Conversation
…asr#2573) If a single phoneme is aligned to the whole utterance, it is counted both as `begin` and `end`, but is added to the total only once. This caused `assert count >= 0` in analyze_phone_length_stats.py to fail. Now only the `begin` is counted in that case.
…ments for OCR tasks (kaldi-asr#2579)
…aldi-asr#2596) OpenFst 1.6.7 does not build with 4.8.1, and 4.8.2 has an stl bug that is fatal for Kaldi.
|
Hi Dan, |
|
ok thanks.
…On Thu, Aug 9, 2018 at 6:47 PM, Gaofeng Cheng ***@***.***> wrote:
Hi Dan,
the results did not change.
This update is a fix that in the first commit I used the wrong xconfig.
This update corrects the xconfig.
Gaofeng
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#50 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/ADJVu7C6BkVSDNNXJBcc0ShJ-67J3us1ks5uPOY2gaJpZM4V1J2P>
.
|
…arbage if PCA failed. (kaldi-asr#2590)
| conv-relu-batchnorm-layer name=cnn3 $cnn_opts height-in=40 height-out=20 height-subsample-out=2 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=128 | ||
| conv-relu-batchnorm-layer name=cnn4 $cnn_opts height-in=20 height-out=20 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=128 | ||
| conv-relu-batchnorm-layer name=cnn5 $cnn_opts height-in=20 height-out=20 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=128 | ||
| conv-relu-batchnorm-layer name=cnn6 $cnn_opts height-in=20 height-out=20 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=128 |
There was a problem hiding this comment.
sorry, I didn't look at this before. Can you try a version where the height-out of cnn5 and cnn6 is 10, not 20, and their num-filters-out is 256? This will leave the compute time about the same (while increasing the parameters), and will allow those layers to see a wider range of frequency. So reducing the height (and increasing the num-filters) actually increases the modeling power.
There was a problem hiding this comment.
tdnn7q_sp cnn_tdnn1a_sp cnn_tdnn1a_more_filters_sp
WER on train_dev(tg) 12.08 12.13 11.97
WER on train_dev(fg) 11.15 11.16 11.12
WER on eval2000(tg) 14.1 14.1 13.9
WER on eval2000(fg) 12.8 12.6 12.5
WER on rt03(tg) 17.5 17.3 17.1
WER on rt03(fg) 15.3 14.9 14.9
Final train prob -0.055 -0.057 -0.056
Final valid prob -0.072 -0.075 -0.075
Final train prob (xent) -0.875 -0.877 -0.871
Final valid prob (xent) -0.9064 -0.9134 -0.9110
Num-parameters 18725244 14597020 15187100
There was a problem hiding this comment.
Great! So that's the setup with more filters and height-out=10 on the last 2 layers, then?
In that case I think you should just change your 1a to be that configuration, and we could mege that.
…ere is no spk info (kaldi-asr#2639)
…asr#2581) This came from Vimal's work on the MGB-3 challenge. Interface is similar to the existing GMM-based cleanup/segmentation scripts.
…_tdnn_f_cgf_local
20 participants
No description provided.