"formosa_speech" recipe and database for Taiwanese Mandarin speech recognition #2474
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Update upstream
|
Ping on this PR. |
|
@underdogliu, since we have not heard from @yfliao in a while, are you interested in applying your suggested fixes yourself and making your own version of this PR? |
|
Oh yeah, it looks like they did not provide the data for us to put on
openslr. So there is not much point merging it until that's available.
Contact him by email, sure.
…On Thu, Dec 13, 2018 at 11:56 AM Xuechen Liu ***@***.***> wrote:
@danpovey <https://github.com/danpovey> definitely I'd like to but I have
no data so although the fixes I'm suggesting are mainly style-wise I'm not
sure if it will work. Probably I can contact @yfliao
<https://github.com/yfliao> in person (through mail maybe?) and see how
it works. What do you think?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#2474 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/ADJVu6LJI01UKwM9SZxdhIhzAcI97JZbks5u4rEEgaJpZM4UaJ5Y>
.
|
revert commit
results: ```bash $ ./run.sh --stage 10 ./run.sh: extract the results WER: test %WER 61.18 [ 83187 / 135972, 5435 ins, 19162 del, 58590 sub ] exp/mono/decode_test/wer_11_0.0 %WER 40.83 [ 55522 / 135972, 5837 ins, 14160 del, 35525 sub ] exp/tri1/decode_test/wer_14_0.5 %WER 40.47 [ 55028 / 135972, 7068 ins, 12165 del, 35795 sub ] exp/tri2/decode_test/wer_15_0.0 %WER 38.68 [ 52597 / 135972, 7443 ins, 10447 del, 34707 sub ] exp/tri3a/decode_test/wer_14_0.0 %WER 35.83 [ 48720 / 135972, 7536 ins, 9490 del, 31694 sub ] exp/tri4a/decode_test/wer_17_0.0 %WER 39.92 [ 54286 / 135972, 7407 ins, 10832 del, 36047 sub ] exp/tri4a/decode_test.si/wer_15_0.0 %WER 32.17 [ 43748 / 135972, 7387 ins, 8396 del, 27965 sub ] exp/tri5a_cleaned/decode_test/wer_17_0.0 %WER 35.93 [ 48853 / 135972, 6489 ins, 10720 del, 31644 sub ] exp/tri5a_cleaned/decode_test.si/wer_14_0.5 %WER 32.30 [ 43923 / 135972, 6105 ins, 10183 del, 27635 sub ] exp/tri5a/decode_test/wer_17_0.5 %WER 36.10 [ 49092 / 135972, 6258 ins, 11197 del, 31637 sub ] exp/tri5a/decode_test.si/wer_14_0.5 %WER 23.95 [ 32568 / 135972, 4457 ins, 10271 del, 17840 sub ] exp/chain/tdnn_1a_sp/decode_test/wer_10_0.0 %WER 24.67 [ 33546 / 135972, 5358 ins, 7920 del, 20268 sub ] exp/nnet3/tdnn_sp/decode_test/wer_13_0.0 CER: test %WER 54.20 [ 116924 / 215718, 4789 ins, 24777 del, 87358 sub ] exp/mono/decode_test/cer_10_0.0 %WER 32.36 [ 69805 / 215718, 6105 ins, 16033 del, 47667 sub ] exp/tri1/decode_test/cer_13_0.0 %WER 32.11 [ 69260 / 215718, 6289 ins, 15492 del, 47479 sub ] exp/tri2/decode_test/cer_13_0.0 %WER 30.44 [ 65661 / 215718, 6338 ins, 14062 del, 45261 sub ] exp/tri3a/decode_test/cer_13_0.0 %WER 27.55 [ 59439 / 215718, 6267 ins, 13074 del, 40098 sub ] exp/tri4a/decode_test/cer_16_0.0 %WER 31.56 [ 68074 / 215718, 6283 ins, 14101 del, 47690 sub ] exp/tri4a/decode_test.si/cer_13_0.0 %WER 24.16 [ 52121 / 215718, 6690 ins, 10620 del, 34811 sub ] exp/tri5a_cleaned/decode_test/cer_14_0.0 %WER 27.84 [ 60050 / 215718, 6778 ins, 11600 del, 41672 sub ] exp/tri5a_cleaned/decode_test.si/cer_12_0.0 %WER 24.43 [ 52710 / 215718, 6722 ins, 11040 del, 34948 sub ] exp/tri5a/decode_test/cer_14_0.0 %WER 28.04 [ 60494 / 215718, 6760 ins, 11980 del, 41754 sub ] exp/tri5a/decode_test.si/cer_12_0.0 %WER 16.83 [ 36305 / 215718, 4772 ins, 10810 del, 20723 sub ] exp/chain/tdnn_1a_sp/decode_test/cer_9_0.0 %WER 17.24 [ 37188 / 215718, 4880 ins, 9721 del, 22587 sub ] exp/nnet3/tdnn_sp/decode_test/cer_12_0.0 ./run.sh: all done ```
