01What this site does
It splits pronunciation into three layers: how a single sound is made, how syllables are built, and how sounds change in connected speech.
Each page answers one question, and answers it with checkable numbers — not "/θ/ needs the tongue between the teeth", which everyone already says, but "how long /θ/ actually measures in this corpus, and how it compares with /s/". That number is recomputed on every build; it is not typed in.
02Where the data comes from
The audio is synthesised (four American voices, two male, two female). Nothing is clipped from films or textbooks, so there is no rights problem and sentences can be added at any time.
Durations, pitch and harmonics-to-noise are computed from the audio using forced alignment (MMS_FA) and acoustic analysis — not copied from a dictionary. The dictionary only decides which sounds a word should contain; how long they actually are is taken from the recording.
Where we cannot measure, we do not print a number. For linking, the page reports how many dB the energy dips at the boundary, not "this is linked" — the second is a judgement, the first is a measurement.
03Why only English is complete
Because we ship one language at a time: take a language from phonemes all the way through connected speech, get it right, then repeat the pipeline for the next one. English currently has 41 phoneme symbols, 41 of them with their own page, 7 connected-speech phenomena (7 live), across 24 content pages.
Spanish has one phoneme (θ) — a pilot to prove the same pipeline can carry another language, not a half-finished product. Japanese has not started, because its pitch accent needs a separate template; copying the English phoneme inventory would only produce a page that looks plausible.
So the unlit cells on the home page are not gaps — they are a queue.
04Why some phoneme cells are grey
Grey means the page has not been built yet, not that the sound does not exist. A phoneme inventory is a closed list; pages are built one at a time.
The reverse also happens: a page exists but has no minimal-pair difference, because of a limit in our measurement — not because learners do not need it. Those two cases are labelled separately and never merged.
05What this site does not have
No accounts, no email collection, no analytics.
Ads load only after a visitor accepts; choosing "essential only" leaves the site fully working.
No autoplay and no recommendation engine — pages are joined by structure, not by an algorithm.
—Contact
If a number is wrong, a sentence is odd, or a clip sounds off, just tell us. Every feedback item is pre-filled with where it came from (page, file, word, transcription). [email protected]