Even if only the acoustic channel is considered, human communication is

highly multi-modal. Non-lexical cues provide a variety of information such as

emotion or agreement. The ability to process such cues is highly relevant for

spoken dialog systems, especially in assistance systems. In this paper we focus

on the recognition of non-lexical confirmations such as "mhm", as they enhance

the system's ability to accurately interpret human intent in natural

communication. The architecture uses a Support Vector Machine to detect

confirmations based on acoustic features. In a systematic comparison, several

feature sets were evaluated for their performance on a corpus of human-agent

interaction in a setting with naive users including elderly and cognitively

impaired people. Our results show that using stacked formants as features yield

an accuracy of 84% outperforming regular formants and MFCC or pitch based

features for online classification.