On what basis or criteria are the different word error rates(WER) calculated in decoding?

[Prabhjit Singh Thind]

So I run my own recipe of .wav audio files on Kaldi. I create the language model, do the monophone and triphone training and decoding according to the tutorial Kaldi for Dummies. Its link is: http://kaldi-asr.org/doc/kaldi_for_dummies.html I get the decoding result in the form of following WERs.

For Triphone:

exp/tri1/decode/wer_10

%WER 2.83 [ 114 / 4034, 60 ins, 31 del, 23 sub ]

%SER 8.28 [ 81 / 978 ]

exp/tri1/decode/wer_11

%WER 2.55 [ 103 / 4034, 51 ins, 31 del, 21 sub ]

%SER 7.46 [ 73 / 978 ]

exp/tri1/decode/wer_12

%WER 2.28 [ 92 / 4034, 42 ins, 30 del, 20 sub ]

%SER 6.54 [ 64 / 978 ]

exp/tri1/decode/wer_13

%WER 2.13 [ 86 / 4034, 35 ins, 33 del, 18 sub ]

%SER 6.13 [ 60 / 978 ]

exp/tri1/decode/wer_14

%WER 2.13 [ 86 / 4034, 35 ins, 33 del, 18 sub ]

%SER 6.13 [ 60 / 978 ]

exp/tri1/decode/wer_15

%WER 1.98 [ 80 / 4034, 31 ins, 32 del, 17 sub ]

%SER 5.73 [ 56 / 978 ]

exp/tri1/decode/wer_16

%WER 1.93 [ 78 / 4034, 30 ins, 31 del, 17 sub ]

%SER 5.52 [ 54 / 978 ]

exp/tri1/decode/wer_17

%WER 1.86 [ 75 / 4034, 29 ins, 30 del, 16 sub ]

%SER 5.21 [ 51 / 978 ]

exp/tri1/decode/wer_7

%WER 3.59 [ 145 / 4034, 98 ins, 23 del, 24 sub ]

%SER 11.25 [ 110 / 978 ]

exp/tri1/decode/wer_8

%WER 3.25 [ 131 / 4034, 80 ins, 27 del, 24 sub ]

%SER 9.71 [ 95 / 978 ]

exp/tri1/decode/wer_9

%WER 3.10 [ 125 / 4034, 72 ins, 29 del, 24 sub ]

%SER 9.10 [ 89 / 978 ]

Now my question is on what criteria or basis are the different WERs and SERs calculated. Which are the parameters that change in each of the cases above?

Please Help!!

Thanks in advance

 

[Daniel Povey]

It's the LM scale that changes.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Prabhjit Singh Thind]

Okay and one more thing I wanted to ask. Is there any way we can calculate WER separately for every speaker in the same recipe

[Jan Trmal]

if you use the scripts steps/score_kaldi.sh, it will compute the per-user stats for you (among other things). Or just read the script to figure out how to do it.

y.

--

[Nani]

steps/diagnostic/analyze_lats.sh: see stats in exp/tri1/decode/log/analyze_lattice_depth_stats.log

exp/tri1/decode/wer_10

%WER 88.51 [ 77 / 87, 3 ins, 26 del, 48 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_11

%WER 88.51 [ 77 / 87, 2 ins, 29 del, 46 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_12

%WER 88.51 [ 77 / 87, 2 ins, 33 del, 42 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_13

%WER 88.51 [ 77 / 87, 2 ins, 34 del, 41 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_14

%WER 87.36 [ 76 / 87, 1 ins, 35 del, 40 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_15

%WER 87.36 [ 76 / 87, 1 ins, 36 del, 39 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_16

%WER 87.36 [ 76 / 87, 1 ins, 37 del, 38 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_17

%WER 87.36 [ 76 / 87, 1 ins, 39 del, 36 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_7

%WER 100.00 [ 87 / 87, 13 ins, 16 del, 58 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_8

%WER 94.25 [ 82 / 87, 8 ins, 17 del, 57 sub ]

%SER 100.00 [ 16 / 16 ]

exp/tri1/decode/wer_9

%WER 93.10 [ 81 / 87, 7 ins, 22 del, 52 sub ]

%SER 100.00 [ 16 / 16 ]

These are my results after doing all....and my doubt is wer is high how can i improve my my model get low WER.

TIA

On Thursday, April 20, 2017 at 4:52:33 PM UTC+1, Yenda wrote:

if you use the scripts steps/score_kaldi.sh, it will compute the per-user stats for you (among other things). Or just read the script to figure out how to do it.

y.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.