Automatic evaluation of children reading aloud in babble noise of classroom context
Learning how to read is the most important step in the intellectual development of a child, who then acquires a skill that will impact his/her whole life. The learning process passes through oralisation, but it is time-consuming for teachers to regularly assess their student’s reading level, as it needs to be done individually with each of the 20-to-30 children in their class. We are developing, as a part of a pedagogical assistant for Kindergarten to 2nd grade teachers, a tool to automatically score the fluency and the pronunciation of students reading aloud, based on a speech recognition HMM-GMM system. Evaluating the performance of 5-7 years old children in the classroom environment is very challenging : lack of data, characteristic non-reader children’s prosody and ‘babble’ noise are the 3 main challenges we encounter. This poster presents the different strategies to address those challenges. The development of an efficient process to quickly collect and annotate a large amount of children’s voice recordings enables us to improve our acoustic models and networks. The slow-reading and underdeveloped enunciation of young children is addressed with personalized language and grammar models. The ‘babble’ noise, that is typical in classrooms, is hard to filter due to variable and often poor quality of audio gear of schools and the targeted child’s voice sometimes being drowned out by noise. Further work will consist of training a neural network that automatically rejects the recordings containing too much noise: if so, the child is asked to speak louder, or, in extreme cases, the reading aloud exercise is stopped.