Archives: Tweets

@eldang I think these qqs might be the curriculum for a ling-575 at UW that I teach (Tuning text synthesis for sociolinguistics?)

@eldang … and there are open qqs on whether training exx length, genre should change over training (or what effects are)

@eldang not necessarily. raw text includes role cues (“CLAUDIUS: Let ….”) so it might be sufficient. Length of training exx matters though

@eldang … so if your training data has sufficient information about the relationship between voices, then feeding it the cues may work.

@eldang I’ve not talked about it here, but you can provide a –seed flag to the samyro write tool that will give it a starting prompt

@eldang I think there may be some hyperparametric sweet spots where it will do the right thing, in fact.