Acoustic modeling of spontaneous speech of Japanese preschool children

Abstract

In recent years, there is an increasing demand for speech recognition of children. However, the recognition of children’s speech, especially preschool children (2 to 5 years of age), is very difficult. For example, recognition accuracy using a children’s acoustic model provided by the Japanese Dictation Toolkit is only 21.4%. Many different variations of child speech with palatal sounds and pronunciation error decrease recognition performance. This paper proposes a recognition method that investigates the characteristics of preschool children’s speech using experimental data and considers phonetic changes. Mapping between standard and altered pronunciations of words is determined. In experiments, a large amount of spontaneous child speech (2 to 15 years of age) was collected with the speech‐oriented public guidance system, ‘‘Takemaru‐kun,’’ which is currently available. Recognition performance increases to 49.2% by acoustic model adaptation of preschool children’s speech. When allowing multiple pronunciation variations per word during recognition, further improvement to 52.0% is achieved.