In recent years, speech recognition systems have been introduced in a wide variety of environments such as vehicle instrumentation. Speech recognition plays an important role in ships' chief engineer systems. In such a system, speech recognition supports engine room controls, and lower than 0-dB signal-to-noise ratio (SNR) operability is required. In such a low SNR environment, a noise signal can be misjudged as speech, dramatically decreasing the recognition rate. Hence, speech recognition systems operating in low SNR environments have not received much attention. Therefore, this study focuses on a recognition system that uses body-conducted signals. Such signals are seldom affected by background noise, and thus a high recognition rate can be expected in low SNR environments such as an engine room. Since noise is not introduced within body-conducted signals that are conducted in solids, even within sites such as engine rooms which are low SNR environments, construction of a system with a high recognition rate can be expected. However, within the construction of such systems, in order to create models specialized for body-conducted speech, learning data consisting of sentences that must be read in numerous times is required. Therefore, 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 only small numbers of vocalizations. Simultaneously, the measure by pretreatment was also worked on.
展开▼