Body-Conducted Speech Recognition and Application to Speech Support System

Abstract

In recent years, speech recognition systems have been used in a wide variety of environments, including internal automobile systems. Speech recognition plays a major role in a dialogue-type marine engine operation support system currently under investigation. In this system, speech recognition would come from the engine room, which contains the engine apparatus, electric generator, and other equipment. Control support would also be performed within the engine room, which means that operations with a 0-dB signal-to-noise ratio (SNR) or less are required. Noise has been determined to be a portion of speech in such low SNR environments, and speech recognition rates have been remarkably low. This has prevented the introduction of recognition systems, and up till now, almost no research has been performed on speech recognition systems that operate in low SNR environments. In this chapter, we investigate a recognition system that uses body-conducted speech, that is, types of speech that are conducted within a physical body, rather than speech signals themselves. Since noise is not introduced into body-conducted signals that are conducted in solids, even within sites such as engine rooms which are low SNR environments, it is necessary to construct a system with a high speech recognition rate. However, when constructing such systems, learning data consisting of sentences that must be read a number of times is required for creation of a dictionary specialized for body-conducted speech. In the present study we applied a method in which the specific nature of body-conducted speech is reflected within an existing speech recognition system with a small number of vocalizations. On the other hand, people with speech disabilities face communication problems in daily conversation. They can communicate with substitute speech, but this does not have the required frequency to be readily understood in daily conversation. Therefore, we have proposed the speech support system with body-conducted speech recognition. The system retrieves speech from the body-conducted speech via a transfer function using recognition to decide on a subword sequence and the duration. Before constructing the system, we examined the effectiveness of body-conducted speech recognition for communication disorders. The first step in constructing the system involved investigating continuous word unit speech recognition, using an acoustic model not suited to body-conducted speech for communication disorders. In this study, we analyzed each parameter of these speeches and experimented with body-conducted speech recognition. We concluded that an adaptation using body-conducted speech recognition to achieve high recognition performance for disorders is valid.