Open greed2411 opened 2 years ago
Improve on the current confidence scoring algorithm here. The current confidence scores aren't very informative in terms of relative importance of the resulting alternatives. This is because Kaldi's AM and LM scores are scaled and unbounded negative log-likelihoods. Write an algorithm to compute the aggregate %confidence scores that can help improve performance on downstream tasks like SLU by removing bat alternatives, e.g. confidence scores similar to that of Google ASR.
Relevant links:
To be updated.