# Random Word Generator

> A random word generator for invented words that are actually pronounceable.
> https://random-word-generator.skillsafe.ai/

Random Word Generator coins words that do not exist and sound like they should — for naming
products, characters, places, or for building a constructed language. It runs
entirely in the browser. There is no AI model, nothing is metered, no account is
needed, and nothing a visitor types leaves the tab.

## What makes it different from a random letter generator

Where the randomness happens. Sampling letters gives you "qxlpht". Random Word Generator
samples *syllables* out of an inventory a particular language would actually
use, under that language's own constraints — which is why the output can be read
aloud on sight.

A "language flavour" here is a phonotactic grammar, not a letter-frequency
table. Each one carries:

- separate onset inventories for word-initial and word-medial position, because
  most languages tolerate clusters in one and not the other
- a vowel inventory including the digraphs that flavour's orthography uses
- separate coda inventories for word-medial and word-final position
- probabilities for open versus closed syllables and for onsetless syllables
- a hard constraint on which onsets may meet which nuclei
- a ceiling on how many consonants may meet at a syllable boundary, counted in
  consonant sounds rather than letters, so "sh" costs one and "str" costs three
- a curated set of word-final elements, each tagged so a vowel-initial ending
  only lands after a closed syllable and a consonant-initial one only after an
  open syllable
- a stress rule, used to mark the pronunciation hint

## The nine flavours

| id | name | what is distinctive about it |
|---|---|---|
| `neobrand` | Neo-brand | The product-naming register: short, a distinctive consonant early, category-signalling endings (-ly, -io, -ex, -ora). |
| `latinate` | Latinate | Open syllables, stop-plus-liquid onsets, case endings (-us, -ium, -ora) doing most of the work. |
| `germanic` | Germanic | Heavy onsets (st-, schr-, str-), heavy codas (-cht, -rst, -ndt), first-syllable stress. |
| `slavic` | Slavic | Initial clusters no other flavour allows (vl-, zm-, mst-, zdr-), y as a full vowel, -ov / -ka / -itsa endings. |
| `nihongo` | Japanese-ish | Strictly (C)V, moraic n the only coda, geminates the only cluster, and the real gaps in the mora chart enforced. |
| `oceanic` | Polynesian-ish | No closed syllables at all; vowel sequences do the work clusters do elsewhere. |
| `semitic` | Arabic-ish | Root-and-pattern: three root consonants poured into a vocalic template. |
| `steppe` | Turkic-ish | Vowel harmony across the whole word, including the suffix. |
| `elvish` | High-fantasy | Sonorant-heavy, scarce voiced stops, vowel-initial endings (-iel, -ion, -riel). |

Two of these differ at the level of *rules* rather than inventory, which is the
real test of whether a flavour control means anything:

- **Turkic-ish enforces vowel harmony.** The first vowel picks back (a, o, u, y)
  or front (e, i) and every vowel after it has to agree — including the one in
  the suffix. That is why -lar and -ler are one suffix and not two, and the app
  selects the agreeing form automatically.
- **Arabic-ish does not assemble syllables at all.** It picks three root
  consonants and pours them into a vocalic pattern, the way Semitic words are
  genuinely built, so two coinages from one root sound like relatives. A
  consequence worth knowing: its syllable count is decided by the pattern rather
  than requested, so it advertises the range it can produce (2 to 4) and says so
  when a request is clamped into it.

## Controls

- **Language flavour** — one of the nine above.
- **Syllables** — a minimum and a maximum, 1 to 6, counted over the *finished*
  word including its ending.
- **Familiar → exotic**, 0 to 100 — an exponent over each segment's tier, so at
  the low end only the everyday material of the flavour appears and at the high
  end the rare material is on level terms. The same dial scales the
  closed-syllable probabilities, multiplicatively, so a flavour with no codas
  still has none at the top of the range.
- **Starts with / ends with** — enforced while the word is built rather than by
  generating and discarding, so a two-sided constraint still returns a full
  batch.
- **How many** — 4 to 60.
- **Seed** — the same seed gives the same batch. It travels in the URL fragment,
  so a batch of coinages is a link.

## Collision checks

Every candidate is checked before it is shown. A hit is discarded and resampled,
never displayed; the counts are reported, and for dictionary collisions the word
that collided is reported too.

- **Offensive terms** — always on, cannot be switched off. A curated list of 163
  stems covering English and several widely-spoken European languages.
- **Existing English words** — 126,188 entries from SCOWL levels 10 to 70. On by
  default. A coinage that is already a word is not a coinage.
- **Existing proper names and places** — 14,404 entries. On by default.
- **Well-known trademarks** — 157 marks famous enough to recognise. On by
  default, switchable.

Stems of four letters or more match anywhere in a word; stems of three or fewer
match only at a word edge, which is what stops the filter rejecting "assam" and
"Scunthorpe".

**These are generated strings, not a clearance.** The offensive list does not
cover most of the world's languages. The trademark list is not a trademark
search — a real one covers millions of registrations across dozens of classes
and jurisdictions. A clean result means no obvious collision was found. It never
means a name is safe to use, and this app is not a substitute for a search of
the register or for advice from someone qualified to read one.

## What each result carries

- the coined word
- a pronunciation hint with the stress marked by the flavour's own rule, derived
  from the syllable boundaries the generator actually used rather than guessed
  from the spelling
- the syllable template it was built from, in CV notation
- the syllable-by-syllable breakdown (onset, nucleus, coda)
- for the Arabic-ish flavour, the three root consonants and the pattern
- for the Turkic-ish flavour, the harmony class
- any real English words hiding inside the coinage — reported, not filtered,
  because that is something a person naming a product wants to notice

## Data and provenance

- Word lists: SCOWL 2020.12.07 by Kevin Atkinson, levels 10–70 only. Licence at
  /data/LICENCE-SCOWL.txt, provenance and per-file SHA-256 at /data/NOTICE.txt.
  Levels above 70 are deliberately excluded for licensing reasons, so this app
  ships nothing derived from the UK Advanced Cryptics Dictionary or from ENABLE.
- Offensive-stem list (/data/avoid-stems.txt) and trademark list
  (/data/brand-stems.txt) are original to this app.

## Cost and account

Completely free. No model, no metering, no credits, no sign-in required. Signing
in is optional and does exactly one thing: it lets kept words live on your
SkillSafe account so they follow you to another browser and can be searched by
meaning.

## Pages

- / — the generator
- /api.html — the data API for kept words, in eight languages
- /tokens.html — session and token panel (noindex)
- /data/NOTICE.txt — sources, licences and checksums
- /data/LICENCE-SCOWL.txt — the word-list licence