PhonoPair scores are deterministic and explainable — not AI-generated
Every factor has a named source. Every score can be reproduced. This page explains exactly how the three-pillar scoring system works and the peer-reviewed research behind it.PhonoPair does not use GPT, Claude, or any neural network to generate or score names. The scoring engine is a deterministic rule system built on top of structured linguistic databases. The same input will always produce the same output — and the output comes with a full breakdown explaining why.
This matters in professional naming contexts: you can show a client, investor, or legal team exactly why a name scored the way it did. There is no hallucination, no unexplainable confidence score, and no model drift over time.
Every word combination is scored across three independent pillars. The pillars are then blended into a final 0–100 score.
Alliteration, assonance, consonance, vowel shape compatibility, rhythm, syllable balance, and articulatory transition ease — the raw sound quality of a name pair.
Word type compatibility (adjective-noun, noun-noun), stress pattern harmony (using each word's real dictionary stress where the two words meet), syllable length preference, grammatical structure, word order naturalness, and visual legibility (letter silhouette, handwriting confusability, industry typographic fit).
Meaning tension, cultural metonymy, conceptual contrast, oxymorons, onomatopoeia, and resonance — how the two words interact at a deeper level of meaning.
For a two-word combination, the final displayed score blends a per-word quality average (40%) with a pairwise compatibility score (60%). The compatibility score starts as the sum of all three pillar scores. Up to 88 it is used as is; above 88 a soft ceiling compresses it, so it approaches 100 but never reaches it. Strong pairs therefore no longer all tie at 100 — a pair that is genuinely better still scores higher.
For three or more words, compatibility is the average across each neighbouring pair. Every pair — two words, longer phrases and the Name Finder — goes through the same scoring code.
Every phonetic score starts from how each word is actually pronounced. PhonoPair checks sources in a fixed order and uses the first match:
Accents and curly apostrophes are normalised first (Zoë, O’Brien), and hyphenated or multi-word names (Mary-Kate, de la Cruz) are pronounced part by part. The dictionaries also give each word's stress pattern — which syllable is strong. Stress feeds the Language pillar (how the stress falls where the two words meet) and each word's quality score (regular rhythms with early stress read and remember more easily). When a word's stress is unknown, first-syllable stress is assumed, as is typical in English.
Names with a middle name: compatibility now counts the full name as it is usually said (first + last) for half, and the pairs side by side (first + middle, middle + last) for the other half, so a three-part name is no longer judged only on its middle joins. The surname always stays last when we suggest a better order.
The semantic pillar works again. Its knowledge graph (ConceptNet) had been unreachable, so meaning-based factors such as opposites and complementary pairs scored nothing. PhonoPair now hosts that data itself. Most scores move by a point or two; strongly related pairs (opposites like hot and cold, complements like salt and pepper) gain more.
Where the names meet now counts properly. A soft beat handing off to a stressed one (LI-ly | SMITH) adds points; two stresses colliding (JAMES | BLACK) or names running together (Jack Kennedy, Ava Adams) cost points — and those penalties count in full, even for names that otherwise score very high. Rhymes and echoes (Ray Gray, Eve Keaveney) still raise the sound match, and the Name Finder tags them so you can hide them. Every Name Finder score now has a “Why this score?” breakdown, and rhythms can be played aloud.
Every result now records the scoring model it was calculated with. Shared links and saved analyses from an earlier model are labelled, and shared links are re-scored with the current model where possible.
We corrected how pronunciations are looked up. Previously most words fell through to a dictionary converter that garbled them (“apple” came out with stray symbols), so many phonetic scores were based on the wrong sounds. Words now come from the full CMU dictionary and the reviewed name table first, and real stress patterns are now part of the score.
We also replaced the hard cap at 100 with the soft ceiling described above, so top-end scores are no longer tied.
Scores are still deterministic — the same input gives the same output under the current rules — but results you saved or shared before this update may differ if you run them again. Re-run a name to see its current score.
These are structured databases and linguistic knowledge graphs — not language models. All sources are publicly documented and have been used in academic linguistics research.
Carnegie Mellon's machine-readable phonetic database mapping ~134,000 words to their ARPAbet phoneme sequences and stress patterns. PhonoPair uses the full dictionary (BSD-style licence).
Used for: Primary pronunciation and stress source — phonemes, syllables, stress, alliteration, rhymePhonoPair's own static table for US first names: about 2,600 names CMU lacks (modern spellings like Kayden or Nevaeh), transcribed with language-model assistance, validated as ARPAbet with one primary stress, plus about 330 hand-audited corrections where CMU's first reading is wrong for the name. Prepared once, offline — no model runs when you score a name.
Used for: Pronunciation and stress for first names, checked before the full CMU dictionaryPublic-domain yearly top-1,000 first-name lists for the United States, 1900–2024, obtained via the aruljohn/popular-baby-names mirror (MIT licence), pinned to a fixed version.
