Efficiency of blind source separation in a real room

The present study is concerned with the effectiveness of the blind source separation (BSS)
in real acoustical environment measured by means of a speech reception threshold, SRT. BSS
is a multisensoral method that leads to a signal extraction from a mixture of sounds. The performance
of such algorithms are most often described by means of the increase in signal-tonoise
ratio, SNR. However, the SNR is not an appropriate measure for speech enhancement,
since the increase in SNR does not lead directly to the improvement of the speech intelligibility.
Thus, the relationship between an increase in SNR and speech intelligibility is not
straightforward.
This work shows some experiments in which the speech masked by a babble noise, music
or concurrent speech was separated out using an algorithm for convolutive BSS. The SRT before
and after the separation was measured for subjects with audiologically normal hearing.
All recordings were carried out in a small office room using an array of two microphones.
Two spatially separated loudspeakers were used as sources of signals: target speech and disturbances.
The results of the experiment revealed a marked improvement in the speech intelligibility:
the decrease in SRT reached about 7 to 24 dB in individual cases.