Closed csukuangfj closed 3 years ago
The results of using 1-best and n-best decoding are given below:
The column in nbest indicates the size of n in n-best.
nbest
n
n-best
Detailed WERs are shown below:
# NO LM Rescoring # 1-best # [test-clean-no_rescore] %WER 7.19% [3781 / 52576, 503 ins, 318 del, 2960 sub ] # [test-other-no_rescore] %WER 18.99% [9941 / 52343, 1129 ins, 930 del, 7882 sub ] # # [test-clean-no_rescore-50] %WER 7.20% [3785 / 52576, 606 ins, 279 del, 2900 sub ] # [test-other-no_rescore-50] %WER 19.23% [10064 / 52343, 1418 ins, 801 del, 7845 sub ] # [test-clean-no_rescore-100] %WER 7.18% [3774 / 52576, 604 ins, 279 del, 2891 sub ] # [test-other-no_rescore-100] %WER 19.23% [10063 / 52343, 1420 ins, 800 del, 7843 sub ] # [test-clean-no_rescore-200] %WER 7.18% [3777 / 52576, 604 ins, 279 del, 2894 sub ] # [test-other-no_rescore-200] %WER 19.24% [10069 / 52343, 1411 ins, 810 del, 7848 sub ] # [test-clean-no_rescore-300] %WER 7.19% [3778 / 52576, 604 ins, 279 del, 2895 sub ] # [test-other-no_rescore-300] %WER 19.20% [10052 / 52343, 1409 ins, 805 del, 7838 sub ] # [test-clean-no_rescore-500] %WER 7.19% [3778 / 52576, 604 ins, 279 del, 2895 sub ] # [test-other-no_rescore-500] %WER 19.25% [10074 / 52343, 1420 ins, 809 del, 7845 sub ]
n-best decoding without an LM does not seem to outperform 1-best decoding when no LM is used. Anyway, it shows how we can use k2 for n-best decoding.
Cool! I suppose it's OK to merge even though it doesn't help, because it might be useful as an example.
The results of using 1-best and n-best decoding are given below:
The column in
nbest
indicates the size ofn
inn-best
.Detailed WERs are shown below:
n-best
decoding without an LM does not seem to outperform 1-best decoding when no LM is used. Anyway, it shows how we can use k2 for n-best decoding.