Short answer: When a language app scores your pronunciation, a speech recognition model first works out which sounds you produced and when, then compares each sound with what a model of native speech expects for that word. The result is usually several scores: accuracy (how close each sound is to the reference), fluency (pauses and breaks), completeness (how much of the target text you said) and sometimes prosody (stress, intonation and rhythm). These scores are useful for spotting specific sounds to fix, but they measure distance from one reference accent, not whether people understand you. Read them as trends and word-level hints, not as a grade of your ability.
How Pronunciation Scoring Works, Step by Step
- Recording. The app captures your audio. Background noise and a poor microphone already affect everything that follows.
- Recognition and alignment. A speech recognition model converts audio into likely sounds (phonemes) and lines them up in time with the words you were supposed to say. In unscripted practice there is no target text, so the model first guesses what you meant to say.
- Comparison. For each phoneme, the system estimates how likely it is that you produced the expected sound rather than another. This idea goes back to the "Goodness of Pronunciation" measure described by Witt and Young (2000), and modern systems use the same logic with neural networks.
- Aggregation. Phoneme scores are combined into word, sentence and overall scores, and timing information becomes a fluency score.
General-purpose recognisers such as OpenAI's Whisper are trained to transcribe speech robustly, which means they are designed to tolerate accents. That is good for understanding learners but not for grading them, which is why pronunciation scoring needs a separate comparison step.
What Each Pronunciation Score Measures
Microsoft's Azure pronunciation assessment, a service used to build language-learning features, documents its scores publicly. They are a good reference for what most apps report, even when the names differ.
| Score | What it measures (Microsoft's definition) | Level |
|---|---|---|
| Accuracy | How closely the phonemes match a native speaker's pronunciation | Phoneme, word and full text |
| Fluency | How closely your use of silent breaks between words matches a native speaker's | Full text |
| Completeness | The ratio of pronounced words to the reference text | Full text (scripted only) |
| Prosody | How natural the speech is, including stress, intonation, speaking speed and rhythm | Full text |
| Overall | A weighted combination of the scores above | Full text |
Scores are given on a 0 to 100 (or 0 to 5) scale. Microsoft's documentation, checked November 10, 2026, also notes that prosody assessment is only available for US English, which is one reason prosody feedback is rarer in other languages.
What Pronunciation Scores Do Not Tell You
Accent Is Not the Same as Being Understood
Research on second-language speech separates three things: accentedness (how different you sound), comprehensibility (how easy you are to understand) and intelligibility (how much listeners actually understand). In a classic study, Munro and Derwing (1995) found that a strong foreign accent does not necessarily make speech less intelligible. A score that measures distance from one native model mostly tracks accentedness, so a "70" may still be perfectly clear to real listeners.
The Reference Accent Is a Choice
"Native pronunciation" in a scoring model usually means one variety, such as US English or Castilian Spanish. If you are learning a different variety, correct sounds may be marked as errors.
Recognisers Are Not Equally Accurate for Everyone
Speech recognition still makes more errors for some groups of speakers. A study in PNAS (Koenecke et al., 2020) found that five commercial systems had an average word error rate of 0.35 for Black speakers compared with 0.19 for white speakers. Some learner accents may be penalised for the same reason.
Short Phrases Give Noisy Scores
One word or a short sentence gives the model very little evidence, so scores can jump between attempts even when you said it the same way.
How to Use Pronunciation Scores Without Frustration
| Instead of... | Do this |
|---|---|
| Chasing 100 on one sentence | Track your weekly average across many sentences |
| Reading only the overall score | Open the word or sound view and find the one sound that is flagged repeatedly |
| Repeating the same phrase 20 times | Fix one sound, then use it in new sentences |
| Treating the score as your level | Test real understanding: can a listener, or a voice assistant, understand you first time? |
| Practising in a noisy room | Use a headset microphone in a quiet place so the score reflects you, not the room |
Pronunciation also improves through listening. Shadowing (repeating native audio in real time) trains rhythm and stress, which single-word scores rarely capture.
Choosing an App by What It Measures
Apps differ in how much of this they show. Pronunciation specialists such as ELSA Speak focus on sound-level scores (see our ELSA Speak review). Conversation tutors focus more on fluency in free speech: Enverson AI, for example, reports separate measurements for vocabulary, grammar, speaking pace, fluency and filler words after a Free Talk session. If people often ask you to repeat yourself, start with a sound-level tool; if you are understood but slow, prioritise fluency practice. Our methodology explains how we test speech recognition when we review apps.
Frequently Asked Questions
How do language apps score pronunciation?
They use speech recognition to identify the sounds you produced, align them with the words you were meant to say, and compare each sound with a model of native speech. The comparisons are combined into accuracy, fluency, completeness and sometimes prosody scores.
Why does my pronunciation score change every time?
Short phrases give the model little evidence, and small differences in noise, microphone distance or timing change the result. Look at averages over many sentences rather than single attempts.
Does a low pronunciation score mean people cannot understand me?
Not necessarily. Scores measure distance from one reference accent. Research shows that a noticeable accent does not always reduce how well listeners understand you.
What is a prosody score?
Prosody covers stress, intonation, speaking speed and rhythm. Microsoft's pronunciation assessment, for example, only offers prosody scoring for US English.
Is AI pronunciation feedback accurate for all accents?
No system is equally accurate for every speaker. Studies have found higher recognition error rates for some groups, so treat scores as hints and check them against real listeners.
Sources
- Microsoft Learn: Use pronunciation assessment (checked November 10, 2026)
- Witt and Young (2000), Phone-level pronunciation scoring and assessment for interactive language learning, Speech Communication
- Munro and Derwing (1995), Foreign accent, comprehensibility, and intelligibility, Language Learning
- Koenecke et al. (2020), Racial disparities in automated speech recognition, PNAS