How Phonema scores pronunciation

Phonema is an iOS app for practising English pronunciation. This page explains how the scoring works and where its limits are, because a pronunciation score is only worth what you know about how it was produced.

The model

Phonema uses a wav2vec2 XLS-R speech model, fine-tuned for phoneme recognition on speechocean762, a publicly released corpus of second-language English speech with human pronunciation annotations. It is converted to CoreML and shipped inside the app.

Everything runs on the device

The model runs on your iPhone. Your recordings are not uploaded, not stored on a server, and not used to train anything. There is no account and no audio upload step — the reason the app works offline is the same reason it is private. See why on-device processing matters.

Scoring is per sound, not per sentence

English is scored across 42 ARPAbet phonemes. For each one, the app computes a Goodness of Pronunciation (GOP) value: how confident the acoustic model is that the sound you produced is the sound the word required, given everything around it. A single sentence-level percentage hides which sound actually went wrong, which is why the app scores each sound separately and reports back at the level of words. What GOP measures goes into detail.

What it cannot do

A score is not a listener. It cannot know your context, your accent goals, or whether a real person would have understood you. Automatic speech scoring and human judgement of pronunciation are related but different things, and the research says so plainly — see intelligibility vs. accent. Phonema is built to tell you which word was unclear, not whether you sound native. That is a deliberate choice, not a limitation we are working around.