Multi-level annotation in SpeeCon Polish Speech Database

SpeeCon Polish Speech Database was collected within the framework of the SpeeCon project partially sponsored by the EC (IST-1999-10003). The database contains two sets of data, which comprise 550 adults' recording sessions and 50 sessions from children, respectively. The adult speakers were recorded in various environments: offices, living rooms, cars and public places. Recordings contain free spontaneous speech passages, elicited spontaneous speech, phonetically compact words and sentences, general-purpose words and phrases, specific application words and utterances. One of the most important problems in the construction of the database is to define bases for multi-level transcription composed of several tiers. They could be grouped into three classes - linguistic, symbolic and physical representation. The orthographic transcription is applied to the sentence, phrase and word tiers, symbolic transcription related to grammar and articulation - to part of speech, phoneme and syllabic tiers and mnemonics - to the description of some characteristic of the measurable physical data. The paper presents the rules applied to text, speech and noise transcriptions and remarks on pronunciation varieties found in the database. The final part of the paper discusses the problem of the lexicon creation, which is an alphabetically ordered list of distinct lexical items occurring in the recorded corpus. The Polish lexicon has been built up by various methods, including hand-annotation and generation by rule with subsequent manual check.