Designing voice for under-eights: what actually changes in the ASR stack

The first demos of a children’s voice product usually go well, and then the product meets a four-year-old. What breaks is rarely one thing. It is a stack of small mismatches between how the recognition stack expects speech to arrive and how speech actually arrives from someone whose front teeth are missing.

Where the mismatch actually is

Pitch. A young child’s fundamental frequency sits well above an adult’s, and formant frequencies shift with it. Models handle this better than they did five years ago, but the tail of the distribution — high-pitched, excited, shouting — is still where accuracy falls off first.

Sentence shape. Adults ask questions. Young children issue commands, fragments and compound run-ons: “not that one the other one the blue one”. A recognition stack tuned on question-answer corpora sees utterances it has effectively never been trained to expect.

Distance and motion. Children do not stand at microphone distance. They talk from a metre away, then from across the room, then from half a metre with their mouth against the plush. A far-field array tuned for a conference room is the wrong tool; a single good microphone with the right placement often beats it.

The sleeve. In a plush product, the noise floor is not room noise — it is fabric. Brushing, squeezing, dragging. These are impulsive, wideband, and always in contact with the microphone.

What changes in the stack

Almost none of this is solved by swapping the ASR engine. The changes that move the number:

• Acoustic front end: expect higher pitch bands; trim the low-frequency response, because a child’s voice does not need it and the fabric noise lives in the low end.
• Language model biasing: weight the decoder toward the age band’s vocabulary and the product’s own content. A story plush should be unusually good at the words in its own stories.
• Confidence and re-ask policy: below threshold, the character asks again in persona instead of answering something else. A wrong confident answer costs more trust than an honest “say that again?”
• Wake word selection: run the candidate word against the show and song library the target age consumes. A wake word that appears in a popular nursery rhyme is a product that lights up all evening.
• Reply pacing: recognition is half of it. Young listeners need slower speech and shorter sentences, and they need the reply to start quickly — silence for two seconds reads as broken.

How to test it

Not with adults imitating children — the imitation is always too slow and too clear. The useful test matrix is: real children across the age band; 1 m and 3 m; carpeted and tiled rooms; television on; a sibling shouting in the background. Log failures by age and by condition. The resulting table — not the average accuracy — is the backlog: which condition, which age, which word, which fix.

One more thing worth writing down: measure how often the toy speaks when nobody meant to talk to it. That number — false wakes per evening — is the one parents complain about, and it has nothing to do with recognition accuracy.