* remove redundant code * simplify local/prepare_data.sh * syntax error * 1. add local/run_cleanup_segmentation.sh 2. some minor fix results: ```bash $ ./run.sh --stage 10 ./run.sh: extract the results WER: test %WER 61.18 [ 83187 / 135972, 5435 ins, 19162 del, 58590 sub ] exp/mono/decode_test/wer_11_0.0 %WER 40.83 [ 55522 / 135972, 5837 ins, 14160 del, 35525 sub ] exp/tri1/decode_test/wer_14_0.5 %WER 40.47 [ 55028 / 135972, 7068 ins, 12165 del, 35795 sub ] exp/tri2/decode_test/wer_15_0.0 %WER 38.68 [ 52597 / 135972, 7443 ins, 10447 del, 34707 sub ] exp/tri3a/decode_test/wer_14_0.0 %WER 35.83 [ 48720 / 135972, 7536 ins, 9490 del, 31694 sub ] exp/tri4a/decode_test/wer_17_0.0 %WER 39.92 [ 54286 / 135972, 7407 ins, 10832 del, 36047 sub ] exp/tri4a/decode_test.si/wer_15_0.0 %WER 32.17 [ 43748 / 135972, 7387 ins, 8396 del, 27965 sub ] exp/tri5a_cleaned/decode_test/wer_17_0.0 %WER 35.93 [ 48853 / 135972, 6489 ins, 10720 del, 31644 sub ] exp/tri5a_cleaned/decode_test.si/wer_14_0.5 %WER 32.30 [ 43923 / 135972, 6105 ins, 10183 del, 27635 sub ] exp/tri5a/decode_test/wer_17_0.5 %WER 36.10 [ 49092 / 135972, 6258 ins, 11197 del, 31637 sub ] exp/tri5a/decode_test.si/wer_14_0.5 %WER 23.95 [ 32568 / 135972, 4457 ins, 10271 del, 17840 sub ] exp/chain/tdnn_1a_sp/decode_test/wer_10_0.0 %WER 24.67 [ 33546 / 135972, 5358 ins, 7920 del, 20268 sub ] exp/nnet3/tdnn_sp/decode_test/wer_13_0.0 CER: test %WER 54.20 [ 116924 / 215718, 4789 ins, 24777 del, 87358 sub ] exp/mono/decode_test/cer_10_0.0 %WER 32.36 [ 69805 / 215718, 6105 ins, 16033 del, 47667 sub ] exp/tri1/decode_test/cer_13_0.0 %WER 32.11 [ 69260 / 215718, 6289 ins, 15492 del, 47479 sub ] exp/tri2/decode_test/cer_13_0.0 %WER 30.44 [ 65661 / 215718, 6338 ins, 14062 del, 45261 sub ] exp/tri3a/decode_test/cer_13_0.0 %WER 27.55 [ 59439 / 215718, 6267 ins, 13074 del, 40098 sub ] exp/tri4a/decode_test/cer_16_0.0 %WER 31.56 [ 68074 / 215718, 6283 ins, 14101 del, 47690 sub ] exp/tri4a/decode_test.si/cer_13_0.0 %WER 24.16 [ 52121 / 215718, 6690 ins, 10620 del, 34811 sub ] exp/tri5a_cleaned/decode_test/cer_14_0.0 %WER 27.84 [ 60050 / 215718, 6778 ins, 11600 del, 41672 sub ] exp/tri5a_cleaned/decode_test.si/cer_12_0.0 %WER 24.43 [ 52710 / 215718, 6722 ins, 11040 del, 34948 sub ] exp/tri5a/decode_test/cer_14_0.0 %WER 28.04 [ 60494 / 215718, 6760 ins, 11980 del, 41754 sub ] exp/tri5a/decode_test.si/cer_12_0.0 %WER 16.83 [ 36305 / 215718, 4772 ins, 10810 del, 20723 sub ] exp/chain/tdnn_1a_sp/decode_test/cer_9_0.0 %WER 17.24 [ 37188 / 215718, 4880 ins, 9721 del, 22587 sub ] exp/nnet3/tdnn_sp/decode_test/cer_12_0.0 ./run.sh: all done ```
|
Dear danpovey and undergdogliu, Alex Hung had revised the recipe, could you check it again to see if it is fine now for merge? Thanks a lot! Best Regards, |
danpovey
reviewed
Mar 5, 2019
egs/formosa/s5/run.sh
Outdated
| if [ $stage -le -1 ]; then | ||
|
|
||
| echo "$0: making mfccs" | ||
| for x in train test; do |
There was a problem hiding this comment.
please merge the run.sh and eval.sh into a single script.
And reduce the amount of vertical whitespace in this script, e.g. at the beginning and end of these if blocks.
egs/formosa/s5/run.sh
Outdated
|
|
||
| if [ $stage -le 6 ]; then | ||
|
|
||
| local/run_cleanup_segmentation.sh |
There was a problem hiding this comment.
it looks to me like maybe you copied the librispeech recipe or something like that.
Did you actually check that the cleanup was helpful? It's kind of slow, and would be better to leave out if it doesn't help
There was a problem hiding this comment.
the average duration of utterances in training set is about 20s, so I copied the script to make them shorter. After processing, it can be reduced to 10s, but there is no significant improvement in CER. I will remove this step.
# %WER 16.83 [ 36305 / 215718, 4772 ins, 10810 del, 20723 sub ] exp/chain/tdnn_1a_sp/decode_test/cer_9_0.0
# %WER 16.44 [ 35459 / 215718, 4216 ins, 11278 del, 19965 sub ] exp/chain/tdnn_1b_sp/decode_test/cer_10_0.0
|
|
||
| # the first splicing is moved before the lda layer, so no splicing here | ||
| relu-batchnorm-layer name=tdnn1 dim=625 | ||
| relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=625 |
There was a problem hiding this comment.
