Cross-lingual detection of mild cognitive impairment based on temporal parameters of spontaneous speech

Abstract

Mild Cognitive Impairment (MCI) is a heterogeneous clinical syndrome, often considered as the prodromal stage of dementia. It is characterized by the subtle deterioration of cognitive functions, including memory, executive functions and language. Mainly due to the tenuous nature of these impairments, a high percentage of MCI cases remain undetected. There is evidence that language changes in MCI are present even before the manifestation of other distinctive cognitive symptoms, which offers a chance for early recognition. A cheap non-invasive way of early screening could be the use of automatic speech analysis. Earlier, our research team developed a set of speech temporal parameters, and demonstrated its applicability for MCI detection. For the automatic extraction of these attributes, a Hungarian-language ASR system was employed to match the native language of the MCI and healthy control (HC) subjects. In practical applications, however, it would be convenient to use exactly the same tool, regardless of the language spoken by the subjects. In this study we show that our temporal parameter set, consisting of articulation rate, speech tempo and various other attributes describing the hesitation of the subject, can indeed be reliably extracted regardless of the language of the ASR system used. For this purpose, we performed experiments both on English-speaking and on Hungarian-speaking MCI patients and healthy control subjects, using English and Hungarian ASR systems in both cases. Our experimental results indicate that the language on which the ASR system was trained only slightly affects the MCI classification performance, because we got quite similar scores (67-92%) as we did in the monolingual cases (67-92% as well). As our last investigation, we compared the proposed attribute values for the same utterances, utilizing both the English and the Hungarian ASR models. We found that the articulation rate and speech tempo values calculated based on the two ASR models were highly correlated, and so were the attributes corresponding to silent pauses; however, noticeable differences were found regarding the filled pauses (still, these attributes remained indicative for both languages). Our further analysis revealed that this is probably due to a difference regarding the annotation of the English and the Hungarian ASR training utterances.