Channel Combination Techniques for
                Far Field Automatic Speech Recognition

                            John McDonough
                 Institut für Theoretische Informatik
                        Universität Karlsruhe
                         Karlsruhe, Germany

Interest within the automatic speech recognition (ASR) research
community has recently focused on the recognition of speech captured
with one or more microphones located in the medium field, rather than
being mounted on a headset and positioned next to the speaker's
mouth. The capacity to reliably recognize such speech is a primary
requirement in making ASR a viable modality for so-called
\emph{ubiquitous computing}.  Such far field ASR is a natural
application for the use of multiple microphones whose signals can be
combined in different ways: On the signal side, combination can be
accomplished by beamforming using a microphone array with a known
geometry, or by blind channel combination techniques if the geometry
is unknown. On the word hypothesis side, combination can be achieved
through confusion network combination. In this talk, we describe the
components of a far field ASR system, including the source localizer,
beamformer, speech activity detector, and ASR engine. We also compare
the effectiveness of the several signal combination techniques, and
compare their performance to that achieved with a close talking
microphone. Finally, we describe a set of extensions modules to the
Python scripting language that implement the algorithms described
here.

Joint work with Hazim Kemal Ekenel, Tobias Gehrig, Ulrich Klee, Kai Nickel and Matthias Wölfel