A major problem in applications such as military communications is the need for transmission of audio over very low bandwidth channels. Speech communication applications employ coding algorithms to effectively deal with limited bandwidth. Recently, the desire for hands-free information access over such channels has generated new interest in robust speech recognition in tactical environments. Due to loss of spectral information, speech coding has a negative effect on the accuracy of speech recognition systems. As the demand for mobile telephony increases, further reductions in codec bit rates are expected, which will further degrade the performance of voice interfaces.

Several approaches have been proposed to deal with this problem. These approaches typically involve either regenerating the speech signal (decoding) prior to applying noise compensation or channel adaptation techniques. Once the robustness of the system to speech compression algorithms has been established, issues such as real-time decoding in limited resources become important. There are numerous strategies for building a real-time system such as using smaller acoustic models, tighter pruning, and fast search algorithms.

In this collaborative research project, we are exploring many of these operational issues. We are focused on three tasks: investigating performance as a function of the speech compression algorithm, real-time recognition systems for vocabularies on the order of 1,000 words in client/server type architectures (e.g., DARPA Communicator), and robustness to variations in the channel and acoustic environments (SPINE).