search

uhh-lt / kaldi-tuda-de

Scripts for training general-purpose large vocabulary German acoustic models for ASR with Kaldi.
Apache License 2.0
172 stars 36 forks source link

Improve robustness and handling of mailabs_data #32

Closed patientzero closed 4 years ago

patientzero commented 4 years ago

adding some | true statements to make deleting files that are possibly not there safer with the -e pipefail option.

Adding relative path to scipts under ./utils as many people will not have them in their PATH.

Slight change to the location and handling of mailabs data, as the extraction and location failed for me on running the script.

bmilde commented 4 years ago

Thanks, appreciate the pull request! Have you been otherwise able to reproduce the WER results?

What happened here: https://github.com/uhh-lt/kaldi-tuda-de/pull/32/commits/6401c8173401fa085b9874462bbff2664adc86b5 :D ?

patientzero commented 4 years ago

hey @bmilde Sorry for the late reply, somehow this was lost...

What happened here: 6401c81 :D ?

ah dammit, that comment made it in... sorry you had to read that, it was not intentional. Took me some time to figure out what was wrong at that point. Sure you know the little 'pains' of kaldi, too. I am extremely grateful for your work, as it helps me a and surely a lot of others. I will report my WER results in the next couple of days.

patientzero commented 4 years ago

hi @bmilde ,

finally thought about looking up the WER results... I could pretty much replicate them. Here the output as a markdown table(github comments somehow doesnt support it, just paste in markdown editor): WER Details System
29.27 [ 20369 / 69600, 3533 ins, 2891 del, 13945 sub ] [PARTIAL] exp/tri3/decode_test_nosp/wer_13_0.0
28.18 [ 18818 / 66768, 3287 ins, 2712 del, 12819 sub ] [PARTIAL] exp/tri3/decode_dev_nosp/wer_13_0.0
27.33 [ 18248 / 66768, 3214 ins, 2589 del, 12445 sub ] [PARTIAL] exp/tri3/decode_dev_pron/wer_14_0.5
28.99 [ 20178 / 69600, 3589 ins, 2753 del, 13836 sub ] [PARTIAL] exp/tri3/decode_test_pron/wer_13_0.5
58.81 [ 40935 / 69600, 2203 ins, 16965 del, 21767 sub ] [PARTIAL] exp/tri2/decode_test_nosp/wer_9_0.0
56.66 [ 37833 / 66768, 1786 ins, 17220 del, 18827 sub ] [PARTIAL] exp/tri2/decode_dev_nosp/wer_10_0.0
15.62 [ 10873 / 69600, 2769 ins, 830 del, 7274 sub ] [PARTIAL] exp/chain_cleaned/tdnn1f_2048_sp_bi/decode_test/wer_11_1.0
14.60 [ 9745 / 66768, 2565 ins, 799 del, 6381 sub ] [PARTIAL] exp/chain_cleaned/tdnn1f_2048_sp_bi/decode_dev/wer_11_1.0
30.96 [ 20673 / 66768, 3555 ins, 2888 del, 14230 sub ] [PARTIAL] exp/tri4_cleaned/decode_dev.si/wer_13_0.5
31.69 [ 22059 / 69600, 3998 ins, 2807 del, 15254 sub ] [PARTIAL] exp/tri4_cleaned/decode_test.si/wer_12_0.5
25.06 [ 17444 / 69600, 3542 ins, 2080 del, 11822 sub ] [PARTIAL] exp/tri4_cleaned/decode_test/wer_13_0.5
23.36 [ 15597 / 66768, 3123 ins, 1971 del, 10503 sub ] [PARTIAL] exp/tri4_cleaned/decode_dev/wer_14_0.5
59.85 [ 41653 / 69600, 2125 ins, 16347 del, 23181 sub ] [PARTIAL] exp/tri1/decode_test_nosp/wer_10_0.0
57.06 [ 38095 / 66768, 2096 ins, 14546 del, 21453 sub ] [PARTIAL] exp/tri1/decode_dev_nosp/wer_10_0.0
34.95 [ 23338 / 66768, 4342 ins, 2881 del, 16115 sub ] [PARTIAL] exp/tri4/decode_dev_pron.si/wer_13_0.5
35.30 [ 24566 / 69600, 4844 ins, 2687 del, 17035 sub ] [PARTIAL] exp/tri4/decode_test_pron.si/wer_12_0.5
24.11 [ 16096 / 66768, 3462 ins, 1858 del, 10776 sub ] [PARTIAL] exp/tri4/decode_dev_pron/wer_14_0.5
25.77 [ 17939 / 69600, 3883 ins, 1953 del, 12103 sub ] [PARTIAL] exp/tri4/decode_test_pron/wer_13_0.5