Talkidz: Phoneme Recognition in Italian Children's Speech with Deep Learning Techniques

Speech therapists in Italy face significant challenges in diagnosing speech impairments due to the lack of efficient and accessible tools, specially for transcribing speech into the International Phonetic Alphabet (IPA), a task that generally takes ten times as long as the length of the speech audio itself. To address this issue, a team of therapists from Fondazione Paideia has developed a standardized test for childern aged 3 to 6, aimed at improving early diagnosis. This thesis presents the implementation of a web-based platform that provides the test, alongside the development of Neural Networks trained to automate speech-to-IPA transcription. The project required extensive reserach to balance accuracy, speed, and usability, ensuring the system effectively generalizes across different therapists' needs. By integrating state-of-the-art Deep Learning models with clinical expertise, the proposed solution enhances speech assessment reliability while reducing the workload for professionals. This thesis, and the whole research behind it, contributes to the advancement of speech therapy tools by offering a novel data-driven approach to standardized speech evaluation.