| 摘要 |
PURPOSE:To execute the correct recognition even against an input pattern to which a fluctuation is applied without increasing the number of a leaning voice sample by setting the upper limit of magnitude of the applied noise, based on a standard deviation calculated from a leaning pattern by an actual voice sample as a reference, at the time of increasing artificially the number of patterns by a word recognizing system. CONSTITUTION:The system is constituted or a voice input part 11 to a learning pattern storage part 16, at the time of generating a leaning pattern, constituted of the learning pattern storage part 16 and a neural network part 17 at the time of learning, and constituted of a voice input part 11, a pre-processing part 12, the neural network part 17, and a deciding part 18. In such a state, a noise superposing part 15 superposes a noise generated by a noise generating part 13 onto the leaning pattern and increases artificially the number of patterns. The applied noise is limited in a noise control part 14, and its upper limit is set, based on a standard deviation calculated at every speaker from a pattern extracted from a voice sample as a reference. |
