Say far out slowly and you hear two words. Say it at speed and the /r/ slides across: fa-rout. Nothing was added. The boundary stopped existing.
Linking has a mechanical cause. Speech is a continuous stream of articulation, and the tongue does not return to a neutral position between words. When one word ends in a vowel and the next begins with one, the tongue has to travel from the first vowel position to the second, and that travel is audible as a glide — a /j/, a /w/ or an /r/ that is written nowhere.
Which glide appears is decided by where the first vowel ends. A vowel that finishes high and front (i, ɪ, eɪ, aɪ) hands over through /j/. A vowel that finishes high and back (u, ʊ, oʊ, aʊ) hands over through /w/. A vowel that finishes mid-central (ɑ, ɔ) often picks up an /r/ in British English.
Consonant to vowel needs no glide at all: the consonant simply attaches to the following vowel, which is why an apple sounds like a-napple.
Vowel into vowel — a glide appears
The dip is smallest here, and these are the examples where you can hear the inserted glide most clearly. The glide is inferred from the vowel the first word ends on, not detected.
60 sentences · 981 measured words · 921 word boundaries, of which 40 are vowel-into-vowel. Below are 10, chosen where the join is cleanest. The rest are the same phenomenon, less clearly recorded.
The shaded band is the measurement window (the join ±60 ms); the thin line is where the aligner put the boundary. The aligner's boundary is an estimate from character timings, not something detected in the audio — read the band, not the line.
Consonant into vowel — they simply join
The consonant moves across to the next word. Nothing is inserted; the closure is just never released into silence.
60 sentences · 981 measured words · 921 word boundaries, of which 196 are consonant-into-vowel. Below are 10, chosen where the join is cleanest. The rest are the same phenomenon, less clearly recorded.
How the join is measured
If two words are linked, the energy between them never falls to silence. So for every word boundary we take the energy envelope (5 ms windows, 2.5 ms steps), find the quietest point within ±60 ms of the boundary, and subtract it from the median energy inside the words on either side. The result is the dip in dB: how far the signal drops at the join.
A large dip means a real gap. A dip near zero — or below it — means the join is as loud as the words themselves, which is what linking sounds like. The waveform under each example is the actual envelope, so you can look at the valley instead of taking our word for it.
We do not use the gap between the aligner's word timestamps. We tried, and it fails: the aligner puts word boundaries on the frames it is confident about, and the audio keeps sounding through the frames it left unassigned. Measured on real recordings, those “gaps” were sometimes louder than the words on either side.