Interpresing Noisy Prose using Heterogeneous Multi-cores

Currently, verbally dictating commands and programs to a computer is not practical because in our experiments, 10% to 15% of the words and 30% to 80% of the commands and statements were misrecognized by the Automatic Speech Recognition (ASR) systems. Interpreting Noisy Prose (INP) is a post ASR algorithm designed to execute on and utilize the extensive computing power of heterogeneous multi-core chips. The algorithm generates several variations of an ASR’s output based on common and user defined homonyms, misunderstandings and “near homonyms.” The variations are compared against domain specific grammars and lexicons. To achieve acceptable speeds and accuracy, INP uses parallel computation to follow the branches of a search tree and data parallel matching at nodes of the tree to evaluate the sets of alternatives and select the “best” interpretation of the ASR’s output. The initial INP results were 99% correct recognition for words and 94% correct recognition for commands. The heterogeneous approach reduces the number of parallel tasks by about a factor of 150.