Computer Science Graduate Seminar: Progress in Decoding for Large Vocabulary Continuous Speech Recognition
Wednesday, April 19, 2017, 10:15am
Location: Computer Science Center, Room 5056, Ahornstr. 55
Speaker: Dipl.-Inform. David Nolden
The subject of this thesis is the search problem in automatic speech recognition. The search is responsible for matching an incoming acoustic speech signal with statistical speech models, in order to find the word sequence which is most likely to have been spoken. In principle, it is necessary to enumerate all possible word sequences, to compute a likelihood for each word sequence according to the models, and to select the best one. When the vocabulary is large, then such a straight-forward approach is not feasible, due to the huge number of possible word sequences; instead, state-of-the-art approaches transform the models into compact search network structures, match the input signal time-synchronously against the search network, and exploit recombination and pruning to limit the search effort. In this work, we analyze existing search strategies, combine them, and introduce novel extensions which further improve their efficiency and precision. We give a holistic overview of the ingredients required for efficient search. We investigate how the search network should be structured, and how the search space can be managed most efficiently. Normally, the search space depends on the language model; we introduce a novel search space management algorithm, which partially decouples the search effort from the language model's order. We introduce a novel framework which explains why pruning is possible, and which helps motivating and finding effective pruning methods; it establishes a direct relationship between pruning and recombination. Then we analyze common pruning methods regarding effectiveness and motivation, introduce novel pruning methods, and propose improved look-ahead techniques which make the pruning more effective. Pruning induces a certain amount of search errors, and usually a specific trade-off between precision and efficiency needs to be selected manually. In a last step, we show how search errors can be detected, and derive a search algorithm which allows efficient search without search errors. All methods are evaluated experimentally on a variety of state-of-the-art speech recognition tasks. On all tasks, a considerable reduction of the search space is achieved using the new methods, and overall, a speedup of the core search by a factor of more than 10 is achieved in comparison to the baseline method.
The computer science lecturers invite interested people to join.