A study on subspace-based estimation of STFT real and imaginary modulation signals for speech enhancement
Author(s)
George, Aidan EW
Pickersgill, Christine
Schwerin, Belinda
So, Stephen
Year published
2016
Metadata
Show full item recordAbstract
In this paper, we investigate the advantages of applying the subspace approach within short-time Fourier transform-based (STFT) modulation domains over the conventional sub-space algorithm for single-channel speech enhancement. Specifically, we consider the subspace processing of the modulations of STFT real and imaginary (RI) parts and the modulations of the STFT magnitude values. Speech enhancement experiments were performed on the Noizeus speech corpus that had been corrupted by varying levels of white and coloured noise. The results indicate that the modulation domain-based subspace methods outperformed the conventional ...
View more >In this paper, we investigate the advantages of applying the subspace approach within short-time Fourier transform-based (STFT) modulation domains over the conventional sub-space algorithm for single-channel speech enhancement. Specifically, we consider the subspace processing of the modulations of STFT real and imaginary (RI) parts and the modulations of the STFT magnitude values. Speech enhancement experiments were performed on the Noizeus speech corpus that had been corrupted by varying levels of white and coloured noise. The results indicate that the modulation domain-based subspace methods outperformed the conventional subspace approach in both enhanced speech quality and intelligibility. Objectively, the RI-modulation subspace method was found to have a slight quality and intelligibility advantage over the magnitude modulation subspace method. In blind subjective testing on speech quality, the modulation subspace methods were significantly preferred over the conventional time-domain-based subspace method.
View less >
View more >In this paper, we investigate the advantages of applying the subspace approach within short-time Fourier transform-based (STFT) modulation domains over the conventional sub-space algorithm for single-channel speech enhancement. Specifically, we consider the subspace processing of the modulations of STFT real and imaginary (RI) parts and the modulations of the STFT magnitude values. Speech enhancement experiments were performed on the Noizeus speech corpus that had been corrupted by varying levels of white and coloured noise. The results indicate that the modulation domain-based subspace methods outperformed the conventional subspace approach in both enhanced speech quality and intelligibility. Objectively, the RI-modulation subspace method was found to have a slight quality and intelligibility advantage over the magnitude modulation subspace method. In blind subjective testing on speech quality, the modulation subspace methods were significantly preferred over the conventional time-domain-based subspace method.
View less >
Conference Title
2016 10TH INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND COMMUNICATION SYSTEMS (ICSPCS)
Subject
Signal processing