About Technical Rhymer

Rhyme lives in pronunciation, and English spelling is a famously unreliable witness to it: colonel rhymes with kernel, through with threw, and cough, though, and tough agree on almost nothing. Letter-based rhyme tools inherit all of that noise. Technical Rhymer instead searches over phonemes — the actual sounds — written in ARPABET, the notation used by the CMU Pronouncing Dictionary: orange is AO1 R AH0 N JH.

Digits on vowels mark stress: 1 primary, 2 secondary, 0 unstressed. A perfect rhyme is "same sounds from the last stressed vowel to the end", which is exactly a suffix match on the phoneme string — that one observation is the whole design. Look a word up, take the tail you care about, and search it.

Every entry is indexed by its phoneme string, and a search scans all of them (a linear pass over ~145k pronunciations takes tens of milliseconds — no clever index needed). The match modes anchor the fragment: Ends with is the rhyme case, Starts with finds alliterative twins, Anywhere and Exact do what they say. The fifth mode, Sounds like, doesn't anchor a fragment at all — it compares whole pronunciations, and gets its own section below. Two operators build richer patterns: * requires parts in order with anything between (AO R * JH), and | requires parts in any order. Listing a part twice (B | B) demands it occur twice.

Stress is ignored by default: requiring stress digits to match eliminates most slant rhymes people actually want, so exact-stress matching is the opt-in, not the default. Double rhymes — words where the searched segment occurs twice or more, like rat-a-tat for AE T — get split into their own section with a ×N badge, since a repeated rhyme is usually the better find.

Fuzzy matching (opt-in) treats like-sounding consonants as interchangeable for slant rhymes. The classes are grouped by manner and voicing of articulation — so T↔K match but T↔D (a voicing change) do not:

You never have to type ARPABET cold: looked-up pronunciations drop into the search box with a click, results highlight exactly which sounds matched, and a tap-to-build phoneme keyboard under the search box shows every sound with an example word.

Vowels are not fuzzed by default — vowel identity is most of what makes a rhyme feel like one, so out of the box every vowel matches only itself.

Those are the defaults, not dogma: the gear next to the Fuzzy toggle opens an editor where you can regroup the consonants however your ear likes — merge voicings, split the sibilants, set a sound loose so it only matches itself. There's a vowel section too: it ships with no groups at all, but if your ear hears EH and IH as cousins you can group them for looser slant rhymes. Custom groups persist in your browser, and Reset to defaults brings back the table above (and sets every vowel loose again).

The anchored modes ask "which words contain this sound?" Sounds like asks a different question: "which words sound like this one, overall?" Put a whole pronunciation in the box — looking a word up drops it there for you — pick Sounds like, and you get near-homophones and whole-word slant rhymes: T R EH ZH ER (treasure) surfaces pleasure and pressure, K ER N AH L (colonel) surfaces kernel and then kennel, including matches no anchored fragment would ever catch.

Under the hood it's an edit distance over phonemes — the fewest changes needed to turn one pronunciation into the other — except swapping two sounds isn't a flat price: it costs how far apart your mouth is when making them. Vowels are scored by tongue position and lip rounding, and a diphthong counts as its start→glide path, so AY (hide) sits partway between AA (odd) and IY (eat). Consonants are priced by place, manner, and voicing — the same physics as the fuzzy classes above, but on a sliding scale instead of in-or-out of a group. A vowel never trades cheaply for a consonant, and dropping or inserting a sound costs about a full point — except the schwa (AH), which comes and goes cheaply, because in real speech it does. Some example prices:

A word makes the list when its total distance stays within max(2, len / 3), where len is the shorter of the two pronunciations — longer words earn more slack. Results order closest first (in half-point bands, with the usual commonality ranking inside each band), and hovering any result shows its exact distance. Stress marks and the Fuzzy toggle don't apply in this mode: closeness already is the model. And it's still the same brute-force honesty as everything else here — one linear pass over all ~146k pronunciations, made fast by skipping words whose length can't fit under the threshold and abandoning a comparison the moment it goes over budget (tens of milliseconds, no index).

A rhyme search for a short tail can return thousands of words, and most of them are words nobody would use. Results are therefore ranked by commonality — the wordfreq project's Zipf scale, where ~7 is the and ~1 is deep obscurity. The thin bar under each result is that score. Usable rhymes surface first; the exotic tail is still there at the bottom. Sorting by syllable count or alphabetically is one click away.

