“2024-01-06”版本间的差异

• Polish high-school AI handbook

• Analyze and summarize the previous experiments
• Project things

• continue on overlap asr task
• submit HUADANIAN challenge proposal

• The general content of the paper has been completed

(The method part still needs to be formalized, and the experimental part needs to be carefully thought about how to write it)[1]

• SSL loss in TSE[2]

• Finish CN-AV-HuBERT code modify, already test on small data
• Simplified the post process of collected data
• CVS3 audit for December

• prepare training data for AVHuBERT, cp it to GA
• Some function demand on process server need to be solve
• Data Audit

• Organize my previous Experiment Result[3]
• Add correct loss, but code encounters some bugs, so i fix it and train again(still training)

• IS25 Proposal almost done(Double check and stat result)

• NS: multi-enroll+margin bce: 1.638% -> 1.521% (5w spk)

• Try using attention instead of FiLM:
  • valid loss: query emb cross-attn: -8.065 vs. query feature cross-attn: -9.127 vs. FiLM: -10.054
  • read a paper show that FiLM is better than attention in sound sep task
• Huawei project testing

• final exam
• wrote the remaining course papers

• finished all course homework and final exams

• Finish Memory mechanism and Policy fusion part, start run and do further debug

• Text enroll KWS model training without CTC loss.

• Midterm defense
• Assessment exam for application
• Experiment on improving Oromo ASR[4]

• report my recent work
• read paper about functionally invariant path, which was reported by Prof. Wang.

• KWS:
  • pre-prod routine work for zh48+yichang5 models.
  • android demo backed with gRPC services deployed on peacock01.
• QAT: exploration with little progress.