Skip to content
Settings

Editorial policy

This page states where every pronunciation on Learn Pronounce comes from, the standard it is written to, and the things we deliberately refuse to guess. It is written so you can check us, not so you have to trust us.

Who writes this

Learn Pronounce is written and maintained by Muhammad Usama, a software engineer, working alone. There is no editorial board, no contributor network and no anonymous "team" — every guide, every transcription and every line of code has the same author, and every change is recorded in the project's commit history with a date and a name against it. The repository itself is private, so that history is not open for public browsing.

We say this plainly because the alternative — implying a newsroom that does not exist — is the thing we would most object to on somebody else's site. A single named maintainer you can write to is worth more than an unnamed collective you cannot.

Where the phonetic data comes from

The built-in transcription set
1507 headwords — 3185 transcriptions in total: 1448 American, 1447 British, 290 Australian and 0 Indian. A headword carries a form for an accent only where the source gives one, so an accent is never padded out from another. They live in one reviewed data file in the site's repository, so a change to any of them is a dated, attributable edit rather than a silent database update.
Wiktionary, for everything else
The IPA lookup behind /api/ipa queries English Wiktionary and returns the transcription it finds there. The word pages also display glosses sourced from the same Wiktionary REST definition endpoint; these are stored in the site's data files at build time, not fetched live. Both the live transcriptions and the stored definitions are derived from Wiktionary text, which is licensed CC BY-SA 4.0, and every transcription or definition that reaches you through this site carries that licence and that attribution. We do not silently re-badge it as our own.
Additional sources
Pronunciation guides also draw on CMUdict (Carnegie Mellon University, BSD licence) for stress patterns and WordNet (Princeton University, WordNet licence) for word senses. These are secondary references — the IPA transcription itself is always checked against Wiktionary as the primary source.
Verification
Each word's IPA is checked against Wiktionary's primary source entry for that word. Where Wiktionary gives variant transcriptions for an accent, the most common form is used and the note on the word page explains the choice. Corrections are treated as bugs and resolved in the project's commit history.
Stress and syllable marking
The syllable breakdown shown on word pages is derived from the IPA itself: the stressed syllable is the one whose vowel follows the ˈ mark. It is computed from the transcription, never typed in by hand alongside it, so the two cannot drift apart.

The transcription standard

Transcriptions are broad and phonemic, written between slashes in the International Phonetic Alphabet (IPA 2020, International Phonetic Association). That means they record the contrasting sounds of a word in a given accent, not a narrow record of one particular speaker's articulation. A narrow phonetic transcription would be written between square brackets; we do not publish those, because they describe a performance rather than the language.

Within that: ˈ precedes the syllable carrying primary stress and ˌ precedes secondary stress. American forms are written rhotically and British forms non-rhotically, which is why /ˈkɜːrnəl/ and /ˈkɜːnəl/ appear for the same word — they are two accents, not two opinions. The Australian form follows the same principle: it is written as Australian English actually contrasts the sounds, rather than as a decorated copy of the American entry. We hold no sourced Indian transcriptions, so none is published.

Where the audio comes from

Two engines, and the page always tells you which one you are hearing. Device voices are the speech voices already installed in your operating system, reached through the browser's Web Speech API; they are instant, work offline, and nothing about them leaves your machine. Server voices are synthesised by Deepgram Aura-2 (@cf/deepgram/aura-2-en) on Cloudflare Workers AI.

Speed is applied in your browser rather than by re-synthesising, so slow, normal and fast are the same recording played at different rates — not three different renderings that might disagree with each other.

What we refuse to guess

An accent is never substituted for another
The server engine has American, British and Australian voices only. Indian, Canadian and Irish therefore play only when your own device supplies that voice. If it does not, the site says so. It will not quietly hand you a British voice while the label reads Indian — which is exactly what a naive fallback does, and the reason ours only ever falls back within the same accent.
No factual claim is templated across pages
A statement about a specific word — where its stress falls, which letter is silent — is either read from the data for that word or it is not published. It is never filled in from a template, because a template that is right about one word is wrong about a thousand others, and a plausible sentence is not the same thing as a true one.
No invented ratings or credentials
The site publishes no star ratings, review counts, testimonials or qualifications, because it has none to publish. Structured data on this site describes only things a visitor can see on the page.
No single "correct" pronunciation
Many words have several accepted pronunciations. Where they differ we show the variants and let you hear them, rather than promoting one and quietly deleting the rest.

How pages are checked before they ship

Every push to the main branch runs a unit suite over the voice-resolution and preference logic and then builds the whole site; if either fails, nothing is deployed. Alongside that the repository carries a strict TypeScript check over every page and a browser smoke suite that loads real pages, plays real audio and asserts what the interface actually claims — including that a device voice never makes a network request, and that the site never describes a synthesis engine it has stopped using. Those two are run by hand before a release rather than on every commit.

These checks are honest about their reach. They can prove the site does what it says mechanically; they cannot referee a vowel. Phonetic accuracy rests on the rules above and on the corrections you send.

Editorial language review

Guides and FAQ answers may use absolute-sounding language — always, never, every, usually beats — and some of that language is defensible: Italian ch before e or i really is always a /k/, and the /d/ of English -ed endings really never changes the spelling. The editorial question is whether the absolute can be backed by the source it sits next to, not whether the word exists on the page.

During review, check each absolute claim against these categories:

Linguistic rule with no exceptions
Keep the absolute. Italian ch before e/i is /k/; German w is /v/. These are standard, well-documented rules.
Linguistic rule with rare exceptions
Use "generally" or "almost always" and note the exception when it is nameable. Japanese g is "generally hard" (nasalised before vowels in some dialects). German stress "almost always" falls on the first syllable.
Comparative or efficacy claim
Requires a citable source or must be rephrased as a general observation. "Ten minutes of listening beats an hour of repeating" became "Research on perceptual learning suggests focused listening practice is a more efficient use of time than repetition alone."
Claim about third-party technology
State what we know and qualify what we do not. We know server voices are Deepgram Aura-2 models; we do not know whether a given OS-provided device voice was trained on native speakers, so say "designed for that variety" rather than "built from native speakers."

The goal is not to eliminate every strong word — that would make the writing hedged and unhelpful — but to ensure each absolute is either defensible by the source or replaced with language the evidence supports.

Corrections

If a transcription, an audio clip or a claim in a guide is wrong, email hello@learnpronounce.xyz with the word and what it should be. Corrections to phonetic data are treated as bugs, not as feedback: they go ahead of new content. The change lands in the project's commit history, dated and attributed.

Funding and independence

Learn Pronounce is free, carries no advertising, no affiliate links and no sponsored content, and sells nothing. Nothing on the site is placed in exchange for payment. The site loads Google Analytics and Microsoft Clarity to count visits, and only if you accept the consent banner — the full detail is in the privacy policy, which is written to say what actually happens rather than what sounds best.

Known limits

A policy that admits nothing is not worth reading, so: this is a one-person project, so there is no second reviewer and no formal sign-off before publication. Synthesised speech, however good, is a model of an accent and not a recording of a native speaker. Where a word is genuinely contested between dictionaries, showing one transcription per accent necessarily flattens a real disagreement.

Any of these can be improved, and reader corrections are how it happens. Read the about page for how the tool itself works.