Web-based topic language modeling for audio indexing

We describe the implementation of a scalable architecture for audio indexing, in which topic-dependent language models (LMs) were trained on Web pages categorized in a portal Web directory and stored on distributed servers. Input speech was decoded in parallel on servers that each had an individual topic LM. From the decoders´ outputs, an optimal hypothesis was chosen for each utterance by a topic-selection criterion minimizing an energy function with three terms: likelihood scores for the utterances; keyword co-occurrence statistics to measure the long-distance correlation; and Web-based hypothesis verification scores, which penalize misrecognized trigrams through web search results. Experimental results showed that the proposed approach outperformed the baseline topic-independent system by 6.0% absolutely (20.0% relatively) in character accuracy.