Cross-Lingual Adaptation of Broadcast Transcription System to Polish Language Using Public Data Sources

摘要

We present methods and procedures designed for cost-efficient adaptation of an existing speech recognition system to Polish. The system (originally built for Czech language) is adapted using common texts and speech recordings accessible from Polish web-pages. The most critical part, an acoustic model (AM) for Polish, is built in several steps, which include: (a) an initial bootstrapping phase that utilizes existing Czech AM, (b) a lightly-supervised iterative scheme for automatic collection and annotation of Polish speech data, and finally (c) acquisition of a large amount of broadcast data in an unsupervised way. The developed system has been evaluated in the task of automatic content monitoring of major Polish TV and Radio stations. Its transcription accuracy (measured on a set of 4 complete TV news shows with total duration of 105 min) is 79,2%. For clean studio speech, its accuracy gets over 92%.