The performance of state-of-the-art automatic speech
recognition (ASR) systems tends to decrease when the distance
between the speaker’s mouth and the microphone
grows, due to both noise and reverberation (Wo¨ lfel and
McDonough, 2009). In many situations the use of closetalking
microphones is not possible or practical, so a different
solution is required. The use of multiple distant-talking
microphones provides several options that may help to
solve this problem.
In this work, we assume a practical, cost-effective and
unconstrained multi-microphone scenario, where the
microphones are arbitrarily located and may show a variety
of characteristics. For instance, in a meeting room,
some microphones may be hanging on the walls, others
standing on the table, or they may be built in the personal
communication devices of the meeting participants.
Moreover, some of them may be omnidirectional, others
directional or noise-canceling, etc. In such situation, where
the positions of the microphones are either not known or
fixed, the application of commonly used multi-microphone
approaches, like array processing (Brandstein and Ward,
2001), becomes difficult.
An alternative is provided by channel selection (CS).
Before any processing, the degree of signal distortion differs
among the channels, depending on the microphone
position and characteristics. Even if speech enhancement
is applied, the processed speech signals will not be distorted
equally, so some of them may be decoded with less recognition
errors than others. Consequently, the ASR system
may benefit if signals of higher quality are selected for further
processing. To do so, a measure of distortion, or a
measure of how well recorded or enhanced signals fit the
set of acoustic models of the ASR system is needed.
As the word error rate (WER) is unknown during recognition,
the main problem is to develop a measure, that
allows to rank the channels in a way as close as possible
to the WER based ranking. In this paper, several new measures
are presented and compared, in terms of recognition