Constructing multi-level speech database for spontaneous speech processing
暂无分享,去创建一个
The paper describes a database called multi level speech database for spontaneous speech processing. We designed the database to cover textual and acoustic variations from declarative speech to spontaneous speech. The database is composed of 5 categories which are, in the order of decreasing spontaneity, spontaneous speech, interview, simulated interview, declarative speech with context, and declarative speech without context. We collected in total, 112 sets from 23 subjects (male: 19, female: 4). The database was firstly transcribed using 15 transcription symbols according to our own transcription rules. Secondly, prosodic information will be added. The goal of this research is a comparative textual and prosodic analysis at each level, quantification of spontaneity of diversified speech database for dialogue speech synthesis and recognition. From the preliminary analysis of transcribed texts, the spontaneous speech has more corrections, repetitions, and pauses than the others as expected. In addition, the average number of sentences per turn of spontaneous speech is greater than the others. From the above results, we can quantify the spontaneity of the speech database.
[1] Hajime Tsukada,et al. Prosodic Features of Utterances in Task-Oriented Dialogues , 1997, Computing Prosody.
[2] Björn Granström,et al. On the Analysis of Prosody in Interaction , 1997, Computing Prosody.
[3] Björn Granström,et al. Speech synthesis in spoken dialogue research , 1995, EUROSPEECH.
[4] Sheri Hunnicutt,et al. The waxholm application database , 1995, EUROSPEECH.