Sixth International Conference on Spoken Language Processing
If the acoustic models in an ASR system have been built using standard pronunciations in the acoustic training database, dialect speakers usually show in a test a lower ASR performance compared to speakers of standard pronunciations. In this paper, this degree of degradation is considered to be a measure for the distance between dialect and standard pronunciation. We relate this ASR-distance with a phonologically based distance between dialect and standard pronunciation. It is concluded that phonological and acoustically based distance measures are in line with each other, but this conclusion is tentative due to the degrees of uncertainty in all measurements. Simple lexical modifications based on phonological knowledge to accommodate with the dialectal pronunciations were only moderately successful.
Bibliographic reference. Bosch, Louis ten (2000): "ASR, dialects, and acoustic/phonological distances", In ICSLP-2000, vol.3, 1009-1012.