h · the only silent letter

Of the 27 Spanish letters, only h makes no sound at all. hola is [ˈola], not [ˈhola]; hora and ora sound exactly alike. This page measured it: across six word pairs differing only by an initial h, five have a waveform correlation of 1.00 — the same stretch of audio.

Four rules: three positions, one exception

The first three are one rule: h is silent everywhere. They are split because Mandarin speakers go wrong in different ways in each position — pronouncing it word-initially, breaking the word medially, and counting an extra syllable in hue/hie.

The number after each row is how many corpus words matched. This corpus has a tiny vocabulary, so several rows show 0: not because that type is rare (ahora, huevo, hacer are among the commonest words in the language) but because these short sentences happen not to use it. The examples are therefore recorded separately (click to hear).

initial h[∅]6 examples · 2 in corpus

Word-initial h is silent; the vowel starts the word: hola = [ˈola], hora = [ˈoɾa].

h inside a word[∅]6 examples · 0 in corpus

H inside a word is silent too, leaving the two vowels in separate beats — a-ho-ra is three syllables with no consonant in the middle.

h + ue / ie[∅]6 examples · 0 in corpus

With the h silent, hue- reads [we]- and hie- reads [je]-: two syllables, not three.

ch[tʃ]6 examples · 12 in corpus

The only case where h “takes part” — but what sounds is the c half: /tʃ/, one unit, with the h itself still silent.

Six homophone pairs: the hardest evidence

Each pair differs only by the initial h. If h took time, the h-word would have to be longer. The two columns on the right are measured: Δ = difference in voiced-segment length (ms), wave = peak correlation of the two aligned waveforms (1.00 means the same audio).

=
Δ +0.0ms · wave 1.00

hello / wave

=
Δ +0.0ms · wave 1.00

hour / he prays

=
Δ +0.0ms · wave 1.00

until / horn

=
Δ +0.0ms · wave 1.00

fact / I throw

=
Δ +0.0ms · wave 1.00

sling·deep / wave

=
Δ +15.0ms · wave 0.47

iron / error

★ The first five show Δ = 0.0 ms and a correlation of 1.00 — the engine produced virtually the same samples twice. The sixth, hierro / yerro, correlates only 0.47: the difference is not in the h (equally silent) but in how ie and ye are read — that belongs to the y / ll axis, see the related pages. It is kept here precisely because it shows the measurement can tell two sounds apart: otherwise a row of 1.00s might just mean the same file was read twice.

Why h is there at all

It used to be a Latin f

hacer comes from facere, hijo from filius, hablar from fabulare. In Old Spanish this h was actually pronounced ([h]); it fell silent around the 16th century, but the spelling never changed. Today's h is a fossil: it marks where a sound used to be.

Silence creates homophones

hola (hello) and ola (wave); hasta (until) and asta (horn) — two spellings, one sound. The same kind of consequence as the ll / y merger: the sound changed, the spelling did not.

In linking, h simply is not there

Taking no time, it cannot block linking either: in mi hora the i and the o are adjacent and merge into one beat — mi hora sounds like mjora. There is no such thing as a pause “because there is an h”.

Three places Mandarin speakers slip

Pronouncing the h

hola becomes “ho-la”, hotel becomes “ho-tel”. English has silent letters, but the habit of reading what you see is strong. In Spanish the h occupies nothing — the following vowel starts the word.

Counting an extra syllable in hue / hie

huevo becomes “hu-e-vo”, three syllables. Once the h is silent, ue is the diphthong [we] and the word has two syllables, [ˈwe.βo]. Same for hielo: [ˈje.lo], not [hi.ˈe.lo].

Assuming the h in ch is silent too

The h in ch is indeed not pronounced, but the c changes: ch is [tʃ], not [k] plus a silent h. The c+h spelling is historical; as a sound it is one unit.

The numbers on this page

Pairs: 6, each word synthesised on its own (es-ES female voice) and measured for voiced-segment length — the first and last frame whose envelope exceeds 5% of the peak.

Criterion: |Δ| ≤ 40 ms counts as “h takes no time”. The threshold is not arbitrary: θ is the Spanish voiceless fricative, the same class as [h] (friction, no periodicity). The 12 measurements on this site run from 40.2 ms (shortest) to 65.4 ms (median): a friction segment that is actually produced cannot be shorter. Measured |Δ|: median 0.0 ms, max 15.0 ms — all inside the threshold.

Waveforms: peak Pearson correlation within ±20 ms after aligning segment onsets. This is a second, independent route — equal duration can be coincidence, equal waveforms cannot. Δ is quantised at 5 ms per frame, so 0.0 means “less than one frame”, not “exactly zero”.

Corpus: 2 words beginning with h, 0 h + ue/ie, 0 with h inside, 12 containing ch. These counts are small because the corpus consists of short sentences; they are not a distribution, which is why the examples are not drawn from it.

★ What was measured is the edge-tts es-ES voice, not “all Spanish speakers”. It shows that this engine gives h no time at all — but that is not what the page rests on: the page rests on the orthographic fact that h is silent, and the measurement merely makes it visible. The same engine reads ll as ʎ where the orthography says ʝ, which is why an engine can never be the arbiter of a language rule.