Connected Letter Recognition with a Multi-State Time Delay Neural Network

We present an MS-TDNN for recognizing continuously spelled letters, a task characterized by a small but highly confusable vocabulary. We pro(cid:173) pose training techniques aimed at improving sentence level perfor(cid:173) mance, including free alignment across word boundaries, word du(cid:173) ration modeling and error backpropagation on the sentence rather than the word level. Architectures integrating submodules special(cid:173) ized on a subset of speakers achieved further improvements.