|
The Carnegie Speech™ Company is proud to announce a new database of children’s speech, YOUTH, available for use by both research and industry sites.The database was recorded in two schools in the suburbs of Pittsburgh during the fall of 2001. It contains the speech of 135 children reading isolated words, phrases and sentences from Nursery Rhymes and other classic tales in American English such as Peter Pan. The utterances were automatically chosen for their richness in triphone coverage. There are an average of 184 utterances per child, making a total of 25,122 utterances for the whole database. The children are 8 – 10 year old (in third and fourth grade) and assessed as “good readers” by their teachers. Some, but not all of the utterances are the same for all speakers. The children come from upper middle class suburban families, many of whom are not native to the region, and therefore do not manifest a Pittsburgh dialect. We did record a few non-native children, a few who have colds, and a few who have speech impediments. There is a spreadsheet included that indicates these details for each speaker. Recording Conditions:The recordings took place in quiet rooms off the libraries. The text was read from the screen. The children wore Andrea NC-72 headsets attached to preamps. Children were taken out of class for one 45-minute session each. Labelling of the database:The content of the utterances was determined by taking large quantities of text from Nursery Rhymes and other classic tales in American English such as Peter Pan. The utterances to be read were then automatically chosen for their richness in triphone coverage. The 179 first utterances are the same for all speakers and the 25 others are the same for every tenth child in the database (child 1 read from Section0.txt, child 2 read from Section1.txt, child 10 read from Section9.txt, child 11 read from Section0.txt and so on). It should be noted that most children did not finish all of the utterances they were to read – they were taken out of class for a specific amount of time and some read faster than others. Others were sick or tired and we never insisted that a child read more than he or she was comfortable with. Each speech file is accompanied by a text file containing the word labels of the speech. We do not label after the fact, but use on-the-fly choice to eliminate (and rerecord) utterances that do not correspond to the exact expected phonetic content of the text. We performed after the fact verification, eliminating the few utterances that did not meet this criterion (about 35 in the whole database). No phonetic labeling was performed since we feel that each user site will want to perform this in its own way. About our speakersThe 135 children were recorded during the fall of 2001 for a project funded by an SBIR from the Department of Education. We wanted the data to reflect a cross-section of the population of the schools we recorded in and we wanted it to be useful for real applications, so we included kids who had colds, kids who are not native speakers, kids who have speech impediments. We noted who they were in the “comments” column on the spreadsheet. There was mouth noise, microphone wire noise and heavy breathing – we do not feel that any application that is for children can keep them from making these noises and so we have kept these noises. The recordings were carried out by persons who had human experimentation clearance (the whole project is approved by an IRB). The children were each paid for their participation and have signed forms giving the Carnegie Speech™ Company the right to use their voices in its products. Recording Conditions:Here are four examples of utterances from the database. They come from different kids and are typical of our content. Girl saying: “He cries’ said Smee” Boy saying: “Oh what a lovely one” Boy saying: “Wednesday’s child is full of woe” Girl saying: “Catch that pig” If you are interested in getting more information or in licensing the database, please contact: Angela Kennedy Email: sales@carnegiespeech.com Tel 1-412-281-3004 |
Awarded an Advanced Technology Program Grant from NIST |