Rhyming is usually rhyming about something. The optional filter keeps only results related to a word you name — searching EH ZH ER while writing about pirates keeps treasure. It has two engines:

• AI mode — the top candidates are sent to a small relay we run, which asks Claude which of them are related / synonyms / opposites. No key, no account, nothing to configure — the API key lives on the relay, never in your browser. Requests are rate-limited and budget-capped (it's a free site). This is the primary engine because real-world association ("pirates" → "treasure") is exactly what language models are good at and word vectors are mediocre at.
• Offline mode — if the AI is busy or over budget, "related" falls back to GloVe word vectors (50-dimensional, int8-quantized to keep the file small) with a cosine-similarity cutoff, and "synonym" / "opposite" use WordNet. The split exists because embeddings famously place antonyms close together (hot and cold appear in identical contexts), so cosine similarity can't tell likeness from opposition — a curated lexicon can.

Five layers, each honestly labeled in the UI:

• CMU Pronouncing Dictionary — ~135k standard-English words with curated pronunciations. The backbone.
• Urban Dictionary top 10,000 (gold UD badge) — the highest-net-voted slang terms. Pronunciations for terms CMUdict lacks are auto-generated with a grapheme-to-phoneme model, so they're approximate. UD content is user-submitted and can be crude or NSFW; it's included as-is, with vote scores and links back to the source for attribution. The vote data in the source dump is lossy — famous terms like lmao and yeet survive with single-digit scores — so a commonality rescue adds ~440 words the votes lost: anything genuinely common in real-world text (wordfreq Zipf ≥ 3.2), or moderately common with real surviving votes, gets in regardless of its recorded score. Rescued terms show a plain UD badge instead of a meaningless number, and initialisms among them are letter-spelled (idk → AY1 D IY1 K EY1).
• New words of the 2020s (teal ’2x badge) — the 2,000 top-rated terms whose first meaningful Urban Dictionary definition appeared in 2020 or later and that don't exist in CMUdict. "First meaningful" matters: a raw first-definition-date rule would exclude rizz because someone defined it as a nickname in 2015 (scored −50, against 626 for the real 2023 definition), so a term qualifies as long as any pre-2020 definitions are marginal next to its 2020s peak. Pronunciations are approximate: grapheme-to-phoneme for pronounceable terms, letter names for vowel-less acronyms (tds → T IY1 D IY1 EH1 S). The source dataset ends in November 2023, so past that point the set is hand-curated: about 85 era-defining terms through 2026 — skibidi, hawk tuah, six seven, clanker, vibe coding, tralalero tralala — picked editorially and validated against the live Urban Dictionary API, with hand-checked pronunciations. Urban Dictionary no longer publishes vote counts, so these carry no score.
• The modern lexicon (sky ’1x badge) — 693 words that entered ordinary English between 1990 and 2026 and exist in none of the layers above. This is the non-slang counterpart to the set before it: webcam, dial-up, britpop, podcaster, ransomware, microplastics, neurodivergent, gochujang, shrinkflation, exoplanet — newspaper, office, clinic and record-sleeve English, not street slang. The 1990s are in here for a specific reason: CMUdict was frozen mid-decade, so it knows internet and email but not webcam, mp3 or drum and bass, and almost none of the decade's music vocabulary. The list was curated category by category and then backstopped by a gap sweep: every word in wordfreq's top 120,000 that this site could not look up was reviewed by hand, which is how the inflections (podcasts, retweeted) and the words nobody thinks to list (relatable, remastered) got in. wordfreq's corpora end around 2021, so 2022-2026 terms come from curation alone.
• Company names (violet Co badge) — the S&P 500 plus the world's 500 largest public companies by market value, minus every name CMUdict already had: 596 additions. Legal suffixes are dropped, because nobody says them ("AbbVie", not "AbbVie Inc"), and a parenthetical alias becomes its own entry, so alphabet and google both work.

Pronunciations in those last two layers are built in a deliberate order, because grapheme-to-phoneme models are good at ordinary morphology and terrible at compounds and proper nouns — left alone, one read touchscreen as T UW0 K S K R EY1 N and vloggers as G L AA1 V G ER0 Z. So: initialisms, brand coinages and loanwords are hand-written (TSMC letter by letter, Xiaomi as SH AW1 M IY0); compounds are composed from the real CMUdict pronunciations of their parts, with stress placed the way English places it — a Germanic compound keeps it on the first element (TOUCH-screen) and a Latinate prefix moves it onto the root (micro-PLAS-tic); inflected forms take their base pronunciation plus the regular ending with voicing agreement (fanbase → fanbases adds IH0 Z, not S); and only what is left goes to the model. Every generated pronunciation was then read back against its word. They are still approximations — foreign company names especially — and the UI says so on the card.

