Key points are not available for this paper at this time.
This paper presents a phonetically motivated analysis of a version of the continuous speech recognition system developed aspart ofthe ARM' (Airborne Reconnaissance Mission) project at the RSRE Speech Research Unit.This version of the system is speaker dependent and is based on phoneme-level hidden Markov models.In order to take aeemmtof thecontext sensitivity ofthe acouch rulisation ol'phonemcs.approximately 1500 word-intemaltriphone models are used.A triphone is a model of a phoneme in its left and right context.The version of the system on which this investigation was based scores an average of 86.8% word accuracy with word level syntax (perplexity 497).The 'ARM system is more fully described in Russell.Ponting those which cannot may indicate where models could be improved.For instance if lpl is misrecog-nised as /b/.this is understandable from a phonetic point of view.as the two areacousticallyvery similar.howevu iflp/ iseonsisten y misrecognised ashlar/@Ul it would bedifl'tcult to explain on acoustic phonetics grounds.and would probably indicate that there is something wrong with the model(s).The following section describes the background to the investigation.and how the phoneme perfonnnnce is evaluated.The reminder of the paper contains an analysis of speci c types of errors.motivated by the desire to nd phonetic explanations of them.The transcriptions in this paper are in the SAM-PA notation (Fourcin etal 2. BACKGROUND2.1 The system For the ptuposes of the analysis described in this paper the system was con gured as a phoneme recogniser with no
Browning et al. (Thu,) studied this question.