Spell Checker

How does it work?

Behind the scenes it relies on the browser native spellcheck, which draws from the dictionaries installed on your OS. Nothing is sent anywhere.

One requirement applies: the language you want checked has to be installed on the system. In Chrome or Edge, head to Settings → Languages → Spell check.

Need something stricter? Dedicated tools such as LanguageTool are worth a look.

Spell Checking: Dictionaries, Edit Distance, and Language Models

A spell checker compares the words in a text against an authoritative lexicon and flags anything missing, optionally proposing replacements. Modern checkers stack several layers: a base dictionary of lemmas, a morphological analyzer that knows "running" is a valid form of "run", an edit-distance engine that ranks candidates, and increasingly a contextual language model that picks between valid alternatives based on the surrounding words. The first interactive checker, Ralph Gorin's SPELL, ran on Stanford's PDP-10 in 1971; UNIX shipped Doug McIlroy's "spell" in 1975. Fifty years later the principles are the same; only the dictionaries and the statistical machinery have grown by orders of magnitude.

This page offers an in-browser checker that delegates to the native HTML spellcheck attribute, using whichever dictionary your operating system or browser provides (Hunspell on Linux, the built-in macOS checker, Windows Spell Check, Chromium's own dictionary). That is a reasonable default but it has limits, especially for Portuguese: native dictionaries vary in quality between platforms and rarely understand grammar or context. The sections below explain how spell checking actually works, why good checkers still miss certain errors, and which tools to reach for when the stakes are higher.

The 1990 Portuguese Orthographic Agreement: Accents, Hyphens, Tréma, Mute Consonants

Any Portuguese checker built after 2009 must implement the Acordo Ortográfico da Língua Portuguesa de 1990, an inter-governmental treaty that unified spelling across the lusophone countries. In Brazil it became mandatory on January 1st, 2016; in Portugal on May 13th, 2015. The reform touched roughly 0.5% of Brazilian vocabulary and 1.6% of European Portuguese, but its impact on dictionaries and editors was significant.

The visible changes: the tréma (two dots over u in freqüência, lingüiça, pingüim) was abolished, leaving frequência, linguiça, pinguim. Differential accents on minimal pairs disappeared, so pára (verb) and para (preposition) are both written para; pêlo (hair) and pelo (preposition) collapse into pelo. Hiatuses preceded by diphthongs in paroxytone words lost the acute accent: jóia became joia, idéia became ideia, assembléia became assembleia. The circumflex on double-o vanished: vôo became voo. The hyphen was systematized: when a prefix ends in a vowel and the second element starts with r or s, the consonant doubles and the hyphen drops (anti-semita to antissemita, contra-regra to contrarregra); but when the prefix and next word share the same vowel the hyphen stays (micro-ondas, anti-inflamatório). In European Portuguese the agreement also eliminated mute consonants Brazilians never pronounced: óptimo to ótimo, acção to ação, eléctrico to elétrico. The alphabet was officially expanded from 23 to 26 letters with the formal admission of k, w and y. A robust checker should accept pre- and post-reform spellings, ideally with a soft warning that the older form is obsolete since 2009.

How Spell Checkers Work: Hunspell and Edit Distance

Almost every modern checker is built around the same triad. A dictionary file lists root words. An affix file describes prefixes and suffixes, so 50,000 lemmas cover millions of surface forms. A candidate generator proposes replacements when a word is missing, ranked by edit distance and frequency. Hunspell, written by László Németh from 2003 on, is the de facto open-source standard, used by LibreOffice, Firefox, Thunderbird, Chrome (for some languages), macOS and Adobe products. It extends MySpell with support for complex morphology (compounding, double affixation, conditional rules). A Portuguese Hunspell dictionary declares 80,000-200,000 roots and covers tens of millions of inflected forms.

When a word is unknown, the checker enumerates dictionary words within a small edit distance. Edit distance, formalized by Vladimir Levenshtein in 1965, is the minimum number of single-character insertions, deletions, or substitutions needed to transform one string into another. The Wagner-Fischer dynamic-programming algorithm computes it in O(m × n) time. Candidates are ranked by distance and, within the same distance, by corpus frequency. Peter Norvig's famous 21-line Python checker (2007) showed that this technique with a probabilistic model trained on a large corpus reaches 80-85% accuracy on common typos. The Damerau-Levenshtein variant adds transposition of adjacent characters as a single edit, capturing one of the most common typing mistakes (teh to the).

