Maas, Andrew

Apr 13, 2015 04:44 PM  (see Semantic Machines... Cal Berk profs in startup.. 15 employees;

Andrew joins them in Berk... has moved there)  stealth mode.

abst for today:

Speech recognition is a well-formalized, data-rich task where computational approaches still cannot match human performance. I will describe two projects on large vocabulary continuous speech recognition for conversational speech. Using an HMM-based approach to speech recognition, my team and I performed an extensive set of experiments using deep neural networks (DNNs) as acoustic models to map small spans of acoustic data to phonetic sub-states. Our results suggest that some aspects of DNN architecture are important for overall system performance, but overall it is difficult to obtain system performance gains in an HMM-based system by DNN modifications alone. This finding leads to the second project, which abandons the traditional HMM-based framework for speech recognition and instead utilizes deep recurrent networks to directly map from acoustic inputs to text characters. We demonstrate the first set of results using such a simple approach to yield reasonable performance on the challenging Switchboard corpus. Our model uses a character-level language model in place of a word-level lexicon and language model. This further reduces system complexity and enables the model to transcribe out of vocabulary words.