Week 101

What I pre-planned

• bhai
  • continue research on different model architectures for lid
  • check if transcription can be confirmed

What new did I learn

• lots of research stuff for bhai
  • TDNN(f)
• some interesting networking for setting up minecraft
• deep discharge damage to bhAI battery, charging nad discharge cycles
  • optimizing low power mode and doing analysis
• realised how the researchers spent their time trying different approaches to optimize some eval
  • it is fun but needs patience

Where/How did I implement it

• mlid model
  • the good models I got depend on a 1gb phoneme frontend which I don’t like
  • trained teacher student model to distill the 1gb model to 30MB even 15MB
  • got decent outcome but not upto the satisfying accuracy, keeps forgetting gujarati phonemes
  • now researching from start again to see if I can reverse engineer vosk and train my own kaldi based model
  • identified complete similarity scores of many indic languages

Any memorable moments of the week

• the expanse s3 finished
• prepared dance for upcoming friends wedding, was fun
• checked my lifelong dream of playing minecraft with friends