In this paper, the fundamental methods of audio content retrieval are researched. Started from fast vocabulary independent keyword spotting, several audio content search algorithms based on acoustic GMM and TRAP-NN framework are explored. Word lattice-based methods are also discussed, and related indexing and retrieval schemes are provided. Based on these techniques, an integrated audio content retrieval system is implemented. The main works of this paper are as follows: 1 Based on acoustic GMM method, a novel phoneme matrix-based vocabulary independent keyword spotting approach is proposed. Very fast detect speed is achieved at the cost of acceptable precision loss. 2 TRAP feature and NN model are explored, and related phoneme recognition results are discussed. Keyword search method based on TRAP-NN model is provided. Compared to GMM approach, the new method require less training corpus with faster indexing speed and same detection rate. 3 Lattice-based keyword search methods are researched. Confusion network and revised word-to-syllable lattice algorithms are both proposed to solve OOV problem, with better precision than previous methods. A novel bi-syllable indexing and search scheme is proposed with less storage consumption cost and no performance degradation. 4 Using techniques above, an integrated audio content retrieval system is designed and implemented. Problems about mass audio processing, index storage and retrieval interfaces are carefully considered and solved.
修改评论