| Popis: |
We investigated a body-conducted speech recognition system for the establishment of a usable dialogue-type marine engine operation support system that is robust in noisy conditions, even in a low SNR environment such as an engine room. By bringing bodyconducted speech close to audio quality, we were able to examine ways to raise the speech recognition rate. However, in an examination by pretreatment, we could not obtain optimal results when using an adaptation filter and a cross-spectral method. We introduced an adaptive processing method and confirmed the effectiveness of adaptive processing via small repetitions of utterances. In an environment of 98 dB SPL, improvements of 50% or above of recognition rates were successfully achieved within one utterance of the learning data and speech recognition rates of 95% or higher were attained. From these results, it was confirmed that this method will be effective for establishment of the present system. In a wireless version of the system, the results showed a worsening of recognition rates because of noise in the speech bandwidth. Even when adaptive processing was performed, a sufficient speech recognition rate could not be obtained. Although more testing of this wireless system within the actual environment of the Oshima-maru will be necessary, it will also be necessary to investigate other wireless methods. |
