Skip to content

Latest commit

 

History

History
112 lines (92 loc) · 4 KB

Systems.md

File metadata and controls

112 lines (92 loc) · 4 KB

Systems

Cascaded Systems

ctc-tedlium2.smallTED

  • This system combines the ctc-tedlium2 system and the small TED system.
  • Languages:
    • English-German
  • Training: There is no script to train this system directly, but you can train both system individually with the scripts in that folder
  • Download: You can download the system with the following script (remember to set the appropriate source and target language to select the correct MT system, if you haven't done that yet):
export sl=en
export tl=de
/opt/SLT.KIT/systems/ctc-tedlium2.smallTED/Download.sh
  • For the individual components used in this system refer to the components of the two individual systems
  • Testing:
/opt/SLT.KIT/systems/ctc-tedlium2.smallTED/Test.sh ${testset}
  • Output of the ASR component: /data/ctc/eval/${testset}.s
  • Segmented and punctuated transcript: /data/monoTransPrepro/eval/ctc.${testset}.s
  • Translation: /data/mt/eval/ctc.${testset}.t
  • Results: /results/ctc-tedlium2.smallTED/ctc.${testset}/Summary.md

ctc-tedlium2.midSize

  • Same as ctc-tedlium2.smallTED except using the midSize system for MT and segmentation/punctuation

ASR Systems

ctc-tedlium2

  • ASR system using the CTC loss trained on Byte Pair Units
  • Language: English
  • Training: The standard network uses BPE units created with 300 merge operations. You can retrain this model with BPE units created with 10000 merge operations:
# Train 300
/opt/SLT.KIT/systems/ctc-tedlium2/Train.sh
# Train 10000
/opt/SLT.KIT/systems/ctc-tedlium2/Train.sh 10000
  • Download: You can download the models (both 300 and 10000) using the following script:
/opt/SLT.KIT/systems/ctc-tedlium2/Download.sh
  • Testing (10000 model):
/opt/SLT.KIT/systems/ctc-tedlium2/Test.sh ${testset}
  • Data
    • Model: /model/ctc/bpe${bpe_steps}.mdl, /model/ctc/bpe${bpe_steps}.log, /model/ctc/units${bpe_steps}.json
    • Test Output: /data/ctc/eval/${testset}.s, /data/ctc/eval/${testset}.logits

las-tedlium2

  • ASR system using listen-attend-spell architecture
  • Language: English
  • Training: You can train the system with the scripts
/opt/SLT.KIT/systems/las-tedlium2/Train.sh
  • Download: You can download the system with the scripts
/opt/SLT.KIT/systems/las-tedlium2/Download.sh

Testing:

/opt/SLT.KIT/systems/las-tedlium2/Test.sh ${testset}

MT/Sentence segmentation

smallTED

  • MT and sentence segmentation trained on the TED corpus
  • Languages:
    • German-English
    • English-German
    • English-French
  • Components:
    • prepro: Preprocessing of the data; Tokenization, true casing
    • monoTransPrepro: Preprocessing for the training of the monolingual translation system. Removing of case and puncuation information, target representation by labels. For details [1]
    • monTrans: NMT-based monolingual translations system to predict punctuation using OpenNMT-py
    • mt: OpenNMT-py based NMT system
  • Training: You can train the system with the scripts
/opt/SLT.KIT/systems/smallTED/Train.sh
  • Download: You can download the system with the scripts
/opt/SLT.KIT/systems/smallTED/Download.sh
  • Testing:
/opt/SLT.KIT/systems/smallTED/Test.sh ${testset}
  • Segmented and punctuated transcript of the provided ASR output in test data: /data/monoTransPrepro/eval/${testset}.s
  • Translation: /data/mt/eval/${testset}.t
  • Translation of the manual transcript: /data/mt/eval/manualTranscript.${testset}.t
  • Results /results/smallTED/${setset}/Summary.md

midSize

[1] Cho, E., Niehues, J., Waibel, A. (2017). NMT-based Segmentation and Punctuation Insertion for Real-time Spoken Language Translation. In Proceedings of the 18th Annual Conference of the International Speech Communication Association (Interspeech 2017). Stockholm, Sweden.