Automatic Speech Recognition for Speech Assessment of Persian Preschool Children

Preschool evaluation is crucial because it gives teachers and parents crucial knowledge about a children's growth and development. The coronavirus pandemic has highlighted the necessity for preschool children to be assessed online. This online testing requires a variety of technologies, from web application development to various artificial intelligence models in diverse criteria such as speech recognition. Because of the acoustic fluctuations and differences in voice frequencies between children and adults, employing Automatic Speech Recognition(ASR) systems is difficult because they are pre-trained on adults' voices. In addition, training a new model requires a large amount of data. To solve this issue, we constructed an ASR for our cognitive test system using the Wav2Vec 2.0 model with a new pre-training objective, called Random Frequency Pitch(RFP), and our new dataset, which was tested on Meaningless Words(MW) and Rapid Automatic Naming(RAN) tests. Due to the peculiarities of these two tests, we explored numerous models, including Convolutional Neural Network(CNN) and Wav2Vec 2.0 models. Our new approach, reaches Word Error Rate(WER) of 6.45 on the Persian section of CommonVoice dataset. Furthermore our novel methodology produces positive outcomes in zero- and few-shot scenarios.