Reduced energy consumption and improved accuracy for distributed speech recognition in wireless environments

The central theme of this dissertation is the study of a multimedia client for pervasive wireless multimedia applications. Speech recognition is considered as one such application, where the computational demands have hindered its use on wireless mobile devices. Our analysis considers distributed speech recognition on hardware platforms with PDA-like functionality (i.e. wireless LAN networking, high-quality audio input/output, a low-power general-purpose processing core, and limited amounts of flash and working memory). We focus on quality of service for the end-user (i.e. ASR accuracy and delay) and reduced energy consumption with increased battery lifetimes. We investigate quality of service and energy trade-offs in this context. We present software optimizations on a speech recognition front-end that can reduce the energy consumption by over 80% compared to the original implementation. A power on/off scheduling algorithm for the wireless interface is presented. This scheduling of the wireless interface can increase the battery lifetime by an order of magnitude. We study the effects of wireless networking and fading channel characteristics on distributed speech recognition using Bluetooth and IEEE 802.11b networks. When viewed as a whole, the optimized distributed speech recognition system can reduce the total energy consumption by over 95% compared to a client-side ASR implementation. We present an interleaving and loss concealment algorithm to increase the robustness of distributed speech recognition in a burst error channel. This improvement allows a decreased reliance on error protection overhead, which can provide reductions in transmit energy of up to 46% on a Bluetooth wireless network. The findings presented in this dissertation stress the importance of energy-aware design and optimization at all levels for battery-powered wireless devices.