Connected speech — Phoneme·Trick

Word boundaries are mostly invisible in the audio. The gap you see in the spelling is rarely a gap you can hear — and that is what this strand measures.

English is written with spaces between words and spoken without them. When a consonant at the end of one word meets a vowel at the start of the next, the two run together; when a vowel meets a vowel, the mouth inserts a glide that was never written; when a word carries little information, it gets shorter than its spelling suggests.

Each page below takes one of those events and puts real recordings under it, along with the measurement that made us pick that particular example. There are no rules to memorise here — you can hear the difference and see the number behind it.

7 pages

Linking

Where two words meet with no seam at all — and the glide that appears when the meeting point is vowel to vowel.

Weak forms

The small words that shrink: how short they actually get, measured against the rest of their own sentence.

Stress

What actually changes inside a stressed syllable — length, loudness and pitch measured as three separate things, on the same word.

Elision

-ing → -in': the tail of something goes from ŋ to n — nearly identical to the ear, different in spelling and habit.

Reductions

going to → gonna: two ways to say the same thing, and the milliseconds between them.

Assimilation

don't you → doncha: the t and the y at the seam fuse into a single tʃ.

Intonation

The pitch of a whole sentence: where it starts, how it falls, and where the last syllable lands. Drawn from the measurement, not sketched.

One sub-class is still missing

d + y assimilation (did you → didja, would you → wouldja) has no page. The reason is the engine, not the schedule: asked to read didja, it produces d ɪ d j ə with the j holding 71 ms of its own — the two sounds never fuse into dʒ. The t + y class does fuse (gotcha's tʃ measures 161 ms), which is why that half has a page. Producing d + y needs conversational human speech; a synthesiser cannot be talked into it, and we would rather leave the gap visible than fill it with something adjacent and call it assimilation.

What these numbers are, and what they are not

Every figure on these pages is computed from the audio, not from a pronunciation model. The two indicators are simple enough to state in a sentence each: boundary dip — how much the energy drops at the join; duration ratio — how long a word lasts compared with the median word in the same sentence.

What we deliberately do not do is declare “this is a weak form”. A word can be short because it was reduced, because it was unstressed, or because the speaker was hurrying — the measurement cannot separate those. So we report the duration and let the audio settle it.

One limit worth stating plainly: alignment runs on 20 ms frames, so every duration here is a multiple of 20 ms. A vowel lasting 40 ms has two frames of evidence. That is enough to say “short”; it is not enough to say “this vowel is a schwa”.

The audio

Recordings come from LibriSpeech dev-clean: public-domain novels read aloud by volunteers, 16 kHz mono, no music and no background noise. Licence CC BY 4.0, which permits redistribution with attribution.

That matters for honesty in one direction. It is read speech, not spontaneous conversation. A narrator reading a novel articulates more carefully than two people arguing, so the linking you hear here sits at the mild end of the range. If a join shows almost no dip in this corpus, casual speech would show even less.