsstufe zum Zwecke
robuster automatischer Spracherkennung. Blinde Quellentrennung (QT) ist eine
Signalverarbeitungstechnik, die es ermöglicht, aus mehreren Aufnahmen
akustischer Überlagerungen (etwa Sprache im Störgeräusch) die
zugrunde liegenden Quellsignale (Sprache getrennt vom Störgeräusch)
zu rekonstruieren. Ein spezieller Algorithmus für QT in verhallter Umgebung
ist bereits vorgestellt worden (Anemüller, 1999). Eine potentielle Anwendung
solcher Algorithmen besteht in der Störgeräuschbefreiung für
die robuste automatische Spracherkennung.
Das Perzeptionmodell
(PEMO) nach Dau et al. (Dau et al., 1996) wurde bereits zur Merkmalsextraktion
in der automatischen Spracherkennung verwendet. Insbesondere in Kombination
mit Neuronalen Netzen hat diese gehörgerechte Vorverarbeitung zu einer
robusten Erkennungsleistung im Störgeräusch geführt (Tchorz,
1999). Wir kombinierten den QT-Algorithmus mit einem Einzelworterkennungssystem
auf Basis des PEMO, um eine weitere Verbesserung der Erkennungsleistung zu
erreichen. Zur Evaluation vergleichen wir die Erkennungsraten bei QT-Vorverarbeitung
mit denen ohne Vorverarbeitung und mit alternativen Störgeräuschunterdrückungssystemen.
Berücksichtigt werden hierbei Aufnahmesituationen in verhallter und unverhallter
Umgebung und bei unterschiedlichen Signal-Rausch Abständen.
Jörn Anemüller,
`Correlated modulation: a criterion for blind source separation', Joint meeting
of the Acoustical Society of America and the European Acoustics Association,
Berlin, Germany, March 14-19, 1999.
Abstract:
The problem of blindly
separating convolutive mixture of modulated signals is considered. Spectrograms
of the signals are computed and separation is performed in the frequency domain.
A new algorithm for blind source separation is proposed, which is based on
correlated modulation in the sources' different frequency channels. For example,
speech contains correlated modulation in different frequency regions. The
algorithm successfully separates mixtures of modulated artificial signals
and of speech.
Jörn Anemüller
and Tino Gramss, `On-line blind separation of moving sound sources', in: `ICA
99: First international workshop on independent component analysis and signal
separation', Aussois, France, January 11-15, 1999, pp. 331-334.
Abstract:
In this paper, we propose
a method for the on-line blind separation of sound sources in the case where
the mixing filters have a delta-shaped impulse response. Our algorithm works
entirely in the frequency domain and exhibits fast convergence due to cross-frequency
couplings.
Specific problems related
to on-line separation of running speech are discussed. The algorithm performs
successful separation of digitally mixed speech signals and of signals from
a moving and a standing speaker recorded in an anechoic chamber.
Jörn Anemüller
and Tino Gramss, `A neural network for sound source separation', in: "Psychophysics,
Physiology and Models of Hearing", Oldenburg, Germany, August 31 - September
4, 1998, pp. 255-258.
Abstract:
In real-world situations,
we are usually confronted with the simultaneous presence of multiple sound
sources. However, the brain is able to separate out particular source signals
and suppress the remaining ones. An adaptive algorithm is proposed which mimics
this capability by exploiting the mutual statistical independence of the
source signals. We improve on existing algorithms from the field of independent
component analysis by moving to the frequency domain and by introducing cross-frequency
couplings. Low frequency components of the signals are used to obtain a first
estimate of the source parameters while high frequency components improve
on the accuracy. The resulting algorithm successfully separates a mixture
of speaker signals recorded in an anechoic chamber. Since the algorithm works
iteratively, it can also be applied to non-stationary signals. This is demonstrated
by separating signals from a moving and a standing speaker.
Jörn Anemüller
and Tino Gramss, `Blinde akustische Quellentrennung im Frequenzbereich', in:
"Fortschritte der Akustik: DAGA 98", Zürich, Switzerland, March 23-27,
1998, pp. 350-351.
