Fake words — Azure demo

Loading…

Test mode. For each fake word, first play the correct sound, then say it correctly, and see if Azure marks it Right. Then say it deliberately wrong — change one sound, or say a real word that is close (for example nast → nest, or toz → toes), or just guess a different word — and see if Azure marks it Wrong.

This page is a standalone test tool. It is not part of the Climbing With Phonics app. The point is whether Azure can tell if the whole fake word was read right or wrong, not how it grades each sound.

How Azure is being asked. Each recording is sent to Azure Speech pronunciation assessment in en-US (eastus2). Azure is given a reference spelling for the intended sounds (a simpler respelling when the printed word would be misread). Azure does not let us feed it custom phonemes for a made-up word, so this is the best available approach. The main result is Right or Wrong on the whole word (a score of 80 or higher counts as Right). Per-sound scores are tucked underneath and are only extra detail. Close is not used as a main result: a middle score was not a reliable third category in testing.

A sanity check already run with a computer voice, not a child. Fourteen of these words were synthesized correctly, then synthesized wrong in three ways: one sound changed, a nearby real word, and a wild guess. Azure used the same whole-word rule as this page.

What was playedAzure’s whole-word call
Correct reading14 of 14 marked Right (100%)
Completely different guess14 of 14 marked Wrong (100%)
One sound changed7 of 14 marked Wrong (50%)
Nearby real word (nast→nest, toz→toes, and so on)7 of 14 marked Wrong (50%)
Overall42 of 56 calls were correct (75%)

In short: Azure was reliable when the recording was clearly the right word, or clearly some other word. It often still said Right when only one sound was off, or when a similar real word was spoken. It usually printed the target fake word as “what it heard,” even when the audio was wrong, so that line is a weak clue. These were adult computer voices (Jenny), not children.