Indexing of posterior lattice for spontaneous Mandarin speech
Peng Yu · Journal of Tsinghua University(Science and Technology) · 2008
Speech indexing is a key speech retrieval task.A sub-word-based indexing and merging method was developed to index Mandarin speech.Instead of the indexing best path,this method uses the lattice as the index which is first represented as the posterior lattice.The characteristics of the posterior lattice are used to break the word-based index into a sub-word index.Using Chinese characters as the sub-word gave a 3.9% FOM(figure of merit) improvement using syllables gave a 4.4% FOM improvement and using toneless syllables gave a 7.4% FOM improvement.The information from different speech recognizers was fused by merging the nodes and arcs inter-lattices and intro-lattices to get the best FOM of 78.8% compared to 68.3% for the word-index baseline and 66.9% for the syllable-index baseline.