Closed lucasgautheron closed 2 years ago
if you mean annotations where a judgment is made as to whether people are engaged in actual conversation (https://github.com/marisacasillas/chattr-basic#an-important-caveat), there isn't any high-level annotation for those datasets. I meant low-level rules that are applied to human annotations of who talks when; and also to automated annotations of who talks when.
One may think -- why would correlations vary as a function of ITI in such a case? I suspect because there are some non-obvious/overt parameters in the who talks when automated routine leading it to group or separate vocalizations by the same class, and/or to miss some vocalizations altogether. Optimizing a correlation coefficient (and/or an error rate) comparing human and automated annotation of 1- or 2-minute samples could help us answer: "which parameter should I use such that I can trust my automated analyses about as much as I'd trust a human who doesn't speak the language/understand what people are talking about?"
The conversation package aims at providing a way of evaluating conversations and extracting metrics from them (turn counts, duration etc.) ChildProject will use this package for everything related to conversations. https://github.com/LAAC-LSCP/conversations
Is your feature request related to a problem? Please describe.
Evaluating conversational turn counts from speech labels, requires a time threshold to discard speaker switches that are too far apart to belong to a conversation. It is expected to be O(1 s), but the impact of the choice of the parameter should be evaluated, and a better value retained if it leads to a noticeable improvement
Describe the solution you'd like
This will help solve #169