Speling werds egspressively with rrkurrent nuril nedwirques!
summary
Trains recurrent sequence-to-sequence neural networks to map phonemes↔letters using the CMU Pronouncing Dictionary / ARPAbet as seed data and a Jupyter notebook interface for manipulation. Demonstrates altering output by changing letter-probabilities (removing vowels, suppressing YZ), adding Gaussian noise to phoneme vectors to synthesize new “magic” words, and resizing inferred phonetic vectors to compress or expand perceived sound; produces plausible spellings, 1950s pulp variants, and nonsense poetry. Code is slated for open source release. sequence-to-sequence models CMU Pronouncing Dictionary Jupyter notebook demo
AI-written summary · about
chapters (5)
- 0:00Start
- 0:48Lewis Carroll's Jabberwocky
- 1:23Sound Symbolism
- 6:15Changing Spelling To Reflect Phonetics
- 8:31Neural Networks to the Rescue
related talks
03h 22mNonsense verse with Python and machine learning
11mMutating text toward better nonsense?!
37mSpeech to text with Elasticsearch
39mExperimental Creative Writing with the Vectorized Word
34mLearning about AI/ML for Text, with Wordle!
01h 30mIntroduction to Deep Learning & Natural Language Processing
31m“Sorry, Could you repeat that again?” - Speech Recognition with Python
01h 24mIntroduction to Language Modeling