The repeated use of out-of-vocabulary (OOV) words in a spoken document seriously degrades a speech recognizer performance. Even though such recurrent OOV words are often important keywords in a spoken document, they are never correctly recognized. We propose a novel method for robustly detecting recurrent OOV words, which focuses on the degree of consistency among them. It first detects recurrent segments, that is recurrent phoneme sub-sequence in the output of a phoneme sequence decoder. Then, we measure the degree of consistency by using the mean and variance (distribution) of features (DOF) derived from the recurrent segments, and use our DOF for IV/OOV classification. Experiments on academic lectures illustrate that the proposed DOF-based method can robustly detect recurrent OOV words in spontaneous speech and achieves over 60% relative reduction in false alarms. It is also confirmed that detection performance improves as the OOV words are repeated more often.
各種検索
サポート
ヘルプ