摘要 |
PROBLEM TO BE SOLVED: To provide a voice recognition performance estimation device capable of estimating an accurate recognition accuracy upper limit.SOLUTION: A multiple closed acoustic model generation part extracts an acoustic featured value from voice data for evaluation, and generates N(N≥2) pieces of closed acoustic models as a model between the acoustic featured value and an initial acoustic model. A learning set voice recognition accuracy calculation part performs the voice recognition of a learning set by using the N pieces of closed acoustic models, and outputs N pieces of voice recognition accuracy. An acoustic model selection part selects the closed acoustic model from which the maximum voice recognition accuracy is obtained, and outputs it as a selected closed acoustic model. A voice recognition accuracy calculation part outputs recognition accuracy obtained by performing the voice recognition of an evaluation set by using the selected closed acoustic model as the estimation value of a recognition accuracy upper limit. |