IWSLT

• English ASR
  • Acoustic model (Peter)
    • Base model: TED; tandem, hybrid, spec+MFCC/PLP features; MLAN (switchboard/AMI/BBC); HTK+KALDI
    • Contrasts: increased base training data (late stage); Pawel-adaptation
  • Language model
    • slightly more in-domain + b/g model - n-grams (Fergus)
    • RNNs - Andrew Liu's training and decoding (Toms)
  • Segmentation (Mark)
    • refined system; +NNs
• Italian/German ASR
  • revisit light supervision (alllow noisier matches)
  • Downloaded TED talks (DE/IT) - look into commercial transcription
  • German braodacast from RBM
  • Euronews, Europarl
  • ask Pervoice for IT broadcast subtitles
  • LM data following IWSLT links

BBC/Sky

• BBC
  • Waiting on Cam 1-week-seg (23 June?)
  • Dry run on existing seg
  • LM on subtitle data - 3/4-gram (20 June) + gigaword/news-crawl
  • try kenlm
• Sky
  • waiting for raw subtitle files, should match audio time better
  • Sky News transcriptions
  • dev and eval sets being prepared

-- SteveRenals - 11 Jun 2014