This is an an extremely old configuration. You could copy the current configuration from tedlium s5_r2, that would probably work much better. But not necessary in order to check this in.
egs/formosa/s5/cmd.sh
Outdated
| # conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information, | ||
| # or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl. | ||
|
|
||
| export train_cmd="run.pl --mem 2G" |
There was a problem hiding this comment.
please have it as queue.pl when you check it in (since if you run using run.pl on a grid setup, it's not immediately obvious and is harmful to the machines.)
|
thanks.
…On Tue, Mar 12, 2019 at 9:35 PM Alex Hung ***@***.***> wrote:
***@***.**** commented on this pull request.
------------------------------
In egs/formosa/s5/run.sh
<#2474 (comment)>:
> + # align tri5a
+ steps/align_fmllr.sh --cmd "$train_cmd" --nj $num_jobs \
+ data/train data/lang exp/tri5a exp/tri5a_ali || exit 1;
+
+ # decode tri5
+ (
+ utils/mkgraph.sh data/lang_test exp/tri5a exp/tri5a/graph || exit 1;
+ steps/decode_fmllr.sh --cmd "$decode_cmd" --nj $num_jobs --config conf/decode.config \
+ exp/tri5a/graph data/test exp/tri5a/decode_test || exit 1;
+ )&
+
+fi
+
+if [ $stage -le 6 ]; then
+
+ local/run_cleanup_segmentation.sh
the average duration of utterances in training set is about 20s, so I
copied the script to make them shorter. After processing, it can be reduced
to 10s, but there is no significant improvement in CER. I will remove this
step.
# %WER 16.83 [ 36305 / 215718, 4772 ins, 10810 del, 20723 sub ] exp/chain/tdnn_1a_sp/decode_test/cer_9_0.0
# %WER 16.44 [ 35459 / 215718, 4216 ins, 11278 del, 19965 sub ] exp/chain/tdnn_1b_sp/decode_test/cer_10_0.0
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#2474 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/ADJVuzpEObyIF9gwJiQL09Uq4jr1EaZCks5vWFXKgaJpZM4UaJ5Y>
.
|
* Resnet-style TDNNF * merge and remove scripts * update tdnnf script, remove cleanup_segmentationsh * Update cmd.sh copied from tedlium/s5_r2/cmd.sh
|
Hello @underdogliu @danpovey , |
danpovey
pushed a commit
that referenced
this pull request
Mar 17, 2019
* [src] Change warp-synchronous to cub::BlockReduce (safer but slower) (#3080) * [src] Fix && and || uses where & and | intended, and other weird errors (#3087) * [build] Some fixes to Makefiles (#3088) clang is unhappy with '-rdynamic' in compile-only step, and the switch is really unnecessary. Also, the default location for MKL 64-bit libraries is intel64/. The em64t/ was explained already obsolete by an Intel rep in 2010: https://software.intel.com/en-us/forums/intel-math-kernel-library/topic/285973 * [src] Fixed -Wreordered warnings in feat (#3090) * [egs] Replace bc with perl -e (#3093) * [scripts] Fix python3 compatibility issue in data-perturbing script (#3084) * [doc] fix some typos in doc. (#3097) * [build] Make sure expf() speed probe times sensibly (#3089) * [scripts] Make sure merge_targets.py works in python3 (#3094) * [src] ifdef to fix compilation failure on CUDA 8 and earlier (#3103) * [doc] fix typos and broken links in doc. (#3102) * [scripts] Fix frame_shift bug in egs/swbd/s5c/local/score_sclite_conf.sh (#3104) * [src] Fix wrong assertion failure in nnet3-am-compute (#3106) * [src] Cosmetic changes to natural-gradient code (#3108) * [src,scripts] Python2 compatibility fixes and code cleanup for nnet1 (#3113) * [doc] Small documentation fixes; update on Kaldi history (#3031) * [src] Various mostly-cosmetic changes (copying from another branch) (#3109) * [scripts] Simplify text encoding in RNNLM scripts (now only support utf-8) (#3065) * [egs] Add "formosa_speech" recipe (Taiwanese Mandarin ASR) (#2474) * [egs] python3 compatibility in csj example script (#3123) * [egs] python3 compatibility in example scripts (#3126) * [scripts] Bug-fix for removing deleted words (#3116) The type of --max-deleted-words-kept-when-merging in segment_ctm_edits.py was a string, which prevented the mechanism from working altogether.
