This project is exploring the vector implementation of speech algorithms on the torrent processor, which was developed under previous NSF support. Torrent is specialized for the low-precision digital signal processing that arises in speech recognition; it combines multiple fixed-point datapaths, a high bandwidth memory system, and a high-speed RISC general-purpose controller. Algorithms for the computationally intensive tasks of speech recognition, including acoustic probability estimation, feature extraction, lexical search, and evaluation of grammatical probabilities, are being vectorized and implemented on Torrent processor. The system will be demonstrated using a 2000-word natural speech recognition task that will run using a free-standing microphone in a noisy environment.
