Early screening for autism using audio indicators

In this study, the differences between the pre-verbal speech characteristics of children with autism and other children and the role of these differences in the early diagnosis of the disorder have been studied. One of the hallmarks of this disorder, which has been reported in autistic children at an early age, is the presence of stretched and uniform sounds and screams unique to this group.

To fully and comprehensively examine the characteristics of children’s voices, a collection of 237 different features that are categorized into eight separate categories is extracted.

To implement an efficient classifier that can separate the voices of children with autism from those of other children, we need to use feature selection methods to remove features that are not accepted in the same two categories of values and do not create much separation. Due to the difficulty of collecting appropriate audio data from children under the age of four, reducing the size of the feature vector will help prevent bias.

Comparison of the three feature selection methods shows that perceptual features along with MFCC features are the best features for separating the collecting data. Also, three different methods have been used to classify the data, each with different strengths and weaknesses.

The design and evaluation of the parameters of these classifiers are based on data from four healthy children and five children with the disorder and using the selected features, and in the best design, it leads to the correct classification rate of about 97%.


  1. Hamid Ebrahimi
  2. Hadi moradi
  3. Hamidreza Pouretemad