y+acompasado
[i] + [a]
+14.6dB
AMARILLA DE INVISIBLE TéRMINO ACOMPAñáNDOLA FIELES EN ESTE CAMBIO DE SITUACIóN REGULAR Y ACOMPASADO COMO EL DE UN PéNDULO
Three numbers
−1.2dBinside a word108 junctions
−1.5dBacross a word boundary385 junctions
0.43p-value of the differencenot significant
The number is the energy dip at the junction: the quietest 5 ms inside a fixed 40 ms window centred on the junction, compared with the level of the two flanking vowels. 0 means nothing drops; more negative means more like a break. If the word boundary really caused a pause, the cross-word group would be clearly lower. It is not.
The two distributions
Each bar is the share of one group falling in a 4 dB band. The groups differ 3.6× in size, so each is normalised to its own percentage — otherwise the chart would only show which group has more samples. Further left = deeper dip = more like a break.
The two sets of bars almost coincide: the median dip at a word boundary is only 0.3 dB deeper. The cross-word group has the longer left tail — that 12% is the third group of examples below.
Three rules
Linking in Spanish is far more regular than in English: it can be computed from the spelling, with no list of exceptions.
01
vowel + vowel → one syllable (sinalefa)
A word ending in a vowel followed by a word starting with a vowel merges into a single syllable, so the syllable count goes down. mi amigo is mia-mi-go (three syllables, not four); todo estaba is to-doe-sta-ba.
02
final consonant → moves to the next word (enlace)
Spanish avoids syllables ending in a consonant, so a final consonant attaches to the following word. un oso → u-no-so.
03
final s → next syllable (and voices)
los amigos is not los | amigos but lo-za-mi-gos — the s becomes the onset of the next syllable and voices to [z] between vowels. This is the one you can hear most easily: a clean [s] plus a small pause is what makes a native speaker hear a word list rather than speech.
Together these mean the actual syllable count is usually lower than the word count. That is also why word-by-word reading sounds mechanical — the syllable count is already wrong.
Twelve junctions, three tiers
Chosen from the 385 cross-word junctions in the corpus. The first two tiers show linking is real; the third shows it is not obligatory.
linked (no energy drop at the junction)
The junction is louder than the flanking vowels — the airstream never stopped.
que+una
[e] + [u]
+12.5dB
TODO ESTABA LO MISMO QUE UNA HORA ANTES CUANDO EL Té HUMEABA EN LA TAZA DE OJEDA AHORA VACíA Y BLANQUEABAN SOBRE LA MESA LOS PLIEGOS
a+aparecer
[a] + [a]
+10.8dB
LOS MISMOS OCULTáNDOSE PARA VOLVER A APARECER CON REGULARIDAD CASI MECáNICA
había+ocurrido
[a] + [o]
+9.4dB
ALGO NUEVO HABíA OCURRIDO EN TORNO DE éL MIENTRAS CON EL PECHO EN EL FILO DE LA MESA Y LOS OJOS SOBRE LOS PAPELES HUíA LEJOS
typical (slight dip, still linked)
The commonest shape: a slight dip, nowhere near silence.
mismo+aspecto
[o] + [a]
−7.8dB
VIO CON EL MISMO ASPECTO EXTERIOR COSAS Y PERSONAS AL SALIR DE SU ABSTRACCIóN
altura+eran
[a] + [e]
−6.8dB
DOS SALONES AGRANDADOS POR LA ESCASEZ DE SU ALTURA ERAN EL CAMPO VISUAL DE OJEDA
tostado+husmear
[o] + [u]
−6.5dB
MANTECA Y MERMELADAS EL PAN TOSTADO HUSMEAR EL PERFUME DEL Té O SEGUIR EL BURBUJEO DE LAS AGUAS MINERALES TEñIDAS DE JARABES Y LICORES
té+o
[e] + [o]
−6.3dB
MANTECA Y MERMELADAS EL PAN TOSTADO HUSMEAR EL PERFUME DEL Té O SEGUIR EL BURBUJEO DE LAS AGUAS MINERALES TEñIDAS DE JARABES Y LICORES
actually broken
There is no punctuation between these words; the speaker breathed. 12% of junctions in read speech look like this — linking is the default, not an obligation.
reflejo+húmedo
[o] + [u]
−49.9dB
LA BLANCURA DE LA MADERA LAQUEADA TEMBLABA CON CIERTO REFLEJO HúMEDO QUE PARECíA VENIR DEL EXTERIOR
rosa+igual
[a] + [i]
−40.9dB
LOS ASIENTOS TAPIZADOS DE SEDA ROSA IGUAL A LA QUE ADORNABA LOS PLANOS DE LAS PAREDES
perspectiva+al
[a] + [a]
−39.6dB
QUE CERRABAN LA PERSPECTIVA AL OTRO LADO DE LOS CRISTALES LIGERAMENTE TURBIOS POR LA HUMEDAD EXTERIOR
pilastra+a
[a] + [a]
−29.5dB
MADERA TENDIDOS DE PILASTRA A PILASTRA Y EL VERDE AMARILLENTO Y VELLUDO
Where these numbers come from
Corpus: VoxForge Spanish — 15 volunteers, 145 recordings, 2293 words. 16 kHz mono, MMS_FA forced alignment, all char-anchored.
Measure: when both sides of a junction are vowels, take a fixed 40 ms window centred on the junction, compute RMS every 5 ms, take the minimum, and compare it with the quieter of the two flanking vowels, in dB. Using the quieter one is conservative — we would rather under-report a dip than invent one.
Control: vowel+vowel inside a word uses the identical window width, algorithm and reference level; the only difference is that the window centre has no word boundary. This rules out confounds such as "vowels are simply louder than consonants". Result: −1.2 dB vs −1.5 dB, Mann-Whitney p = 0.43, and 14 of 15 speakers differ by less than 6 dB.
Counter-examples reported: 12% of cross-word junctions fall below −20 dB, the deepest at −49.9 dB. One speaker pauses between almost every word (51 words per minute, the slowest in the corpus). So the claim is "linking is the default", not "Spanish always links".
Three failed attempts (they decide the number above, so they belong here): ① we first tried to measure the "inter-word gap in milliseconds" — all four groups came out at exactly 60 ms and 92.3% of values landed on integer multiples of the aligner's frame, because word boundaries are trimmed inward: the quantity does not exist; ② then final-/s/ voicing — 0.20 before vowels and 0.20 before voiceless consonants, within-speaker median difference −0.06: no signal; ③ then "vowel+vowel vs consonant+consonant" — a large gap (−1.6 vs −8.6 dB) but confounded, because stops have closures of their own. Only then did we settle on this inside-vs-across fixed-window design.
Measurements, not verdicts. Failing to link is usually not about speed — it is about treating each word as a separate unit, which is the syllable-timed habit carried over from Mandarin.