Technology
How does speech recognition work?
You speak, and a second later there are words on a screen. In between, sound turns into numbers, numbers into a picture of frequencies, and a neural network trained on huge amounts of speech turns that picture into text.
In short. A microphone samples sound thousands of times a second; software turns the samples into a spectrogram, a picture of frequencies over time; a neural network turns that picture into words. Modern systems learn from enormous amounts of speech: one widely used family was trained on 680,000 hours. Accuracy is measured as the word error rate, and it still rises with noise, distance, accents and mumbling.
Step by step
- Sound to numbers. A microphone turns air pressure into an electrical signal, and a converter measures it thousands of times a second.
- Numbers to a picture. Short slices of the signal are analysed for which frequencies they contain. Stacked in time, they make a spectrogram: speech sounds have recognisable shapes in it.
- Picture to words. A neural network reads the spectrogram and predicts the most likely text, one piece at a time, using what it learned about both sounds and language.
- Words to meaning. For an assistant, the text then goes to a language model that works out what you want.
Learning from huge amounts of speech
Older systems were built from separate parts for sounds, pronunciations and words. Today’s are mostly single neural networks trained on transcribed audio. In 2022, researchers at OpenAI described a system trained on 680,000 hours of multilingual audio with transcripts from the internet. The resulting models handled standard tests well without being tuned for them, and came close to human accuracy and robustness.
How accuracy is measured
The word error rate counts the words a system gets wrong, leaves out or adds, as a share of the words spoken. A rate of 5% means about one word in twenty is off. The same system can score very differently on read speech in a quiet room and on a mumble in a busy street, which is why a single number says little without its conditions.
Why it still goes wrong
- Noise and distance. The further the microphone, the more the room and the street get in.
- Voices it rarely heard. Accents, dialects and ages that were rare in the training data get more errors: voice assistants and accents.
- Names and rare words. A name it has not seen before is easy to mishear as a common word.
- Mumbling and talking over each other. Less signal, more guessing.
On the device or in the cloud?
Small models can run on a phone or a dock; the largest run in data centres. The trade-off is between privacy and battery on one side and accuracy and languages on the other: edge AI vs cloud AI.
How we approach it
A ring hears you from your hand, not from your mouth, so we test speech recognition through a simulated ring microphone, with noise, distance and mumbles, in six languages: can a ring understand a mumble? On Lychnos Ring, your voice is sent only while you hold the button, becomes text in our cloud in the EU, and the clip is deleted right after. It is in development.
Source, read 4 November 2026: Radford and colleagues, robust speech recognition via large-scale weak supervision, 2022 (arXiv).
Questions
How does speech recognition work?
A microphone turns sound into numbers; software turns those into a picture of which frequencies are present over time; a neural network trained on large amounts of transcribed speech turns that picture into the most likely words.
How accurate is speech recognition?
It is measured with the word error rate, the share of words wrong, missing or added. Modern systems can come close to people on clean speech, but noise, distance, accents and mumbling still raise the error rate.