Parameter sharing plays an important role in statistical modeling since training data are usually limited. On the one hand, we would like to use models that are as detailed as possible. On the other hand, with models too detailed, we can no longer reliably estimate the parameters. Triphone generalization may force two models to be merged together when only parts of the model output distributions are similar, while the rest of the output distributions are different. This problem can be avoided if clustering is carried out at the distribution level. In this paper, a shared-distribution model is proposed to replace generalized triphone models for speaker-independent continuous speech recognition. Here, output distributions in the hidden Markov model are shared with each other if they exhibit acoustic similarity. In addition to detailed representation, it also gives us the freedom to use a large number of states for each phonetic model. Although an increase in the number of states will inc...

We present a post-processing technique for correcting errors committed by an arbitrary continuous speechrecognizer. The technique leverages our observation that consistent recognition errors arising from mismatched training and usageconditions can be modeled and corrected. We have implemented a post-processor called SPEECHPP to correct word-level errors, and we show that this post-processing technique applies successfully when the training and usage domains differ even slightly; for the purposes of the recognizer, such a difference manifests itself as differences in the vocabulary and in the likelihoods of word collocations. We hypothesize that other differences between the training and usage conditions yield recognition errors with some consistency also. Hence, we propose that our technique be used to compensate for those mismatches as well. 1.

This document contains draft information from a section of a preliminary version of a VERBMOBIL deliverable (TP 5.3-P1). It is distributed in this form to assist partners in advance planning. This document contains a simple version of the core DATR inference engine in Prolog in order to illustrate the principles of DATR inference to Prolog programmers. Note that in minor details it departs slightly from DATR conventions: - nonstandard nodenames are permitted; - the knowledge base must be pre-sorted to permit 'longest path first' inference; - queries include the theory name. Note also that this is not a directly usable implementation: there is no user interface, no DATR-Prolog interpreter, no DATR-specific trace or debugging, no attention paide to efficiency, etc. The aim is to provide a minimal 'core DATR standard inference' interpreter in logical style. 1 Illustration of a DATR theory: a 'microlexicon' MINILEX.DTR Tablecloth: !? == Compound !ilex? == lemma !relation? == (for covering) !modifier? == "Table:!?" !head? == "Cloth:!?". Table: !? == Simplex !ilex? == lemma !meaning? == (horizontal surface to put things on) !orthography? == (t a b l e). Cloth: !? == Simplex !ilex? == lemma 32 Dafydd Gibbon !meaning? == (variety of textile) !orthography? == (c l o t h). Compound: !? == Word !ilex? == generalisation !type? == compound !meaning? == ("!head meaning?" "!relation?" "!modifier meaning?") !orthography? == ("!modifier orthography?" "!head orthography?"). Simplex: !? == Word !ilex? == generalisation !type? == simplex. Word: !ilex? == generalisation !type? == word. Theorems: Tablecloth:!relation?=(for covering). Tablecloth:!meaning?=(variety of textile for covering horizontal surface to put things on). Tablecloth:!orthography?=(t a b l e c l o t h). Table:!orthography...