Presented by:
Soo-young Suk and Hiroaki Kojima

---

[attachment=5717]

---

Introduction

People with severe speech and motor impairment due to cerebral palsy are great difficult to move independently and also cannot control home electric devices. Computer has much to offer people with disability, but the standard human-machine interface (e.g. keyboard and mouse) is inaccessible to this population. In this chapter, we describe a speech recognition interface for the control of powered wheelchair and home automation systems via severely disabled person’s voices. In particular, we consider that our system can be operated by inarticulate speech produced by persons with severe cerebral palsy or quadriplegia in realenvironment.
The aim of our research is divided two targets. One is easy to control of various home appliances by voice, and the other is to enable severely disabled person’s movement independently using voice activated powered wheelchair. At first, Home automation system product for intelligent home is increasingly getting very common by the help of intelligent home technologies that increased easy, safety and comfort. Moreover, home automation is an absolute benefit and can improve the quality of life for the user. Home automation houses have been developed to apply new technologies in real environment, such as Welfare Techno Houses (Tamura et al., 2007), Intelligent Sweet Home (Park et al., 2007), Smart House (West et al., 2005). Interfaces based on gestures or voices have been widely used for home automation. However, gesture recognition based on vision technologies depends critically on the external illumination conditions. And gesture recognition is difficult or impossible for people suffering from severe motor impairments, such as paraplegia and tremors. Recently, a voice-activated system using commercial voicerecognition hardware in a low-noise environment has been developed for disabled persons capable of clear speech (Ding & Cooper, 2005).
The next, powered wheelchairs provide unique mobility for the disabled and elderly with motor impairments. Sometimes, the joystick is a useless manipulation tool because the severely disabled cannot operate it smoothly. Using natural voice commands, like ``move forward'' or ``move left'' relieves the user from precise motion control of the wheelchair. Voice activated powered wheelchair is required safety manipulation with high speech recognition accuracy because the accident can occur by a misrecognition. Although current speech recognition technology has reported high performance, it is not sufficient for safe voice-controlled powered wheelchair movement by inarticulate speech affected by severe 528 Speech Recognition, Technologies and Applications cerebral palsy or quadriplegia, for instance. To cope with the pronunciation variation of inarticulate speech, we adopted a lexicon building approach based on Hidden Markov Model and data mining (Sadohara at al., 2005), in addition to acoustic-modeling-based speaker adaptation (Suk at al., 2005). We also developed noise-canceling methods, which reduce mechanical noise and environmental sounds for practical use on the street (Sasou at al., 2004). However, though our voice command system has improved recognition performance by various methods, the system requires a guarantee of safety for wheelchair users in two additional conditions. - To move only in response to the disabled person’s own voice. - To reject non-voice command input.
The first problem is to prevent operation of the wheelchair by unauthorized persons near the disabled user. A speaker verification method can be applied to solve this problem, but it is difficult to verify when using short word commands. Therefore, we are now developing a speaker position detection system using a microphone array (Jonson at al., 1993; Sasou & Kojima, 2006). The second problem is that a lot of other noise is input when the voice command system is being used. Also, a voice-activated control system must therefore reject noise and non-voice commands such as coughing and breathing, and spark-like mechanical noise in the preprocessing stage. A general rejection method has achieved a confidence measure using a likelihood ratio in a post-processing step. However, this confidence measure is hard to use as a non-command rejection method because of the inaccuracy of likelihood when speech recognition deals with unclear voice and non-voice sounds. Thus, a non-voice rejection algorithm that classifies Voice/Non-Voice (V/NV) in a Voice Activity Detection (VAD) step is useful for realizing a highly reliable voice-activated powered wheelchair system.
The chapter first presents the F0 estimator and the non-voice rejection algorithm. Next, the inarticulate speech recognition is described in Section 3. In Section 4, we present a developed voice activated control system. And we evaluate the performance of our system in Section 5. Lastly, we offer our conclusions in Section 6.