Accuracy & Methodology

Every word passes through up to three layers, in order:

1. Dictionary lookup. The word is checked against a pronunciation dictionary derived from the CMU Pronouncing Dictionary (~126,000 entries), which records real, attested pronunciations rather than guessing from spelling. Words with more than one accepted pronunciation (like "fire" or "every") keep every valid count.
2. Curated overrides. A small, hand-reviewed table corrects or extends specific entries — for example, adding the common fast-speech pronunciation of "chocolate" that the base dictionary doesn't capture. This table grows from user-submitted corrections after editorial review.
3. Rule-based fallback. Words in neither the dictionary nor the override table (proper nouns, coinages, rare compounds) are estimated using vowel-group counting with adjustments for silent final "e," "-ed" and "-es" endings, "-le" endings, hyphenation, and common prefixes.

Numbers and acronyms get dedicated handling: numbers are estimated by how they'd be read aloud ("1998" → "nineteen ninety-eight"), and all-caps acronyms are either matched against a short list of ones commonly said as a word (like "NASA") or spelled out letter by letter.

Every result carries a confidence label — high, medium, low, multiple pronunciations, or manually corrected — so results are never presented as more certain than they are.

Structure and syllables are checked deterministically. Literary analysis — seasonal words, imagery, juxtaposition, and the haiku-vs-senryu estimate — works differently:

• Seasonal words (kigo) are matched against a curated reference list tagged by season and confidence. The list is a starting point, not exhaustive — a missing match doesn't mean a poem has no seasonal quality, just that it isn't in the current list.
• Sensory imagery is detected via word lists across sight, sound, smell, touch, taste, movement, and temperature.
• Abstract vs. concrete language flags words and phrases (like "beautiful," "I feel," "very") that tend to state rather than show.
• Cuts (juxtaposition) are detected primarily through end-of-line punctuation (a dash, comma, colon, semicolon, or ellipsis), with a softer fallback that checks whether the first line shares no words with the rest of the poem.
• Haiku vs. senryu is estimated by weighing nature/seasonal word matches against everyday human/social word matches, always reported with a confidence percentage rather than a flat verdict.

None of this is AI-generated — every check is a documented word list or structural rule, which means the same input always produces the same output.

The Haiku Score is a 100-point rubric split across seven categories: structure (25), syllable accuracy (20), imagery (15), seasonal context (10), juxtaposition (10), brevity and clarity (10), and natural language (10). Each category is computed from a documented formula — never a random number — and Modern Haiku mode specifically avoids penalizing a poem for not matching 5-7-5, scoring syllable accuracy against a typical brevity range instead.

The Haiku Rater uses a separate, more literary rubric (originality, sound, ending strength, and more) built the same way — documented rules, not a black box.

• Rare proper nouns, brand names, and invented words rely on an estimate, not a dictionary lookup.
• The pronunciation dictionary reflects North American English; other dialects may pronounce some words differently.
• The seasonal word list reflects common English-language haiku convention and isn't exhaustive or globally universal — seasons and their associations vary by region.
• Literary analysis (imagery, juxtaposition, haiku-vs-senryu) is a structured estimate, not a substitute for human editorial judgment — poetry interpretation is inherently subjective.
• The cut/juxtaposition detector relies heavily on punctuation; a cut created purely through subtle phrasing without any punctuation may be missed.

Anyone can report a syllable count they believe is wrong, including a suggested count, pronunciation, and reason. Reports are queued for editorial review before being added to the override table — nothing is added to the live dictionary automatically. See Contact to submit a correction.

• August 4, 2026 — Initial launch: syllable engine, haiku analysis engine, Haiku Checker, Syllable Counter, Haiku Generator, and Haiku Rater.