This is the CSU-K toolkit for spoken call shared task 2. It contains several scripts, models and other data. Those files have been used to develop our two CSU-K ST2 systems, which you can find here:
Furthermore there are (not officially published) papers, which explain the Toolkit and our DNN-Based-System in more detail:
If you use parts of this toolkit (including papers), please make sure to cite our paper at Interspeech 2018. See https://www.isca-speech.org/archive/Interspeech_2018/abstracts/1000.html.
Cite as: Jülg, D., Kunstek, M., Freimoser, C.P., Berkling, K., Qian, M. (2018) The CSU-K Rule-Based System for the 2nd Edition Spoken CALL Shared Task. Proc. Interspeech 2018, 2359-2363, DOI: 10.21437/Interspeech.2018-1000.
@inproceedings{Jülg2018,
author={Dominik Jülg and Mario Kunstek and Cem Philipp Freimoser and Kay Berkling and Mengjie Qian},
title={The CSU-K Rule-Based System for the 2nd Edition Spoken CALL Shared Task},
year=2018,
booktitle={Proc. Interspeech 2018},
pages={2359--2363},
doi={10.21437/Interspeech.2018-1000},
url={http://dx.doi.org/10.21437/Interspeech.2018-1000}
}
| Order | ScriptName | Input | Output | notes |
|---|---|---|---|---|
| 1, c | generateDistanceVectors.py | Doc2Vec, ReferenceGrammer, RecResult | X-Vectors as CSV | |
| a | generateDoc2VecModel.py | Dir to Training Text | a Doc2Vec Model | |
| b | split_file.py | Training CSV, Modulator | Train and Test | you can use another split script |
| d | makeLabelsFromCSV.py | Test or Test CSV | Y-Vectors as CSV | |
| opt | splitBasedOnIdFile.py | Txt containing ids, file you want to split | Two files one only contraining ids from input txt | alternative to b |
| opt | split_csv_by_mod_10.py | Training CSV | Train and Test where Test ist 10% of Train | alternative to b |
| opt | makeLabelsFromCSV_v2.py | Test CSV from ST2 | Y-Vectors as CSV | |
| help | validateUniqueIds.py | As many files as you like | IDs that are shared | check if you have clean train and test |
| help | extractBaseOnId.py | Txt containing ids, some file | Two files like the secound input file. But one just containg the ids from txt | if you have messed up something |
| help | merge_sc_1_training_data_with_asr.py | Train CSV, ASR, ASR_Data | sc1 train entries are have asr output as rec_result | |
| help | merge_sc_2_training_data_with_asr.py | Test CSV, ASR, ASR_Data | sc2 train entries,are have asr output as rec_result | |
| help | normalize_sc_1_test.py | ABC CSV files, Outputfile | CSV file formatted like sc1_train | |
| help | normalize_sc_2_abc.py | Test CSV, Outputfile | CSV file formatted like sc1_train | |
| help | get_intersected_ids.py | A lot of CSV | intersacting ids | |
| help | fix_id_in_csv.py | Data CSV, Outputfile | same as input but with fixed ids | |
| exp | generateWord2VecModel.py | Dir with Training Text | a Word2Vec Model | |
| exp | makeMeaningLabelsFromCSV.py | Data CSV | File with Meaning Labels | |
| exp | makeMeaningLabelsFromCSV.py | Data CSV | File with Meaning Labels | |
| exp | makeGrammarLabelFromCSV.py | Data CSV | File with Grammer Labels | |
| exp | split_ABC_intelligent.py | Data CSV A B and C | Several files needed by ASR. Creates scp files, spk2utt, utt2spk and text | Use with caution. You should already have most of those files |