| 摘要 |
PURPOSE:To obtain a more precise recognition result at the time of initial speech input and speech re-input, to shorten the processing time regarding speech recognition, and to reduce appearance by not regarding an unnecessary section as an object of recognition. CONSTITUTION:At the time of the 1st speech input, a 1st likelihood calculation part 11a refers to feature parameters by analytic frames from the rear to the front of a speech signal. At the time of 2nd speech input after a process error, on the other hand, a 2nd likelihood calculation part 11b performs a process similar to the 1st speech input from the front to the rear of the speech signal at the time of the re-input in consideration of the tendency that a key word is present nearby the start end of a vocalized sentence. |