chenzhehuai
added a commit
to chenzhehuai/kaldi
that referenced
this pull request
Jun 3, 2019
* [build] Allow configure script to handle package-based OpenBLAS (kaldi-asr#2618) * [egs] updating local/make_voxceleb1.pl so that it works with newer versions of VoxCeleb1 (kaldi-asr#2684) * [egs,scripts] Remove unused --nj option from some scripts (kaldi-asr#2679) * [egs] Fix to tedlium v3 run.sh (rnnlm rescoring) (kaldi-asr#2686) * [scripts,egs] Tamil OCR with training data from yomdle and testing data from slam (kaldi-asr#2621) note: this data may not be publicly available at the moment. we'll work on that. * [egs] mini_librispeech: allow relative pathnames in download_and_untar.sh (kaldi-asr#2689) * [egs] Updating SITW recipe to account for changes to VoxCeleb1 (kaldi-asr#2690) * [src] Fix nnet1 proj-lstm bug where gradient clipping not used; thx:@cbtpkzm (kaldi-asr#2696) * [egs] Update aishell2 recipe to allow online decoding (no pitch for ivector) (kaldi-asr#2698) * [src] Make cublas and cusparse use per-thread streams. (kaldi-asr#2692) This will reduce synchronization overhead when we actually use multiple cuda devices in one process go down drastically, since we no longer synchronize on the legacy default stream. More details here: https://docs.nvidia.com/cuda/cuda-runtime-api/stream-sync-behavior.html * [src] improve handling of low-rank covariance in ivector-compute-lda (kaldi-asr#2693) * [egs] Changes to IAM handwriting-recognition recipe, including BPE encoding (kaldi-asr#2658) * [scripts] Make sure pitch is not included in i-vector feats, in online decoding preparation (kaldi-asr#2699) * [src] fix help message in post-to-smat (kaldi-asr#2703) * [scripts] Fix to steps/cleanup/debug_lexicon.sh (kaldi-asr#2704) * [egs] Cosmetic and file-mode fixes in HKUST recipe (kaldi-asr#2708) * [scripts] nnet1: remove the log-print of args in 'make_nnet_proto.py', thx:mythilisharan@gmail.com (kaldi-asr#2706) * [egs] update README in AISHELL-2 (kaldi-asr#2710) * [src] Make constructor of CuDevice private (kaldi-asr#2711) * [egs] fix sorting issue in aishell v1 (kaldi-asr#2705) * [egs] Add soft links for CNN+TDNN scripts (kaldi-asr#2715) * [build] Add missing packages in extras/check_dependencies.sh (kaldi-asr#2719) * [egs] madcat arabic: clean scripts, tuning, use 6-gram LM (kaldi-asr#2718) * [egs] Update WSJ run.sh: comment out outdated things, add run_tdnn.sh. (kaldi-asr#2723) * [scripts,src] Fix potential issue in scripts; minor fixes. (kaldi-asr#2724) The use of split() in latin-1 encoding (which might be used for other ASCII-compatible encoded data like utf-8) is not right because character 160 (expressed here in decimal) is a NBSP in latin-8 encoding and is also in the range UTF-8 uses for encoding. The same goes for strip(). Thanks @ChunChiehChang for finding the issue. * [egs] add example script for RNNLM lattice rescoring for WSJ recipe (kaldi-asr#2727) * [egs] add rnnlm example on tedlium+lm1b; add rnnlm rescoring results (kaldi-asr#2248) * [scripts] Small fix to utils/data/convert_data_dir_to_whole.sh (RE backups) (kaldi-asr#2735) * [src] fix memory bug in kaldi::~LatticeFasterDecoderTpl(), (kaldi-asr#2737) - found it when running 'latgen-faster-mapped-parallel', - core-dumps from the line: decoder/lattice-faster-decoder.cc:52 -- the line is doing 'delete &(FST*)', i.e. deleting the pointer to FST, instead of deleting the FST itslef, -- bug was probably introduced by refactoring commit d0c68a6 from 2018-09-01, -- after the change the code runs fine... (the unit tests for src/decoder are missing) * [egs] Remove per-utt option from nnet3/align scripts (kaldi-asr#2717) * [egs] Small Librispeech example fix, thanks: Yasasa Tennakoon. (kaldi-asr#2738) * [egs] Aishell2 recipe: turn off jieba's new word discovery in word segmentation (kaldi-asr#2740) * [egs] Add missing file local/join_suffix.py in TEDLIUM s5_r3; thx:anand@sayint.ai (kaldi-asr#2741) * [egs,scripts] Add Tunisian Arabic (MSA) recipe; cosmetic fixes to pbs.pl (kaldi-asr#2725) * [scripts] Fix missing import in utils/langs/grammar/augment_words_txt.py (kaldi-asr#2742) * [scripts] Fix build_const_arpa_lm.sh w.r.t. where <s> appears inside words (kaldi-asr#2745) * [scripts] Slight improvements to decode_score_fusion.sh usability (kaldi-asr#2746) * [build] update configure to support cuda 10 (kaldi-asr#2747) * [scripts] Fix bug in utils/data/resample_data_dir.sh (kaldi-asr#2749) * [scripts] Fix bug in cleanup after steps/cleanup/clean_and_segment_data*.sh (kaldi-asr#2750) * [egs] several updates of the tunisian_msa recipe (kaldi-asr#2752) * [egs] Small fix to Tunisian MSA TDNN script (RE train_stage) (kaldi-asr#2757) * [src,scripts] Batched