Why did my pronunciation score change?

You recorded the same sentence twice. The first attempt scored 82. The second scored 64. Your pronunciation ability did not disappear in thirty seconds.

Before treating either number as a diagnosis, work out what changed between the recordings. A pronunciation checker hears the audio it receives. A different microphone position, a clipped first syllable, background speech, or an exaggerated second attempt can move the result even when your underlying skill is much the same.

This is a troubleshooting guide, not another explanation of whether AI pronunciation scoring is accurate. The job here is to isolate the variable behind a surprising result.

Start by sorting the change into one of four patterns

Write down the scores rather than relying on how the session felt.

Pattern	More likely explanation	Next check
One low score between two similar scores	A stray recording or momentary production error	Replay the low attempt and record once more without changing your technique
Every word drops at the same time	The room, microphone, browser, or speaking volume changed	Run the recording checklist below
The same sound stays weak across several words	A repeatable pronunciation target	Test that sound in a new word and a short phrase
Single words are stable but a full sentence drops	Pace, omissions, clusters, or sentence rhythm may be involved	Record the sentence in two chunks, then reconnect it

A sequence of 80, 63, 79 points somewhere different from 63, 65, 61. The first asks for a recording check. The second gives you a stable pattern to investigate.

The recording checklist

Keep the next three attempts as comparable as possible.

• Use the same device and browser.
• Keep the microphone in the same position.
• Record in one quiet room with no other speaker nearby.
• Use a normal volume. Shouting changes the signal as well as your articulation.
• Start recording before the first sound and stop after the final sound.
• Keep the target language and test page the same.
• Avoid switching from a natural pace to slow, separated syllables halfway through the comparison.
• Replay at least one recording through headphones. Listen for clipping, echo, background speech, and missing word endings.

Microsoft, whose pronunciation assessment QwikSpeak uses, recommends a quiet environment, one speaker, a microphone close to the speaker, normal speed and volume, and audio quality of at least 16 kHz. It also advises repeating low-scoring recordings in difficult conditions. (Microsoft’s pronunciation-assessment limitations and recording guidance)

Phone and laptop audio processing can also behave differently. Noise suppression may help in a steady room but react badly to music, fans, or another voice. The easiest test is not to diagnose the software. Record the same prompt three times without changing devices or location.

Minute 0–1: choose a control prompt

Use one short sentence you can read comfortably. Do not choose a tongue twister or a word you have just learned. You are testing the recording process first.

Minute 1–3: record three natural attempts

Keep your distance, pace, and volume steady. Do not correct yourself mid-word. Note all three scores.

If the scores sit within a narrow range, use their general range as the baseline. There is no need to average them to two decimal places. If they swing widely, replay the audio before working on mouth position.

Minute 3–4: listen without looking at the numbers

Ask:

1. Did the first sound get cut off?
2. Is the final consonant audible?
3. Did I skip or replace a word?
4. Did background sound overlap the target?
5. Did one attempt become much slower or louder?

A number can make a clean recording feel bad and a noisy recording feel persuasive. Playback gives you another source of evidence.

Minute 4–6: change one variable

If the audio was noisy, move rooms but keep your delivery natural. If the first syllable was clipped, wait a beat after pressing record. If the sentence became exaggerated, return to normal pace.

Record two more attempts. Changing the room, microphone distance, pace, and target word at once tells you nothing about which change helped.

Minute 6–7: decide what kind of problem you have

• Scores stabilize after the setup change: treat the original outlier as a recording problem.
• The same word remains weak in clean recordings: open that word and inspect its sounds.
• Only the full sentence remains weak: practise phrasing or consonant clusters rather than drilling one vowel.
• The recordings still jump without an audible reason: collect another session on a different day before naming a weakness.

When the change probably reflects pronunciation

A low result becomes more credible when the same sound appears in several failures.

Suppose English street, strong, and storm are repeatedly weaker than simpler words recorded in the same room. The shared clusters deserve attention. If only street failed once and its first sound was clipped, the evidence is weaker.

Use three levels:

1. Same word: Reproduce the result before treating it as stable.
2. Different word, same sound: Check whether the problem travels with the sound.
3. Short phrase: Confirm that the sound survives when speech speeds up.

Practice should begin only after the pattern survives those checks. This prevents a noisy take from deciding what you study for the next week.

When to stop retaking

Repeatedly chasing the highest number changes the task. By attempt nine, you may be imitating one recording in an unnatural voice rather than improving speech you can use elsewhere.

Stop once you have either:

• three reasonably consistent recordings; or
• the same weak sound in two or more words.

Then move from measurement to practice. Return later with a fresh word to see whether the change held.

QwikSpeak’s checker uses three short prompts and shows the result from each recording. Free accounts have five scored recordings per day, so the troubleshooting sequence should be deliberate: retake a surprising result, then spend the remaining recording on the pattern most likely to be real.

Run a controlled pronunciation check in QwikSpeak →

Keep the room, device, and delivery steady. If the pattern survives, you have something worth practising.