Analyzing Children's Speech: An Acoustic Study of Consonants and Consonant-Vowel Transition

Towards age-independent acoustic modeling

“…Increased variability in formant frequencies results in greater overlap among phonemic classes for children than for adult speakers, and makes the speech classification problem inherently more difficult. This study was later extended to consonants in (Gerosa et al, 2006b), obtaining similar results to the ones obtained for vowels. In (Gerosa et al, 2007) the effect of inter-speaker spectral variability was measured on the ChildIt corpus, the distance between probability distributions modeling vowel sounds for different age groups was measured.…”

“…However, even in the case of adequate amounts of age-specific training data, recognition performance reported for children is usually significantly lower than that reported for adults and it improves as the children's age increases (Wilpon and Jacobsen, 1996;Li and Russell, 2002;Potamianos and Narayanan, 2003;Hagen et al, 2003). This correlates well with studies showing that intra-and inter-speaker spectral variability decrease as age increases (Lee et al, 1999;Gerosa et al, 2006b). Furthermore, experiments of human perception of speech from children aged 6-11 show that the human word recognition error rate increases as the age of the child decreases .…”

“…Table 3 show that, as expected, under unmatched conditions (for example, in the case of children's speech recognized with acoustic models trained on adult speech), recognition results are much worse than those achieved under matched conditions. This is mainly due to different characteristics of adult and children's voices (Wilpon and Jacobsen, 1996;Potamianos and Narayanan, 2003;Lee et al, 1999;Gerosa et al, 2006b). …”

Atualidades na Investigação em Psicologia e Psicanálise

“…Para elas, é atrativo brincar com as vocalizações, imitar as vozes dos outros, os sons dos objetos, dos animais etc. Aos 5 anos, apresentam uma variabilidade espectral e temporal na fala muito maior que adultos ou crianças mais velhas, seja entre uma criança e outra da mesma idade, seja a mesma em diferentes momentos e para diferentes palavras e sons (Gerosa, Lee, Giuliani, & Narayanan, 2006). Essa variação nas características vocais é efeito tanto das mudanças anatômi-cas e fisiológicas no desenvolvimento orgânico quanto da crescente habilidade na produção e articulação de sons.…”

Tandem decoding of children's speech for keyword detection in a child-robot interaction scenario

“…Hence, the recognizer does not rely on preprocessing steps such as voice or feature transformation but uses acoustic models that inherently capture the properties of children's speech. As the characteristics of coarticulation in children's speech strongly differ from coarticulation effects in adult speech [Gerosa et al 2006], we apply BLSTM networks as an efficient and comparably novel method of context modeling. Children develop coarticulation skills with increasing age, leading to strong variations in the amount of temporal context that needs to be considered to capture coarticulation for context-sensitive speech feature generation and acoustic modeling [Repp 1986;Mayo et al 2003].…”