nnet3 computation (kaldi-asr#2726) This PR adds the underlying utilities for much faster nnet3 inference on GPU, and a command-line binary (and script support) for nnet3 decoding and posterior computation. TBD: a binary for x-vector computation. This PR also contains unrelated decoder speedups (skipping range checks for transition ids... this may cause segfaults when graphs are mismatched). * [build] Add python3 compatibility to install scripts (kaldi-asr#2748) * [scripts] tfrnnlm: Modify TensorFlow flag format for compatibility with recent versions (kaldi-asr#2760) * [egs] fix old style perl regex in egs/chime1/s5/local/chime1_prepare_data.sh (kaldi-asr#2762) * [scripts] Fix bug in steps/cleanup/debug_lexicon.sh (kaldi-asr#2763) * [egs] Add example for Yomdle Farsi OCR (kaldi-asr#2702) * [scripts] debug_lexicon.sh: Fix bug introduced in kaldi-asr#2763. (kaldi-asr#2764) * [egs] add missing online cmvn config in aishell2 (kaldi-asr#2767) * [egs] Add CNN-TDNN-F script for Librispeech (kaldi-asr#2744) * [src] Some minor cleanup/fixes regarding CUDA memory allocation; other small fixes. (kaldi-asr#2768) * [scripts] Update reverberate_data_dir.py so that it works with python3 (kaldi-asr#2771) * [egs] Chime5: fix total number of words for WER calculation (kaldi-asr#2772) * [egs] RNNLMs on Tedlium w/ Google 1Bword: Increase epochs, update results (kaldi-asr#2775) * [scripts,egs] Added phonetisaurus-based g2p scripts (kaldi-asr#2730) Phonetisaurus is much faster to train then sequitur. * [egs] madcat arabic: clean scripts, tuning, rescoring, text localization (kaldi-asr#2716) * [scripts] Enhancements & minor bugfix to segmentation postprocessing (kaldi-asr#2776) * [src] Update gmm-decode-simple to accept ConstFst (kaldi-asr#2787) * [scripts] Update documentation of train_raw_dnn.py (kaldi-asr#2785) * [src] nnet3: extend what descriptors can be parsed. (kaldi-asr#2780) * [src] Small fix to 'fstrand' (make sure args are parsed) (kaldi-asr#2777) * [src,scripts] Minor, mostly cosmetic updates (kaldi-asr#2788) * [src,scripts] Add script to compare alignment directories. (kaldi-asr#2765) * [scripts] Small fixes to script usage messages, etc. (kaldi-asr#2789) * [egs] Update ami_download.sh after changes on Edinburgh website. (kaldi-asr#2769) * [scripts] Update compare_alignments.sh to allow different lang dirs. (kaldi-asr#2792) * [scripts] Change make_rttm.py so output is in determinstic order (kaldi-asr#2794) * [egs] Fixes to yomdle_zh RE encoding direction, etc. (kaldi-asr#2791) * [src] Add support for context independent phones in gmm-init-biphone (for e2e) (kaldi-asr#2779) * [egs] Simplifying multi-condition version of AMI recipe (kaldi-asr#2800) * [build] Fix openblas build for aarch64 (kaldi-asr#2806) * [build] Make CUDA_ARCH configurable at configure-script level (kaldi-asr#2807) * [src] Print maximum memory stats in CUDA allocator (kaldi-asr#2799) * [src,scripts] Various minor code cleanups (kaldi-asr#2809) * [scripts] Fix handling of UTF-8 in filenames, in wer_per_spk_details.pl (kaldi-asr#2811) * [egs] Update AMI chain recipes (kaldi-asr#2817) * [egs] Improvements to multi_en tdnn-opgru/lstm recipes (kaldi-asr#2824) * [scripts] Fix initial prob of silence when lexicon has silprobs. Thx:@agurianov (kaldi-asr#2823) * [scripts,src] Fix to multitask nnet3 training (kaldi-asr#2818); cosmetic code change. (kaldi-asr#2827) * [scripts] Create shared versions of get_ctm_conf.sh, add get_ctm_conf_fast.sh (kaldi-asr#2828) * [src] Use cuda streams in matrix library (kaldi-asr#2821) * [egs] Add online-decoding recipe to aishell1 (kaldi-asr#2829) * [egs] Add DIHARD 2018 diarization recipe. (kaldi-asr#2822) * [egs] add nnet3 online result for aishell1 (kaldi-asr#2836) * [scripts] RNNLM scripts: don't die when features.txt is not present (kaldi-asr#2837) * [src] Optimize cuda allocator for multi-threaded case (kaldi-asr#2820) * [build] Add cub library for cuda projects (kaldi-asr#2819) not needed now but will be in future. * [src] Make Cuda allocator statistics visible to program (kaldi-asr#2835) * [src] Fix bug affecting scale in GeneralDropoutComponent (non-continuous case) (kaldi-asr#2815) * [build] FIX kaldi-asr#2842: properly check $use_cuda against false. (kaldi-asr#2843) * [doc] Add note about OOVs to data-prep. (kaldi-asr#2844) * [scripts] Allow segmentation with nnet3 chain models (kaldi-asr#2845) * [build] Remove -lcuda from cuda makefiles which breaks operation when no driver present (kaldi-asr#2851) * [scripts] Fix error in analyze_lats.sh for long lattices (replace awk with perl) (kaldi-asr#2854) * [egs] add rnnlm recipe for librispeech (kaldi-asr#2830) * [build] change