Noise immunity learning for word‐spotting speech recognition
Yoichi Takebayashi, Hiroshi Kanazawa · Systems and Computers in Japan · 1992
Abstract This paper discusses noise immunity learning in speech recognition by word‐spotting. In the proposed method, the learning speech data with superposed noise is constructed by adding the noise data to the clean speech data collected beforehand. Then the recognition dictionary for word‐spotting is trained to improve noise immunity. In the learning, the recognition by word‐spotting is attempted for the artificially synthesized learning data while gradually increasing the ratio of the contained noise. The recognition is executed using the learning word feature vector automatically extracted based on the similarity. The noise immunity is realized by the simulation in the noisy environment and the automatic learning. The recognition and learning use the word as the unit and the multiple similarity method, which can cope with a wide range of pattern deformation. An evaluation experiment is executed using 13‐word speech data (plus noise in a railway station). For the case of 96‐dimensional word feature vector (eight dimensions for frequency and 12 dimensions for time), and S/N ratio of 10 dB, the recognition rate is improved from 85.5 percent in the word‐spotting method without learning to 94.1 percent with learning. This indicates the usefulness of the proposed method.