Used for: Name Finder — the 6,671-name list, popularity, style and gender usageA large-scale semantic network with over 34 million statements about how words relate — antonyms, associations, "is-a" relationships, and more. PhonoPair hosts its own English subset of ConceptNet 5.7 (about 670,000 terms and 2.5 million relationships, © Commonsense Computing Initiative, CC BY-SA 4.0), so these lookups no longer depend on the public API, which has been unavailable.
Used for: Semantic factors — metonymy, oxymoron, complementarity, cultural resonanceA linguistics research API returning words by sound, spelling, meaning, and context — rhymes, collocations, sounds-like, and alliterations.
Used for: Phonetic matching, collocations, related wordsEncyclopedic cultural knowledge used to assess whether a word or pair has real-world cultural significance beyond its literal meaning (page-existence lookups only).
Used for: Cultural resonance and metonymy scoring; per-name signal in the Name FinderAuthoritative American English dictionary providing definitions, part-of-speech tags, and etymology.
Used for: Language pillar — POS detection, grammar checks; fallback pronunciation (accepted only when it converts to valid phonemes)The factors we score are not invented. Each one maps to peer-reviewed findings in psycholinguistics, cognitive psychology, and branding science.
Subjects shown alliterative word pairs recalled them significantly faster and more accurately. Effect held whether text was read silently or aloud.
Psychological Science, 2008 ↗Names whose sounds matched product attributes (sharp sounds for precise/fast products, round sounds for soft/warm products) scored higher on preference and purchase intent.
Journal of Consumer Research — Yorkston ↗Vowel position in the mouth (front vs. back) consistently predicts consumer perception of product size, speed, and personality — across English, Chinese, and Navajo speakers.
Multiple studies including ResearchGate / Journal of Global Marketing, 2023 ↗Rounded/soft sounds are perceived as friendly and approachable; angular/sharp sounds as precise and efficient. Brand phonetics that match visual identity score higher on coherence.
Bouba/kiki effect — 99designs, Wikipedia, multiple replication studies ↗Beautiful-sounding words tend to have 3+ syllables with first-syllable stress and liquid consonants (l, r, m, n). These preferences are implemented directly in PhonoPair's rhythm and consonant scoring.
David Crystal, "Phonaesthetically Speaking" (1995)Run the same name through PhonoPair tomorrow and get the same score. No model updates quietly changing your baselines.
Show a client, investor, or legal team exactly why a name scored the way it did — factor by factor, with source data.
Scores are directly comparable across names. A 78 means the same thing regardless of which names you ran before or after it.
The Name Finder applies the same scoring to first names: give it a last name and it ranks thousands of first names by how well each one flows with it.
The list is 6,671 first names drawn from the US Social Security Administration's yearly top-1,000 lists, 1900–2024. Every name has a dictionary or reviewed pronunciation — none are guessed from spelling. Each first name is scored against your last name with the analyzer's pair scoring, using the personal-name weighting profile.
To rank every name in seconds, the Name Finder scores offline: it skips the live per-pair knowledge-graph and Wikipedia lookups and uses each name's stored Wikipedia signal instead. For most pairs the result matches the analyzer exactly; for famous full names, where those live signals fire, it can differ by a few points.
Sound-alikes are found by sound distance: how many phoneme edits turn one name into another, with similar sounds (two vowels, two flowing consonants) counting as smaller changes and extra weight on the first and last sounds, which the ear notices most. Spellings that sound identical are grouped under the most popular one.
Style comes from popularity history (e.g. popular = top 100 in 2024; classic = in the top 1,000 almost every year since 1900; modern = first charted in 1990 or later). A name counts as a girl's or boy's name when it ranks at least twice as high for that gender, otherwise it is shown as unisex. Rhythm, ending and feel filters come from the name's pronunciation.
The scoring system is not static, but the rules it follows are. Every addition stays deterministic, sourced, and explainable — and some land as analysis insight before they are ever allowed to move the score, because a factor should earn its weight against known names before it counts.
Each word is broken into syllables and scored on six documented sound-symbolic dimensions — sharpness, brightness, weight, energy, openness, tension — then read as a contour across the whole name. It ships first as analysis insight, explaining why a name feels the way it does without changing the number. It also consolidates the existing sound-shape signal into a single source of truth rather than running a second heuristic alongside it. It will only affect the score once validated against known names.
A new component of the Language pillar scores the written silhouette of a name — the balance of ascenders, descenders, and x-height letters that determines how quickly it is recognised on a page. It is font-agnostic, deterministic, and intentionally small: a tiebreaker that fine-tunes the score, never one that dominates it, always shown with the reasoning.
The scoring system is documented — but the fastest way to understand it is to run an analysis.
Scores are objective. What you do with them takes judgment. Our studio takes on select naming engagements.