A little more curious, every day.
How machines are learning to listen, measure and help, and where their limits still lie.
7 storiesDeep learning can now pull a voice out of a noisy room better than any filter before it. The hard part is making it fit behind an ear, run on a battery the size of a lentil, and answer in under ten milliseconds.
From free NHS devices to £3,000 private clinics and £40 Amazon amplifiers: an audio-electronics analyst takes them apart.
From two tiny microphones a few millimetres apart to neural networks that take a fingerprint of the room, here is how a hearing aid decides which sounds deserve your attention, and how scientists check that it got it right.
Rings and watches now score our nights down to the minute of REM. A new study suggests your own morning verdict still carries information the sensors miss.
When a gravitational-wave alert goes out, telescopes around the world have minutes to decide whether to chase it. A machine-learning system called GWSkyNet-II now helps them tell a cosmic collision from a hiccup in the detector.
Google trained a version of its Gemini model to read a month of wearable data and coach people on sleep and fitness. The results are impressive in places, and instructive in others.
In 1952 a machine at Bell Labs could recognise the digits zero to nine, spoken by one man. Seventy years on, software transcribes almost anyone, in almost any language. Almost is the interesting part.
