Update: [28 Nov 2016] I had written this blog post quickly, without really explaining the context; see my notes for a better explanation.

People have used sequence-to-sequence recurrent neural networks to "translate" words into pronunciation for speech synthesis. I have been trying to go the other direction. I was inspired by the name "Daneel Olivaw" in Asimov's stories. It's similar to "Daniel Oliver" but it's a little different. The idea is to take English names like Susannah, Michael, etc., convert them to pronunciation phonemes, alter those phonemes in some way (such as the Great Vowel Shift of Middle English), and then convert the altered pronunciation into a new spelling. Then someone could use these new names for a story or game.

I think of it as a "Spelling Bee" neural network. It hears the name and has to come up with a spelling for it. Unlike a regular spelling bee, this is for made-up words. These are results I've gotten so far:

  • Changing the N in Jennifer → Jemifer Gengnifer Gethopher Jeffepher Jessifer Geshifer Gethopher Jeviffer Jesapher Jesifer
  • Changing the C in Christopher → Bristougher Dristofer Gristopher Prestofer Tristofer Threstougher Fristopher Srystofer Shrystofer Thristopher Vristofer Zristopher Ghrystofer
  • Changing the first E in Stephanie → Stophony Staphony Stophanie Stophony Stophony Styphony Sterfaney Staphony Stiffony Stephony Stophani Steuphony Stuphony Stuphony
  • Changing the IE in Daniel → Danall Danall Dannell Dannall Danhowl Danile Danelle Dannerl Danail Dannyll Danielle Danole Danoyle Danule Daneule

I'm also going to try prefixes and suffixes. It's been a fun mini project. I got to learn some TensorFlow and recurrent neural networks, even though for the most part I'm just patching together code I've found without really understanding it. The results so far seem like plausible spellings for words, but most of them aren't sufficiently name-like.

Something that I hadn't thought about before: there can be lots of different spellings for the same sounds in English (what a messed up language!). For example michael → M AY K AH L but the AY K AH reverse maps to ichae in michael, icu in bicuspid, ica in formica, iche in lichen, yca in lycan, yco in glycogen, yche in psychedelic, yc in recycle, so which of these "should" it be using when spelling that sound? How would the neural network be able to learn something if there isn't a good answer?

See my notes for a longer explanation of what I was trying to do.

Labels: ,

1 comment:

Jozef K. wrote at November 24, 2016 2:39 PM

I think this might interest you regarding to the article's subject

https://fatiherikli.github.io/language-evolution-simulation/