The 2nd arg to get_tot_scores() here, representing log_semiring, should be false, because ARPA-type language models are constructed in such a way that the backoff prob is included in the direct arc. I.e. we would be double-counting if we were to sum the probabilities of the non-backoff and backoff arcs.
Change log_semiring to tropical_semiring indeed improves the WER. For the test-clean dataset, when num_paths is 100
and lm_scale=1.2, the WER decreases from 6.06 to 5.98.
See https://github.com/k2-fsa/snowfall/pull/201#discussion_r647975506
Change log_semiring to tropical_semiring indeed improves the WER. For the test-clean dataset, when num_paths is 100 and lm_scale=1.2, the WER decreases from 6.06 to 5.98.