Figure 2.

NUVA: An utterance verification system for assessing naming attempts in anomia treatment. Given a naming attempt, e.g. target word /tree/, the aphasic's utterance is recorded and processed through our DNN) to generate posteriorgrams (1). The system keeps posteriorgrams of previously recorded healthy speakers' utterances for each target word, (2a and 2b). Posteriorgrams are compared using the DTW) algorithm yielding a distance number between 0 and $+\infty$ (3a and 3b). The minimum of both distances is selected (4) and compared to a set threshold (5) calibrated per speaker, in this example 0.675. If the distance is less than the threshold, then the decision is that the aphasic speaker has said the target word correctly, otherwise it is classified as incorrect.