Voice analysis rehabilitation platform based on LSTM algorithm

The proposed work discusses the results of a research project based on the recognition of correctly pronounced words and phrases by implementing a web platform implementing an acoustic training model. The acoustic training model is performed by a long short-term memory - LSTM - algorithm, able to recognise the speech disorder by assigning a score for each test type. The paper discusses the platform design and implementation. The tests are performed for different kind of exercises in rehabilitation patterns. The adopted approach is based on the formulation of acoustic model integrating a training dictionary of correct phonemes to pronounce. The platform enables a real time automatic score of the performed exercises and the test planning. The LSTM training dataset can be enriched by adding new exercise to learn. The output graphical dashboards enforce clinical evaluations and reporting.