978-1-4244-1694-3/08/$25.00 ©2008 IEEE Abstract In this paper, we present the design and implementa- tion of a distributed sensor network application for embedded, isolated-word, real-time speech recognition. In our system design, we adopt a parameterized-data- flow-based modeling approach to model the functional- ities associated with sensing and processing of acoustic data, and we implement the associated embedded soft- ware on an off-the-shelf sensor node platform that is equipped with an acoustic sensor. The topology of the sensor network deployed in this work involves a clus- tered network hierarchy. A customized time division multiple access protocol is developed to manage the wireless channel. We analyze the distribution of the overall computation workload across the network to improve energy efficiency. In our experiments, we dem- onstrate the recognition accuracy for our speech recog- nition system to verify its functionality and utility. We also evaluate improvements in network lifetime to dem- onstrate the effectiveness of our energy-aware optimi- zation techniques. 1. Introduction Speech recognition involves converting acoustic signals, captured by a microphone or an acoustic sen- sor, to a set of words. Then, these words are compared with some pre-defined words and some sort of indica- tion is given if there is a match. The recognized words can then be analyzed and used for back-end applica- tions such as command and control, commercial infor- mation retrieval, and linguistic processing for speech understanding. Figure 1 presents a design flow for basic speech recognition systems. From an embedded system design aspect, a major design challenge for speech recognition is to assure the processing of large amounts of data in real-time. Vari- ous prior efforts on embedded speech recognition, e.g. [1, 12, 13], focused on implementing various speech recognition algorithms on embedded platforms, such as programmable digital signal processors (PDSPs) and comparing their performance. These efforts typically have not explored further optimization for real-time operations and energy usage beyond what is already available through a standard PDSP-based design flow. These existing design approaches therefore are not fully suited for heavily resource-limited, distributed systems, such as wireless sensor networks. A wireless (distributed) sensor network (WSN) sys- tem is composed of resource-limited sensor nodes, which consist of components for sensing, data process- ing, and wireless communication (e.g., see [7]). WSN systems have a variety of potential applications [9], such as environmental monitoring and intrusion detec- tion. Sensor nodes are often deployed in inaccessible or, in the case of certain military and security-related appli- cations, dangerous areas and communicate with each other through self-organizing protocols. To maximize the useful life of these systems, power consumption must carefully be considered during sensor node design. Integrating speech recognition into a WSN system enables a new class of applications for speech recogni- tion that have distributed configurations. We refer to such speech-recognition-equipped WSN systems as dis- tributed automatic speech recognition (DASR) systems. The DASR system developed in this paper is an iso- lated-word and speaker-dependent speech processing system, where templates of extracted coefficients of words have to be created and stored at a central node. The system functionality is to have all sensor nodes col- lect speech data within their sensing ranges, and trans- mit this data periodically — in the form of recognized words (or simple indicators for the absence of any words) — to the central node. Any application-specific analysis and usage of the recognized words is handled as back-end processing on the central node. Based on different requirements on recognition accuracy, we describe two practical application scenar- ios in which our developed DASR system can be applied. The first scenario involves using a DASR sys- tem as a speech-based command and control system in Design and Optimization of a Distributed, Embedded Speech Recognition System Chung-Ching Shen, William Plishker, and Shuvra S. Bhattacharyya Dept. of Electrical and Computer Engineering, and Institute for Advanced Computer Studies University of Maryland at College Park, USA {ccshen, plishker, ssb}@umd.edu Figure 1. Basic design flow for automatic speech recognition systems. Front-end Signal Processing Back-end Applications Input speech Feature Vectors Feature Analysis Sensing / Sampling Start Detection / Framing Recognition In Proceedings of the International Workshop on Parallel and Distributed Real-Time Systems, Miami, Florida, April 2008.