aprilv0 model accuracy

[abb128]

Issues that were caused by poor text normalization in the training data:

• Number transformation - "five twenty" may become "five hundred and twenty"
• Years from 2000 to 2009 are completely wrong
• Decimal numbers are very unreliable ("two point five six" -> "two five hundred and six")
• Currency often gets skipped ("twenty five dollars" -> "twenty five")

Other issues:

• Word "aberrations" in particular seems to cause repeating of r for some reason
• "KDE" often becomes "kitty" or "katy" or similar
• Speaker change can cause poor accuracy for a bit
• The model can sometimes be very vulgar

If you have any additional issues please add a comment, a link to some videos where it performs poorly may help.

[kdwk]

• Microphone: the word "variance" cannot be picked up. It keeps transcribing as "variant" or "variants"
• Microphone: repeating the word "something" will only result in one "something". Try "something something". This doesn't happen with e.g. "hello hello"

[kdwk]

• Microphone: "tactics' cannot be picked up. Keeps transcribing as "technics", which I believe is not a word

[whitequark]

This video has a fair amount of weird transcriptions (what the heck is "oxadizer"? that's not even a word), but I wanted to highlight the mistranscription of "nineteen eighty eight" as "198".