摘要 |
PURPOSE:To improve the recognition performance by adapting a recognition device to a new task by using learning data of the new task that a small number of speakers voice and learning data of a task that many speakers voice. CONSTITUTION:The voicing data of the task A of the speaker SP who voices both of two different tasks A and B and the voicing data of the speaker SQ who voices only the task A are inputted to a correspondence means 101, which makes them correspond to each other by DP matching. A mapping generating means 102 generates mapping from the speaker SP to the speaker SQ from a pair of data made to correspond each other A learning data generating means 103 estimates the data of the task B by the speaker SQ from the data of the task B by the speaker SP by using the generated mapping and outputs it. A standard pattern learning means 104 learns and outputs a standard pattern by using the data of the task B by the speaker SP and the data of the task B by the speaker SQ which is generated by the learning data generating means 103. |