Speech Recognition for Smart Homes

Abstract

When Christopher Sholes created the QWERTY keyboard layout in the 1860s (often assumed to be for slowing down fast typists), few would have imagined that his invention would become the dominant input device of the 20th century. In the early years of the 21st century (the so called 'speed and information' century), its use remains dominant, despite many, arguably better, input devices having been invented. Surely it is time to consider alternatives, in particular the most natural method of human communications – spoken language. Spoken language is not only natural, but in many cases is faster than typed, or mousedriven input, and is accessible at times and in locations where keyboard, mouse and monitor (KMM) may not be convenient to use. In particular, in a world with growing penetration of embedded computers, the so-called 'smart home' may well see the first massmarket deployment of vocal interaction (VI) systems. What is necessary in order to make VI a reality within the smart home? In fact much of the underlying technology already exists – many home appliances, electrical devices, infotainment systems, sensors and so on are sufficiently intelligent to be networked. Wireless home networks are fast, and very common. Speech synthesis technology can generate natural sounding speech. Microphone and loudspeaker technology is wellestablished. Modern computers are highly capable, relatively inexpensive, and – as embedded systems – have already penetrated almost all parts of a modern home. However the bottleneck in the realisation of smart home systems appears to have been the automatic speech recognition (ASR) and natural language understanding aspects. In this chapter, we establish the case for automatic speech recognition (ASR) as part of VI within the home. We then overview appropriate ASR technology to present an analysis of the environment and operational conditions within the home related to ASR, in particular the argument of restricting vocabulary size to improve recognition accuracy. Finally, the discussion concludes with details on modifications to the widely used Sphinx ASR system for smart home deployment on embedded computers. We will demonstrate that such deployments are sensible, possible, and in fact will be coming to homes soon.