| 摘要 |
PROBLEM TO BE SOLVED: To provide a method for delivering an acoustic model for voice recognition, which is reduced in the expected value of calculation cost required for providing a user with the acoustic model, with a high probability. SOLUTION: The method uses a system which is constituted of a user's personal computer 1 having a digital sound recorder 2 like an A/D converter of a sound board, a host computer 4 which uses a large quantity of sound data to learn an acoustic model for unspecific speakers, and a network which connects them. Filtering processing to correct frequency characteristics on the basis of recording frequency characteristic data 3 presented from the user with respect to the large quantity of sound data preserved in a sound database 6 connected on the host computer 4 and down sampling processing are performed, and a learning history database 5 after these processing is used to learn the acoustic model, and the model after learning is provided for the user. Thereby, the acoustic model for unspecific speakers which is adapted to recording frequency characteristics of the user's personal computer is provided. |
