CROSS-CORRELATION BASED AUTOMATIC SEGMENTATION OF MEDIAL PHONEMES

This paper describes a cross-correlation based method aimed at extracting homogeneously-trimmed target medial phonemes from a reference pronunciation (a word or logatome pronounced by the Speech Language Pathologist (SLP)) and a sample utterance (the same word/logatome pronounced by a subject). The newly-generated audio segments are then fed to an Information Entropy based classification stage in order to assess the (dis)similarities between them and to serve as a valid, automated Speech Sound Disorder (SSD) Screening tool. The input for the study consisted in audio recordings of a Romanian word containing the target phoneme /r/ in medial position pronounced by a population of 44 preschoolers and primary schoolers, aged 6-7.