configure version from 9 to 10 (kaldi-asr#2853) (kaldi-asr#2855) * [src] fixed compilation errors when built with --DOUBLE_PRECISION=1 (kaldi-asr#2856) * [build] Clarify instructions if cub is not found (kaldi-asr#2858) * [egs] Limit MFCC feature extraction job number in Dihard recipe (kaldi-asr#2865) * [egs] Added Bentham handwriting recognition recipe (kaldi-asr#2846) * [src] Share roots of different tones of phones aishell (kaldi-asr#2859) * [egs] Fix path to sequitur in commonvoice egs (kaldi-asr#2868) * [egs] Update reverb recipe (kaldi-asr#2753) * [scripts] Fix error while analyzing lattice (parsing bugs) (kaldi-asr#2873) * [src] Fix memory leak in OnlineCacheFeature; thanks @Worldexe (kaldi-asr#2872) * [egs] TIMIT: fix mac compatibility of sed command (kaldi-asr#2874) * [egs] mini_librispeech: fixing some bugs and limiting repeated downloads (kaldi-asr#2861) * [src,scripts,egs] Speedups to GRU-based networks (special components) (kaldi-asr#2712) * [src] Fix infinite recursion with -DDOUBLE_PRECISION=1. Thx: @hwiorn (kaldi-asr#2875) (kaldi-asr#2876) * Revert "[src] Fix infinite recursion with -DDOUBLE_PRECISION=1. Thx: @hwiorn (kaldi-asr#2875) (kaldi-asr#2876)" (kaldi-asr#2877) This reverts commit 84435ff. * Revert "Revert "[src] Fix infinite recursion with -DDOUBLE_PRECISION=1. Thx: @hwiorn (kaldi-asr#2875) (kaldi-asr#2876)" (kaldi-asr#2877)" (kaldi-asr#2878) This reverts commit b196b7f. * Revert "[src] Fix memory leak in OnlineCacheFeature; thanks @Worldexe" (kaldi-asr#2882) the fix was buggy. apologies. * [src] Remove unused code that caused Windows compile failure. Thx:@btiplitz (kaldi-asr#2881) * [src] Really fix memory leak in online decoding; thx:@Worldexe (kaldi-asr#2883) * [src] Fix Windows cuda build failure (use C++11 standard include) (kaldi-asr#2880) * [src] Add #include that caused build failure on Windows (kaldi-asr#2886) * [scripts] Fix max duration check in sad_to_segments.py (kaldi-asr#2889) * [scripts] Fix speech duration calculation in sad_to_segments.py (kaldi-asr#2891) * [src] Fix Windows build problem (timer.h) (kaldi-asr#2888) * [egs] add HUB4 spanish tdnn-f and cnn-tdnn script (kaldi-asr#2895) * [egs] Fix Aishell2 dict prepare bug; should not affect results (kaldi-asr#2890) * [egs] Self-contained example for KWS for mini_librispeech (kaldi-asr#2887) * [egs,scripts] Fix bugs in Dihard 2018 (kaldi-asr#2897) * [scripts] Check last character of files to match with newline (kaldi-asr#2898) * [egs] Update Librispeech RNNLM results; use correct training data (kaldi-asr#2900) * [scripts] RNNLM: old iteration model cleanup; save space (kaldi-asr#2885) * [scripts] Make prepare_lang.sh cleanup beforehand (prevents certain failures) (kaldi-asr#2906) * [scripts] Expose dim-range-node at xconfig level (kaldi-asr#2903) * [scripts] Fix bug related to multi-task in train_raw_rnn.py (kaldi-asr#2907) [scripts] Fix bug related to multi-task in train_raw_rnn.py. Thx:tessfu2001@gmail.com * [scripts] Cosmetic fix/clarification to utils/prepare_lang.sh (kaldi-asr#2912) * [scripts,egs] Added a new lexicon learning (adaptation) recipe for tedlium, in accordance with the IS17 paper. (kaldi-asr#2774) * [egs] TDNN+LSTM example scripts, with RNNLM, for Librispeech (kaldi-asr#2857) * [src] cosmetic fix in nnet1 code (kaldi-asr#2921) * [src] Fix incorrect invocation of mutex in nnet-batch-compute code (kaldi-asr#2932) * [egs,minor] Fix typo in comment in voxceleb script (kaldi-asr#2926) * [src,egs] Mostly cosmetic changes; add some missing includes (kaldi-asr#2936) * [egs] Fix path of rescoring binaries used in tfrnnlm scripts (kaldi-asr#2941) * [src] Fix bug in nnet3-latgen-faster-batch for determinize=false (kaldi-asr#2945) thx: Maxim Korenevsky. * [egs] Add example for rimes handwriting database; Madcat arabic script cleanup (kaldi-asr#2935) * [egs] Add scripts for yomdle korean (kaldi-asr#2942) * [build] Refactor/cleanup build system, easier build on ubuntu 18.04. (kaldi-asr#2947) note: if this breaks someone's build we'll have to debug it then. * [scripts,egs] Changes for Python 2/3 compatibility (kaldi-asr#2925) * [egs] Add more modern DNN recipe for fisher_callhome_spanish (kaldi-asr#2951) * [scripts] switch from bc to perl to reduce dependencies (diarization scripts) (kaldi-asr#2956) * [scripts] Further fix for Python 2/3 compatibility (kaldi-asr#2957) * [egs] Remove no-longer-existing option in tedlium_r3 recipe (kaldi-asr#2959) * [build] Handle dependencies for .cu files in addition to .cc files (kaldi-asr#2944) * [src] remove duplicate test mode option from class GeneralDropoutComponent (kaldi-asr#2960) * [egs] Fix minor bugs in WSJ's flat-start/e2e