Given ONLY ONE microphone (i.e. one input) containing speech recordings from two people, could we "separate" each person's speech?

If we have a look at the log spectrogram of the "mixture" (i.e. recordings from the ONLY ONE microphone) and the element-wise maximum of log spectrograms of each "source" (i.e. speech from each person),

Figure 1      The Log-Max Approximation

Full Screen View

From Top to Bottom, Log spectrogram of mixture, Element-wise maximum of log spectrograms of sources, and differences between two plots

these spectrograms are almost the same!

The idea of Refiltering is to use "masking signals" to isolate single sources from the mixture and reconstruct signals given some "prior information" (or "Training").

Due to limited disk quotas (department system admin has already kindly increased mine once), sound files in wave format are not available at the moment. They will be on-line when I finish writing up my report.

For now, please see the following log spectrogram plots.

Figure 2      Mixture

Figure 3      Estimation 1

Figure 4      Estimation 2

For comparisons, here are clean recordings from Source 1 and Source 2.

To save you some time doing "image analysis" on sources and estimations, here are differences between them:

Differences between source 1 and estimation 1

Differences between source 2 and estimation 2

M.J. Reyes-Gomez, D. Ellis, N. Jojic (2004), Multiband Audio Modeling for Single Channel Acoustic Source Separation, Proceedings of ICASSP-04, Montreal, May 2004

S.T. Roweis (2003), Factorial Models and Refiltering for Speech Separation and Denoising, Proceedings of Eurospeech 2003, Geneva, pp.1009-1012

S.T. Roweis (2002), One Microphone Source Separation, Neural Information Processing Systems 13, pp.793-799

More references are available in Selected Literature