NOISE LEVEL ESTIMATING METHOD, SPEECH SECTION DETECTING METHOD, SPEECH RECOGNIZING METHOD, SPEECH SECTION DETECTING DEVICE, AND SPEECH RECOGNITION DEVICE

摘要

PROBLEM TO BE SOLVED: To reduce and overcome the influence of unsteady node, variation in noise level, added noise, an answer speech, etc. ;SOLUTION: A speech analysis part 4 analyzes an inputted signal and extracts its power level and feature quantity vector. On the basis of the extracted power level, a frequency information update part 9 generates a power distribution chart in a frequency information storage part 13. A noise level decision part 10 detects the maximum frequency from the power distribution chart and selects a grade which is in the specific upper and lower section of the maximum frequency and has larger than a specific ratio to the maximum frequency, and a threshold value setting part 11 sets a threshold value with the mean value of power regarding the selected grade. A start and end detection part 12 cuts a speech section out by using the set threshold value and a recognition part 7 performs speech recognition as to the cut speech section by using the feature quantity vector extracted by the speech analysis part 4. An answer speech from an answer speech output part 2 is canceled by an echo cancellation part 3 from the input signal and a leak quantity decision part 5 decides the leak quantity of the answer speech and limits the threshold value.;COPYRIGHT: (C)1998,JPO