Start here

Scoring definitions

Every score is counted from the sounds of the recording, with no weights and no adjustments. This page gives each one as a formula, so any number in a response can be checked by hand.

These definitions apply to every recording scored from 18 November 2026.

#The reference, and what 100 means

Every score compares the recording with one reference: the standard pronunciation of the language variant the call names (British or American English, the Spanish of Spain, and so on), as written in Spelly’s pronunciation reference, which follows published descriptions of each language’s sounds. It is not one speaker’s voice, and it is not an accent rating.

The scores are not CEFR levels and do not follow CEFR descriptors. CEFR appears only in Spelly’s own app, to place a learner (see the last section).

  • A recording scores 100 in accuracy when every expected sound was heard right; native_likeness reaches 100 only when each one also has its native detail.
  • A word scores 100 in accuracy when each of its expected sounds was heard right.
  • A sound has no percentage: it gets a verdict, described below, and a confidence from 0 to 1.
  • There is no overall score: combining the scores would take weights someone has to choose.

#What each expected sound gets

The text is turned into its expected sounds, from the reference above. The model listens to the recording freely, and what it heard is lined up with the expected sounds. Each expected sound gets one verdict:

  • Right: heard as the expected sound.
  • Close (is_close): heard as another sound that shares at least two of three features with the expected one: where it is made, how it is made, and whether the voice is on. /s/ for /z/ is close.
  • Other sound: heard as another sound that shares fewer.
  • Not said (predicted is empty): nothing the model heard was lined up with it. This is not the same as silence: the model heard nothing it could pair with this sound.

A sound heard that the text does not have is an extra sound (is_insertion). Extra sounds are listed with the word they were heard in, and lower no score.

#Confidence and faintly heard sounds

The recording is also aligned to the text: each expected sound is placed where it should be, and its confidence is the probability the model gives that sound at its place, from 0 to 1. For an extra sound, confidence is how sure the model was of what it heard.

A sound not said whose confidence is 0.15 or more is faintly heard (is_faint): the model still finds it at its place, weakly. It is a label only, and every score counts it as not said.

#The scores

Each recording gets five scores from 0 to 100, and each word its own accuracy and clarity.

ScoreHow it is counted
accuracyExpected sounds said right ÷ expected sounds.
completenessExpected sounds said (right, close or other) ÷ expected sounds.
native_likenessExpected sounds said right with their native detail (the English /tʰ/, not only /t/) ÷ expected sounds.
fluencyGaps between neighbouring words with no stop of 250 ms or more ÷ gaps counted. A gap runs from the end of one word to the start of the next. A gap next to a word the alignment could not time is not counted. With no gap counted (one word, or no two neighbouring words timed), fluency is null.
clarityThe median confidence of the sounds heard, extra sounds included.
analyzed_text[].accuracyThe word’s expected sounds said right ÷ its expected sounds.
analyzed_text[].clarityThe median confidence of the word’s sounds heard.

#Rounding

A recording’s scores are given as the nearest whole number, halves up: 157 sounds right out of 200 is 78.5, given as 79. The counts are divided exactly before rounding, so a half is exactly a half. Word scores are given unrounded.

#What the scores leave out

  • Fluency counts stops between words, not speed.
  • Native speakers stop too, at commas and between phrases, so a fluency below 100 is not a mistake in itself.

#In Spelly’s app

Spelly’s own app shows these scores with one addition for a learner placed at a level: a close sound earns part of a point, a whole point at A1 down to none at C2. The API always answers the scores as defined on this page.

Questions?

Tell us what you have in mind and we will get back to you.

Get in touch