In recent years, server-based automatic speech recognition (ASR) systems have become ubiquitous, and unprecedented amounts of speech data are now available for system training. The availability of such training data has greatly improved ASR accuracy, but how to maximize the ASR performance in new domains or domains where ASR systems currently fail (thus limiting data availability) is still an important open question. In this paper, we propose a framework for mapping large speech corpora to different acoustic environments, so that such data can be transformed to build high-quality acoustic models for other acoustic domains. In our experiments using a large corpus, our proposed method reduced errors by 18.6%.

抄録全体を表示

Large variations occur in actual meteorological conditions according to the time and place, on which the atmospheric absorption of sound depends strongly. To examine the temporal variability of atmospheric sound absorption during the year, the attenuation coefficients for atmospheric absorption were calculated from meteorological data in various world regions. The hourly meteorological data at 10 international airports and the half-daily aerological data obtained at 3 observatories were used in the calculation. The results show that significant differences in atmospheric absorption with the place can be found. The seasonal changes in the attenuation coefficients depend on both the frequency of the sound and the place. In addition, the effects of atmospheric absorption on aircraft noise propagation were examined by performing simple simulated calculations. The A-weighted sound pressure levels and sound spectra of aircraft noise change instantaneously owing to atmospheric absorption, which depends on actual meteorological conditions.

抄録全体を表示