Phonetic Matching: Soundex, Metaphone, N-grams

Edit distance fails when a writer misspells a word phonetically. Someone who writes "exceção" as "esceção" produces a string several edits away from the correct form but phonetically identical. Phonetic algorithms reduce both the input and dictionary entries to a normalized sound code, then match codes instead of letters. Soundex, patented in 1918 for US Census name matching, keeps the first letter and encodes the rest as digits based on consonant classes. Metaphone (Lawrence Philips, 1990) improved Soundex by considering letter context and pronunciation rules; Double Metaphone (2000) emits two codes for ambiguous pronunciation. For Portuguese, BuscaBR and Metaphone-PT-BR adapt the rules: they collapse s, ss, ç, sc, xc into one "s" code and normalize nasal endings (am, ão, an).

A pure dictionary checker cannot catch real-word errors: words that are valid in isolation but wrong in context. "I'd like a peace of cake" passes a dictionary check because "peace" is a real word. Detecting this requires a model of which words appear together. The classical solution is the n-gram language model: count how often each bigram, trigram or longer sequence appears in a reference corpus, then score sentences by those counts. If "piece of cake" appears 100,000 times and "peace of cake" once, the checker proposes the swap. Microsoft Word's contextual spelling feature (2007) was the first mainstream product to use this approach. Since 2018, transformer-based neural language models (BERT, GPT) have begun to replace n-grams, scoring any token given the surrounding context and catching rare real-word errors that n-grams miss.

Limitations: Homophones, Grammar, Style

No checker, however sophisticated, catches everything. Portuguese is rich in homophones that confuse even careful writers. Classical pairs include concerto (a musical performance) vs conserto (a repair), mas (but) vs mais (more), há (verb haver) vs a (preposition) vs à (preposition with article), cessão (yielding) vs seção (section) vs sessão (meeting), discriminar (to distinguish) vs descriminar (to decriminalize), ratificar (to confirm) vs retificar (to correct). All are valid words; only context decides. A contextual checker catches the obvious cases ("vou ao concerto da banda") but misses subtle ones in unusual contexts.

Beyond homophones, checkers do not enforce subject-verb agreement, do not detect missing words, and do not flag stylistic issues like redundancy, passive voice or vague pronouns. They over-flag proper names, technical jargon and neologisms unless you maintain a custom dictionary. They struggle with code, URLs and brand names. Treat a checker as the first pass, not the last; a human proofreader still catches what machines miss.

Brazilian vs European Portuguese; Professional Tools

Brazilian (pt-BR) and European Portuguese (pt-PT) share unified spelling after 2009 but diverge in vocabulary and syntax. Brazilians say ônibus, trem, geladeira, banheiro, celular, mouse, sorvete, terno; Europeans say autocarro, comboio, frigorífico, casa de banho, telemóvel, rato, gelado, fato. Clitic pronouns sit before the verb in Brazil, after in Portugal; "estou fazendo" vs "estou a fazer". A checker tuned for one variant flags valid words from the other. Hunspell ships pt_BR and pt_PT separately; LanguageTool, Word, Google Docs and Pages all expose the choice.

For serious writing in Portuguese the four leading checkers in 2026 are: LanguageTool (open source LGPL, free public API, self-hostable; the Portuguese rule set covers over 4,000 patterns including AO90 conformance and false cognates); Grammarly (closed source, dominant for English, Portuguese support added late 2023 but still less mature); Microsoft Editor (ships with Office 365 and Edge, benefits from Microsoft's enormous document corpus); ProWritingAid (targets fiction with style reports and readability scores; weakest Portuguese support of the four). For programmatic use, LanguageTool is the only one with a stable open API.

Frequently Asked Questions

Does this checker work offline?

Yes. The tool uses the HTML spellcheck attribute, which relies on the dictionary already installed on your operating system or browser. No text is sent to a server.

Why does it miss obvious homophone errors?

Dictionary-only checkers verify only that each word exists. Errors like "concerto" instead of "conserto" or "mas" instead of "mais" require a contextual language model — for that, use LanguageTool, Grammarly or Microsoft Editor.

Is the 1990 Orthographic Agreement still in force?

Yes. It became mandatory in Brazil on January 1st, 2016 and in Portugal on May 13th, 2015. Pre-reform spellings (idéia, freqüência, óptimo) are now incorrect for official, academic and journalistic use.

Why does the checker underline brand names and proper nouns?

Native dictionaries do not contain every brand or surname. Right-click an underlined word in most browsers to add it to a personal dictionary, preventing future false positives.