Automatic Transcription of Polish Radio and Television Broadcast Audio

摘要

This paper describes a Large-Vocabulary Continuous Speech Recognition (LVCSR) system for the transcription of television and radio broadcast audio in Polish. This work is one of the first attempts of speech recognition of broadcast audio in Polish. The system uses a hybrid, con nectionist recognizer based on a recurrent neural network architecture. The training is based on an extensive set of manually transcribed and verified recordings of television and radio shows. This is further boosted by a large collection of textual data available from online sources, mostly up-to-date news articles. The paper describes and evaluates some of the key components of the architecture. The system is also compared to a conventional HMM-based architecture. An application of the described system in indexing and search of terms within audio and video transcripts is also described.