recipe (kaldi-asr#2968) * [egs] Fix to BSD compatibility of TIMIT data prep (kaldi-asr#2966) * [scripts] Fix RNNLM training script problem (chunk_length was ignored) (kaldi-asr#2969) * [src] Fix bug in lattice-1best.cc RE removing insertion penalty (kaldi-asr#2970) * [src] Compute a separate avg (start, end) interval for each sausage word (kaldi-asr#2972) * [build] Move nvcc verbose flag to proper location (kaldi-asr#2962) * [egs] Fix mini_librispeech download_lm.sh crash; thx:chris.keith.johnson@gmail.com (kaldi-asr#2974) * [egs] minor fixes related to python2 vs python3 differences (kaldi-asr#2977) * [src] Small fix in test code, avoid spurious failure (kaldi-asr#2978) * [egs] Fix CSJ data-prep; minor path fix for USB version of data (kaldi-asr#2979) * [egs] Add paper ref to README.txt in reverb example (kaldi-asr#2982) * [egs] Minor fixes to sitw recipe (fix problem introdueced in kaldi-asr#2925) (kaldi-asr#2985) * [scripts] Fix bug introduced in kaldi-asr#2957, RE integer division (kaldi-asr#2986) * [egs] Update WSJ flat-start chain recipes to use TDNN-F not TDNN+LSTM (kaldi-asr#2988) * [scripts] Fix typo introduced in kaldi-asr#2925 (kaldi-asr#2989) * [build] Modify Makefile and travis script to fix Travis failures (kaldi-asr#2987) * [src] Simplification and efficiency improvement in ivector-plda-scoring-dense (kaldi-asr#2991) * [egs] Update madcat Arabic and Chinese egs, IAM (kaldi-asr#2964) * [src] Fix overflow bug in convolution code (kaldi-asr#2992) * [src] Fix nan issue in ctm times introduced in kaldi-asr#2972, thx: @vesis84 (kaldi-asr#2993) * [src] Fix 'sausage-time' issue which occurs with disabled MBR decoding. (kaldi-asr#2996) * [egs] Add scripts for yomdle Russian (OCR task) (kaldi-asr#2953) * [egs] Simplify lexicon preparation in Fisher callhome Spanish (kaldi-asr#2999) * [egs] Update GALE Arabic recipe (kaldi-asr#2934) * [egs] Remove outdated NN results from Gale Arabic recipe (kaldi-asr#3002) * [egs] Add RESULTS file for the tedlium s5_r3 (release 3) setup (kaldi-asr#3003) * [src] Fixes to grammar-fst code to handle LM-disambig symbols properly (kaldi-asr#3000) thanks: armando.muscariello@gmail.com * [src] Cosmetic change to mel computation (fix option string) (kaldi-asr#3011) * [src] Fix Visual Studio error due to alternate syntactic form of noreturn (kaldi-asr#3018) * [egs] Fix location of sequitur installation (kaldi-asr#3017) * [src] Fix w/ ifdef Visual Studio error from alternate syntactic form noreturn (kaldi-asr#3020) * [egs] Some fixes to getting data in heroico recipe (kaldi-asr#3021) * [egs] BABEL script fix: avoid make_L_align.sh generating invalid files (kaldi-asr#3022) * [src] Fix to older online decoding code in online/ (OnlineFeInput; was broken by commit cc2469e). (kaldi-asr#3025) * [script] Fix unset bash variable in make_mfcc.sh (kaldi-asr#3030) * [scripts] Extend limit_num_gpus.sh to support --num-gpus 0. (kaldi-asr#3027) * [scripts] fix bug in utils/add_lex_disambig.pl when sil-probs and pron-probs used (kaldi-asr#3033) bug would likely have resulted in determinization failure (only when not using word-position-dependent phones). * [egs] Fix path in Tedlium r3 rnnlm training script (kaldi-asr#3039) * [src] Thread-safety for GrammarFst (thx:armando.muscariello@gmail.com) (kaldi-asr#3040) * [scripts] Cosmetic fix to get_degs.sh (kaldi-asr#3045) * [egs] Small bug fixes for IAM and UW3 recipes (kaldi-asr#3048) * [scripts] Nnet3 segmentation: fix default params (kaldi-asr#3051) * [scripts] Allow perturb_data_dir_speed.sh to work with utt2lang (kaldi-asr#3055) * [scripts] Make beam in monophone training configurable (kaldi-asr#3057) * [scripts] Allow reverberate_data_dir.py to support unicode filenames (kaldi-asr#3060) * [scripts] Make some cleanup scripts work with python3 (kaldi-asr#3054) * [scripts] bug fix to nnet2->3 conversion, fixes kaldi-asr#886 (kaldi-asr#3071) * [src] Make copies occur in per-thread default stream (for GPUs) (kaldi-asr#3068) * [src] Add GPU version of MergeTaskOutput().. relates to batch decoding (kaldi-asr#3067) * [src] Add device options to enable tensor core math mode. (kaldi-asr#3066) * [src] Log nnet3 computation to VLOG, not std::cout (kaldi-asr#3072) * [src] Allow upsampling in compute-mfcc-feats, etc. (kaldi-asr#3014) * [src] fix problem with rand_r being undefined on Android (kaldi-asr#3037) * [egs] Update swbd1_map_words.pl, fix them_1's -> them's (kaldi-asr#3052) * [src] Add const overload OnlineNnet2FeaturePipeline::IvectorFeature (kaldi-asr#3073) * [src] Fix syntax error in egs/bn_music_speech/v1/local/make_musan.py (kaldi-asr#3074) * [src] Memory optimization for online feature extraction of