Cost-Efficient Development of Acoustic Models for Speech Recognition of Related Languages

Abstract

When adapting an existing speech recognition system to a new language, major development costs are associated with the creation of an appropriate acoustic model (AM). For its training, a certain amount of recorded and annotated speech is required. In this paper, we show that not only the annotation process, but also the process of speech acquisition can be automated to minimize the need of human and expert work. We demonstrate the proposed methodology on Croatian language, for which the target AM has been built via cross-lingual adaptation of a Czech AM in 2 ways: a) using commercially available GlobalPhone database, and b) by automatic speech data mining from HRT radio archive. The latter approach is cost-free, yet it yields comparable or better results in LVCSR experiments conducted on 3 Croatian test sets.