Week 102

What I pre-planned

• bhai
  • train a ASR model on kaldi architecture
  • reevaluate all approaches and think ootb
  • continue the research with faster outcome deadlines
  • replace battery/optimize firmware power issues
• challenged myself to make the medusa from Dr. Stone in CAD

What new did I learn

• how vosk works so well with such a small footprint
  • kaldi architecture basics and training my own ASR on colab using my curated datasets
• completed work annual compliance training and it was comoparable to a netflix show lol
• math behind a trfoil knot, simple ways/tricks to make it from scratch
  • new CAD tools learnt like 3d spline, got a lot faster on onshape shortcuts
• esp32cam modules are quite bad at the only thing they are meant to do well. anywhere else they are not sufficient
• read a really fine blog on founders and how startups work in a fun and engaging story about an oven startup
• learnt how to play zoltraak - frieren on piano, flute and ukelele

Where/How did I implement it

• mlid model
  • trained full GMM-HMM base graphs and tested deconding output and got a bad 70%+ WER
  • tried various approaches even to the extent of reverse engineering the vosk c++ source code to inject an LID model on the HCLG graphs, but stepped back
  • finally trained a zipa based phoneme model to go back to square one to rethink about the problem in a different way. this was quite successful and still refining. expected to get a ~70MB model for full frontend model
• medusa CAD
  • took me 4 hours of effort in total to make the full replica in CAD
• made an excalidraw workspace for myself to solve my everyday problem of losing excalidraw’s and wanting to organize my diagrams for work and projects

Any memorable moments of the week

• the expanse s4 finished
• made another one of my collectibles on a camera based augmented piano concept
• played a lot of peak and minecraft with shank and sanskar till late night
• went shopping with sis and had a nice dinner
• finally got some hope on the research for getting edge transcription working (fingers crossed)
  • teased the wearable and waitlist site on insta