“ASR Status Report 2017-6-5”版本间的差异

• created a database for speech overlap
• experiments with speech overlap

here

• align data before overlap
• test in lstm,tdnn

• sparse dnn with different threshold

• continue the experiment

• mix-lingual structure (without LID)
  • the performance of thch30 is bad than baseline 24.84
  • the performance of kazak is better than baseline 24.37
  • the performance of uyghur is very bad 109.01
• result analysis
  • pdf mix
  • the language model of uyghur is too weak

• work for APSIPA2017

• OLR17(Ten lans) i-vector / L-vector + svm baseline
• "cough, laugh, hello" data for speaker identification

• go on the two tasks.

• End-to-end ASV [1]
• Deep speaker feature learning.

• Speaker diarization.

• basic recipes for OLR17 and MixASR17

• improve the recipes and publish

• after introducing beamforming, zhiyong lets me learn something about how to sound localization

• Create a database for overlapping speech

• work for Apsipa17
  • thchs baseline
  • kazakh baseline
  • uyghur baseline
  • 3 language merge GMM baseline

• work for Apsipa17
• graduation thesis

• Submit the Interspeech paper.
• Deep speaker feature learning with a larger ASR model.
• Prepare the end-to-end ASV.

• End-to-end ASV.

• Experiments for OC17 mix-asr challenge,baseline done.
• Spk feature as LID input.
• Data preparation for AP17 orl challenge.

• Two challenges done.