Open pronunciation research · v1

Inside the Echoes pronunciation graph

Think of pronunciation as a map. Each sound sequence is a place on the map. Two places connect when you can change one speech sound to travel between them. That simple rule creates the network behind Echoes.

By · Published · Updated

14,402pronunciation nodes
54,345one-phoneme links
2,879connected components
75.4%in the largest component

Graph rule

What counts as one link

A phoneme is one speech sound. We remove stress marks from each primary CMUdict pronunciation, then connect two nodes only when one sound sequence can become the other by inserting, deleting or replacing exactly one phoneme.

Homophones share a node because they sound the same in this dictionary. Spelling may change a lot across one sound hop, while two similarly spelled words may not connect at all.

Scope and limits

A pinned US-English dictionary graph

The graph uses primary pronunciations from CMU Pronouncing Dictionary commit 74790861f652 under its BSD-2-Clause licence. It does not use machine-estimated spelling-to-sound pronunciations, also called grapheme-to-phoneme or G2P rows, from other site data.

The common-word universe and display-spelling choices use wordfreq (Robyn Speer) 3.1.1. Its code is MIT; the frequency data are CC BY-SA 4.0.

It cannot represent every accent, alternate pronunciation, inflection or English word. The common-word threshold and safety exclusions also shape the available paths. A missing link describes this released dataset, not every speaker's pronunciation.

Worked route: PULLS → POLLS → POSE → CHOSE → CHOKE. Each step changes one phoneme in the released dictionary pronunciation.

Reusable figure

The released graph at a glance

The Echoes graph contains 14,402 pronunciation nodes and 54,345 one-phoneme links.

Download SVG · Download PNG. Both carry the release scope; retain the source line when reusing them.

Bounded sample

One sound away from CAT

Download sample JSON

The sample begins at CAT and includes every node one phoneme edit away, plus all links among those 38 nodes. It is small enough to inspect in a lesson or code example. It is not a replacement for the complete internal graph.

Released pronunciation nodes one sound edit away from CAT
Displayed wordPronunciationDistance from CAT
BATB AE T1
CABK AE B1
CADK AE D1
CADGEK AE JH1
CALK AE L1
CALFK AE F1
CAMK AE M1
CANK AE N1
CANTK AE N T1
CAPK AE P1
CAPPEDK AE P T1
CASK AE S1

Classroom-ready activity

Build a four-hop sound ladder

The two-page A4 sheet asks students to order the route from PULLS to CHOKE and name the sound added, removed or replaced at each step. It also asks them to compare the dictionary evidence with accents spoken in the room. Page two gives the route, a 20-minute lesson, sources, licence, method and data cutoff.

Download colour PDF · Download black-and-white PDF

The graph universe includes accepted five-letter CMUdict words plus common 3–7-letter CMUdict words at wordfreq Zipf 3.2 or above. The sheet uses primary US-English dictionary pronunciations as one inspectable source, not as a correctness test for every accent.

Immutable release v1

Downloads, methods and checksums

The manifest binds every artifact to a SHA-256 checksum and records its licence. Graph-derived tables, samples, figures and worksheets are CC BY-SA 4.0; standalone original documentation is CC BY 4.0; CMUdict source material remains BSD-2-Clause. Creator credit is required for the Creative Commons files, but no hyperlink is required.

Citation

Cite this released graph

Matthew. Echoes pronunciation graph research release v1. FiveLetterWords, 2026. https://www.fiveletterwords.io/data/echoes-phoneme-graph/v1/manifest.json