Why voice assistants struggle in noise.

Three things that make it hard

• Noise. A coffee machine, traffic or music covers parts of the sound that tell one word from another, especially the quiet consonants.
• Distance. Your voice fades quickly with distance, while the room keeps its echo and its other sounds. A microphone across the room hears more room and less of you.
• Other voices. When people talk at once, the system has to work out who said what, and a neighbour’s sentence can end up in your request.

What the research measured

Distance, in meetings. OpenAI’s 2022 paper on robust speech recognition tested its models on the AMI corpus of recorded meetings. The largest model made about 17 errors per 100 words with microphones worn by the speakers and about 36 per 100 with a single distant microphone. Across all ten model sizes, the distant microphone roughly doubled the errors.

Noise. The same paper mixed white noise and the chatter of a crowded pub into clean read speech. Every model got worse quickly as the noise rose. Models trained mainly on clean audiobooks did better than OpenAI’s model when the noise was faint, but worse once pub noise came close to the level of the speech. OpenAI’s model was trained on a very wide range of real recordings, from many environments, recording set-ups, speakers and languages.

Dinner parties. The CHiME-6 challenge recorded twenty dinner parties of four friends each, in twenty real homes, with microphone arrays around the rooms and the ordinary sounds of kitchens and air conditioning. Even when the systems were told where each person’s turn began and ended, the 2020 baseline systems got about half the words wrong; when they also had to work out who was speaking, between three quarters and four fifths. These were research baselines, not commercial assistants, and the speech was relaxed conversation rather than commands, but they show how far the problem is from solved.

Languages. In the same paper, the error rate for a language fell steadily with the amount of training audio in it, and most languages had less than a thousand hours. Accents tell a similar story: voice assistants and accents.

What helps

• A microphone close to your mouth. The cheapest improvement there is: the closer the microphone, the more of you and the less of the room it hears.
• Telling the device when you speak. A button you hold while talking marks exactly when the request starts and ends, so it does not have to guess from the noise: hold to talk versus wake words.
• Asking again rather than guessing. When an assistant is not sure what it heard, a short “say that again” costs less than doing the wrong thing; anything that sends, pays or deletes should ask first anyway: AI that asks first.
• Models trained on real-world audio, with noise, distance, accents and mumbles in it, not only clean recordings: how speech recognition works.

On a ring

A ring has a small microphone, and it is used in exactly the places this guide is about: streets, cafés, kitchens. It also has an advantage: you raise your hand towards your mouth and hold a button while you speak, so the microphone is close and knows when you are talking. Understanding speech in noise is the first thing we are working on for Lychnos Ring, ahead of everything else, because an assistant that mishears wastes your time and its battery. What we have learned so far, including a speech model that invented words: speech on a ring and what our ring and robot learn.

Sources, read 5 November 2026: Radford and colleagues, robust speech recognition via large-scale weak supervision, arXiv, 2022; Watanabe and colleagues, CHiME-6 challenge: tackling multispeaker speech recognition for unsegmented recordings, arXiv, 2020.