long recordings (kaldi-asr#3038) * [build] fixed a bug in linux_configure_redhat_fat when use_cuda=no (kaldi-asr#3075) * [scripts] Add missing '. ./path.sh' to get_utt2num_frames.sh (kaldi-asr#3076) * [src,scripts,egs] Add count-based biphone tree tying for flat-start chain training (kaldi-asr#3007) * [scripts,egs] Remove sed from various scripts (avoid compatibility problems) (kaldi-asr#2981) * [src] Rework error logging for safety and cleanliness (kaldi-asr#3064) * [src] Change warp-synchronous to cub::BlockReduce (safer but slower) (kaldi-asr#3080) * [src] Fix && and || uses where & and | intended, and other weird errors (kaldi-asr#3087) * [build] Some fixes to Makefiles (kaldi-asr#3088) clang is unhappy with '-rdynamic' in compile-only step, and the switch is really unnecessary. Also, the default location for MKL 64-bit libraries is intel64/. The em64t/ was explained already obsolete by an Intel rep in 2010: https://software.intel.com/en-us/forums/intel-math-kernel-library/topic/285973 * [src] Fixed -Wreordered warnings in feat (kaldi-asr#3090) * [egs] Replace bc with perl -e (kaldi-asr#3093) * [scripts] Fix python3 compatibility issue in data-perturbing script (kaldi-asr#3084) * [doc] fix some typos in doc. (kaldi-asr#3097) * [build] Make sure expf() speed probe times sensibly (kaldi-asr#3089) * [scripts] Make sure merge_targets.py works in python3 (kaldi-asr#3094) * [src] ifdef to fix compilation failure on CUDA 8 and earlier (kaldi-asr#3103) * [doc] fix typos and broken links in doc. (kaldi-asr#3102) * [scripts] Fix frame_shift bug in egs/swbd/s5c/local/score_sclite_conf.sh (kaldi-asr#3104) * [src] Fix wrong assertion failure in nnet3-am-compute (kaldi-asr#3106) * [src] Cosmetic changes to natural-gradient code (kaldi-asr#3108) * [src,scripts] Python2 compatibility fixes and code cleanup for nnet1 (kaldi-asr#3113) * [doc] Small documentation fixes; update on Kaldi history (kaldi-asr#3031) * [src] Various mostly-cosmetic changes (copying from another branch) (kaldi-asr#3109) * [scripts] Simplify text encoding in RNNLM scripts (now only support utf-8) (kaldi-asr#3065) * [egs] Add "formosa_speech" recipe (Taiwanese Mandarin ASR) (kaldi-asr#2474) * [egs] python3 compatibility in csj example script (kaldi-asr#3123) * [egs] python3 compatibility in example scripts (kaldi-asr#3126) * [scripts] Bug-fix for removing deleted words (kaldi-asr#3116) The type of --max-deleted-words-kept-when-merging in segment_ctm_edits.py was a string, which prevented the mechanism from working altogether. * [scripts] Add fix regarding num-jobs for segment_long_utterances*.sh(kaldi-asr#3130) * [src] Enable allow_{upsample,downsample} with online features (kaldi-asr#3139) * [src] Fix bad assert in fstmakecontextsyms (kaldi-asr#3142) * [src] Fix to "Fixes to grammar-fst & LM-disambig symbols" (kaldi-asr#3000) (kaldi-asr#3143) * [build] Make sure PaUtils exported from portaudio (kaldi-asr#3144) * [src] cudamatrix: fixing a synchronization bug in 'normalize-per-row' (kaldi-asr#3145) was only apparent using large matrices * [src] Fix typo in comment (kaldi-asr#3147) * [src] Add binary that functions as a TCP server (kaldi-asr#2938) * [scripts] Fix bug in comment (kaldi-asr#3152) * [scripts] Fix bug in steps/segmentation/ali_to_targets.sh (kaldi-asr#3155) * [scripts] Avoid holding out more data than the requested num-utts (due to utt2uniq) (kaldi-asr#3141) * [src,scripts] Add support for two-pass agglomerative clustering. (kaldi-asr#3058) * [src] Disable unget warning in PeekToken (and other small fix) (kaldi-asr#3163) * [build] Add new nvidia tools to windows build (kaldi-asr#3159) * [doc] Fix documentation errors and add more docs for tcp-server decoder (kaldi-asr#3164)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Dear Dan and Kaldi community,
Since, the language habits of Taiwanese people are different from other Mandarin speakers (both accents and cultures). WE are now running a corpus collection project "Formosa Speech in the Wild (FSW)" to boost the development of Taiwan-specific speech recognition techniques.
FSW is a Large-Scale database of Real-Life/Multi-Gene/Taiwanese Spontaneous Speech Collected and Transcribed from various sources (radio, TV, open courses, etc.). And we would like to provide a baseline recipes for new researchers/students to develop their systems easily and quickly.
This recips is modified from AiShell and based ont the "NER-Trs-Vol1" corpus (126.6 hours speech selected from National Education Radio archive, which could be applied for "Non-Commenrical Use Only". For more details, please visit:
Any comments and suggestions are welcomed.
Thanks a lot !
YuanFu Liao