I initially built this as a toy and it turned into an interesting small phonetics project.
Enter a sentence and it finds other ways to break down the same sounds into English words: “kiss the sky” gives you “kiss this guy”. “Four candles” turns to “fork handles”. “Ice bank mice elf” ... I think you get the idea
Each word is looked up in the CMU Pronouncing Dictionary. The sentence is converted into a sequence of phonemes, and a beam search breaks it back down into words. Exact re-parsing is rare, so sounds may be distorted under certain conditions. For example, keep in mind that a double consonant at a word boundary sounds like a single one (mistake/miss steak). A /ɡ/ before a nasal almost completely disappears (recognize -> “reco'nize”). The sound alone is not enough to distinguish “wreck a nice beach” from “reckon eyes beach,” so variations that actually appear in 40 million lines of movie subtitles are also taken into account.
The system was calibrated using 42 known oronyms (38 in the top 12, 25 in first place), and I disabled each rule individually to see how much it contributed.
What doesn’t work: long sentences, proper nouns (“Euthanasia” -> “Youth in Asia” is impossible, since names are filtered out), and it only works with American English, since CMUdict only contains that.
rickintoplace•1h ago
Enter a sentence and it finds other ways to break down the same sounds into English words: “kiss the sky” gives you “kiss this guy”. “Four candles” turns to “fork handles”. “Ice bank mice elf” ... I think you get the idea
Each word is looked up in the CMU Pronouncing Dictionary. The sentence is converted into a sequence of phonemes, and a beam search breaks it back down into words. Exact re-parsing is rare, so sounds may be distorted under certain conditions. For example, keep in mind that a double consonant at a word boundary sounds like a single one (mistake/miss steak). A /ɡ/ before a nasal almost completely disappears (recognize -> “reco'nize”). The sound alone is not enough to distinguish “wreck a nice beach” from “reckon eyes beach,” so variations that actually appear in 40 million lines of movie subtitles are also taken into account.
The system was calibrated using 42 known oronyms (38 in the top 12, 25 in first place), and I disabled each rule individually to see how much it contributed.
What doesn’t work: long sentences, proper nouns (“Euthanasia” -> “Youth in Asia” is impossible, since names are filtered out), and it only works with American English, since CMUdict only contains that.
Static site, no server, no LLM. Code: https://github.com/rickintoplace/wordlab