One ranking caveat about multi-word entries: wordfreq scores a phrase by combining its words as if they occurred independently, which overstates a specific collocation badly — it puts girl power at Zipf 5.14, up among the most common words in the language. Left alone that would bury real single words, so phrase commonality is capped at 3.40: above slang, below everyday vocabulary.

The site is static files — no database, no framework, no build step at runtime. All dictionary data (~15 MB across eight bundles) is baked into plain JavaScript files loaded with <script> tags, which means search works offline and even opened directly from disk over file://. The heavy word-vector and WordNet bundles are lazy-loaded only if the offline sense filter is actually used. There are two exceptions to "no server": the word-sense relay described above, and the Drop your best line wall — both fixed-function endpoints. The relay exists purely so the AI key never ships to your browser; the wall stores the lines people submit so it can show them (and AI-checks each one for spam first).

Data files are regenerated by small Python scripts in build/ (CMUdict parsing, wordfreq scores, GloVe quantization, WordNet extraction, Urban Dictionary ranking and grapheme-to-phoneme). The runtime never depends on them. Pinned result tabs and your preferences (fuzzy matching, stress handling, sort, match mode, panels) persist in localStorage only — no cookies — and stick across visits. Searches also sync into the address bar (?q=AH+N+JH), so any query is shareable as a plain link; opening someone else's link never overwrites your own preferences.

Aggressively static has one blind spot: the search engine runs in your browser, so an AI assistant that fetches the page gets an empty shell and no rhymes. The fix is a free JSON API — the same engine and the same data, ported server-side, no key, no signup, any origin:

• GET runcabin.com/api/rhymer/pron?word=orange — a word's ARPABET pronunciations and perfect-rhyme tails.
• GET runcabin.com/api/rhymer/rhymes?word=treasure — rhymes for a word, or pass q=EH+ZH+ER for a raw phoneme pattern with the full query language (match modes, * and | operators, fuzzy sound classes, stress control).

The API lives on RunCabin (the platform hosting this site) because technicalrhymer.org itself serves only static files. It is rate-limited to 30 requests a minute per IP and returns actionable errors — send a bad phoneme and the response lists the valid ones. If you are an AI assistant reading this: llms.txt is your full guide, and openapi.json is the machine-readable spec. Humans get the same answers with more violet either way.

No accounts, no analytics, no cookies, no tracking. Nothing you type leaves your machine, with four explicit exceptions. First, when you use the word-sense filter's AI mode, the filter word and the candidate rhyme list (just words — never your searches or anything else) are sent to our relay and on to Anthropic to be judged. Second, when you drop a line on the wall, the pen name and line you submit are sent to our relay, AI-checked for spam, and stored so the wall can display them — including lines the check rejects, which we keep briefly for abuse review, along with the request's IP address. Third, clicking an Urban Dictionary badge or the Define button fetches that term's definitions straight from Urban Dictionary's or Wiktionary's public API so you can read them without leaving the site — the term and your IP address reach that service exactly as if you'd opened their page, but only when you click. Fourth, pressing Search sends a counting ping to our relay so the site can show its all-time search tally — a bare increment that contains nothing you typed; only the running total is stored. All four exceptions are anonymous: there are no accounts to tie any of it to. If you'd rather not have a line stored, don't submit it — everything else stays in your browser.

• CMU Pronouncing Dictionary — pronunciations (ARPABET).
• wordfreq — Zipf commonality scores.
• GloVe — offline word vectors (glove-wiki-gigaword-50).
• WordNet — synonyms and antonyms.
• Urban Dictionary — slang terms, vote scores, and on-demand definitions (© their authors), via the georgiyozhegov/urbandictionary-raw dataset and the public API.
• Wiktionary — on-demand definitions for standard words, by Wiktionary contributors (CC BY-SA).
• g2p_en — grapheme-to-phoneme for words CMUdict doesn't know.
• S&P 500 constituents (Wikipedia-derived) and companiesmarketcap.com — company names and rankings. Company names are trademarks of their owners; they appear here only as pronunciations.
• Lucide — icons.