K. Murakami, K. Araki, M. Hiroshige, and K. Tochinai (Japan)
Speech translation, Inductive learning, Time-varying characteristics of speech
Conventional speech translation is an integration of three processing techniques: speech recognition, machine translation, and speech synthe sis. Every part consists of several complicated parts. However, any technique has not obtained the satisfying results yet. Even if the accuracy of each processing achieves 90%, the final result is only 53% from the total results of speech recognition, morphological analysis, syntactic analysis, syntactic transform, speech synthesis, and more. Progress of each process is an only way to achieve total high accuracy because lack of accuracy in each process causes declining of the final result. In this paper, we propose a new speech translation method that is not dependent on individual character of any language. The structure of proposed translation system is much simpler than others due to omitting several complicated processes to realize speech translation. We focused on the time-varying characteristics of acoustic speech waveform of source and target languages, and our method realize the speech translation by learning translation rules that have correspondence between two utterances inductively. The approach will be able to not only translate speech samples with in dependence of any language, but also will deal with all processes in short time because the system does not have a lot of processing stages. In this paper, we deal with a machine translation between Japanese and English.
Important Links:
Go Back
