A four-year-old does not finish sentences

Conversation held together on adult voices and fell apart the moment a child got the microphone. We went looking for the reason, and most of the answer turned out to be sitting in published work already. Nobody puts it in one place.

It starts with physics. A five-year-old sounds in a different register: in the measurements by Lee, Potamianos and Narayanan, a five-year-old boy's vowel in a word like "bought" runs at 266 Hz against an adult man's 135, with a first formant of 1166 against 723. The models we take off the shelf learned mostly from the second voice.

Then the data, where the gap is indecent. Whisper was trained on 680,000 hours of speech. Everything publicly available in children's English is measured in hundreds: MyST at 179 hours, the two OGI corpora together at around a hundred, CMU Kids at nine. In a published benchmark, zero-shot Whisper-large returns 43.9 per cent word error on ages four to seven, and 18.6 on teenagers 🎙️

Sounds are their own story. "Wabbit" for "rabbit" arrives on time: ASHA puts the disappearance of that substitution at six to seven years, and the Crowe and McLeod review of American data puts the consonant "r" in the sixth year of life. A system listening to a four-year-old has to work with someone who physically cannot produce half the consonants yet.

What caught us hardest was something else: how a machine decides a person has finished. Almost everywhere it is silence on a timer. In OpenAI's Realtime API half a second of quiet counts as the end of a turn, Pipecat sets the threshold at two tenths, LiveKit at a half.

Now the same thing from the child's side. In Melissa Redford's work, five-year-olds stopped in the middle of a story nearly twice as often as adults, 1.98 times per hundred words, and that effect holds. Eighteen per cent of their stops landed where the phrase is grammatically incomplete: after a preposition, a conjunction, an article. For adults it was seven. The sample is tiny, ten children, so it points a direction and settles nothing 🧒

Put the two together. A child takes a breath after "and then he fell into", the system counts its half second and walks confidently into the middle of the sentence. What follows is simple: a four-year-old who gets interrupted goes quiet.

We have no answer yet, only a question worth asking properly: what tells you a child has finished, once silence no longer does. Everything that comes to mind ends in waiting, and every extra fraction of a second is latency. That is the next post ⏱️