Difference between revisions of "Lantian Li 2016-04-18"

From cslt Wiki
Jump to: navigation, search
Line 3: Line 3:
 
:* Joint-training. [http://192.168.0.51:5555/cgi-bin/cvss/cvss_request.pl?account=lilt&step=view_request&cvssid=516 here].
 
:* Joint-training. [http://192.168.0.51:5555/cgi-bin/cvss/cvss_request.pl?account=lilt&step=view_request&cvssid=516 here].
  
  # Finish the TDNN-dvector system, train two nnetworks (one is based on WSJ_Train, the other is based on Fisher-1000), test on both WSJ_test and SRE08.
+
# Finish the TDNN-dvector system, train two nnetworks (one is based on WSJ_Train, the other is based on Fisher-1000), test on both WSJ_test and SRE08.
  
  # All the i-vector baseline experiments have been done! (Fisher-1000, Fisher-2000, Fisher-4000 and WSJ).
+
# All the i-vector baseline experiments have been done! (Fisher-1000, Fisher-2000, Fisher-4000 and WSJ).
 
   
 
   
 
:* Replay detection. Complete the alignment between normal / replayed utterances.
 
:* Replay detection. Complete the alignment between normal / replayed utterances.

Revision as of 04:59, 18 April 2016

  • Weekly Report
  • Joint-training. here.
  1. Finish the TDNN-dvector system, train two nnetworks (one is based on WSJ_Train, the other is based on Fisher-1000), test on both WSJ_test and SRE08.
  1. All the i-vector baseline experiments have been done! (Fisher-1000, Fisher-2000, Fisher-4000 and WSJ).
  • Replay detection. Complete the alignment between normal / replayed utterances.
  • Deep speaker embedding. Hold
  • This Week
  • Joint-training (speaker recognition) -> Finsher the LSTM-dvector (based on nnet3-TZY).
  • Deep speaker embedding -> start! cooperate with LTY.