This paper deals with acoustic properties of backchannels – those turns within a dialogue which do not convey information but signify that the speaker is listening to his/her interlocutor (uh-huh, hm etc.). The research is based on a Russian corpus of dialogue speech, SibLing, a part of which (339 min of speech) was manually segmented into backchannels and non-backchannels. Then, a number of acoustic parameters was calculated: duration, intensity, fundamental frequency, and pause duration. Our data have shown that in Russian speech backchannels are shorter and have lower loudness and pitch than non-backchannels. After that, two classifiers were tested: CART and SVM. The highest efficiency was achieved using SVM (F 1 = 0.651) and the following feature set: duration, maximum fundamental frequency, melodic slope. The most valuable feature was duration.
|Name||Lecture Notes in Computer Science|
|Conference||22nd International Conference on Speech and Computer|
|Abbreviated title||SPECOM 2020|
|Period||7/10/20 → 9/10/20|