The conditions for devoicing

Devoicing is not 'speaking faster'. It needs three conditions: the vowel is i or u, both neighbours are voiceless, and the vowel is not word-initial. Remove one and the vowel should stay voiced — which is the half learners miss: すき becomes 'ski', but すず must not become 's-zu'.

Seven words

Click a cell: the first three lose the vowel.

or use the + on each cell

In a sentence

How the word sounds in a sentence. Tap the line to hear it all, tap the highlighted word to hear just that part.

Voiced ratio for seven words

The verdict is computed from the data (threshold 0.40), not pasted from a rule list.

WordRomajiContextVoiced ratioVerdict
すきsukii/u between two voiceless consonants0.22devoiced
くつkutsusame (k and ts)0.29devoiced
きくkikuboth morae are i/u between voiceless sounds0.37devoiced
すずsuzuu between s (voiceless) and z (voiced)0.54stays voiced
ふじfujiu between f (voiceless) and j (voiced)0.65stays voiced
いぬinucontrol: i in word-initial position before voiced n0.87stays voiced
あさasacontrol: a is not a high vowel0.53stays voiced

3 devoiced words: 0.22–0.37; 4 voiced words: 0.53–0.87. The two groups do not overlap, and 0.40 sits in the gap — that gap is the basis for the threshold.

How to practise

Where these numbers come from

Fraction of frames Praat judges voiced inside the speech span. The 0.40 line is not hand-written: the devoicing words measure 0.22 / 0.29 / 0.37 and the others 0.53 / 0.54 / 0.65 / 0.87, so 0.40 sits in the gap.

One rule did NOT reproduce: textbooks also devoice word-final です / ます, but this synthesis reads です at 0.56 — the same band as the voiced group — so the page leaves that claim out.