Research
The EWA project focused on three key areas: data acquisition, speech parameter selection through data processing, and parameter evaluation to differentiate healthy speech from speech patterns indicative of cognitive impairment.
Data aquisiton
In collaboration with neurologists and psychologists, we developed a methodology for designing images appropriate for the study, determining key recording characteristics, categorizing data from healthy and sick individuals, and continuously optimizing this process. To facilitate data collection, we subsequently developed an intuitive application for recording participants' speech descriptions of the images.
Data processing
Spontaneous speech goes beyond just words. Laughter, hesitations, coughs, pauses – these all play a part. Annotation means the manual transcription of a vocal utterance into a text, with these accompanying sounds indicated. Annotation is one of the stages in the preparation of an artificial intelligence tool for speech to text - ASR (Automatic Speech Recognition). Subsequently, the speech utterances were subjected to acoustic analysis, which extracted hundreds of parameters such as length, volume, different types of frequencies, voice color, and others from the audio recordings. From the linguistic analysis, the so-called NLP (Natural Language Processing) parameters such as the number of full-meaning words, different word types, various lexical and semantic parameters were determined.
Evaluation of parameters
From one recording that contained descriptions of 65 images, i.e. 65 different speech utterances, over 100,000 parameters were generated and used in various computer algorithms and artificial intelligence methods such as linear regression, Random Forest and neural networks. By leveraging these algorithms and the knowledge of whether the speaker was healthy or diagnosed, the research achieved a high level of computer accuracy in differentiating between healthy and impaired speech. Additionally, the study revealed interesting statistical characteristics.