Recognising spoken words from few recordings
Four command words (go, up, no and on) recognised from short recordings. Brief words, and two that sound very alike.
Short words, so hard ones
A voice command lasts barely half a second. There is little sound to analyse, and any voice or background noise weighs heavily.
The project's vocabulary adds a difficulty: “no” and “on” use almost the same sounds in a different order, and “go” rhymes with “no”. A model must listen to the order of sounds, not just their presence.
Few recordings, a method
When recordings are scarce, we multiply them without inventing anything: the same word is slightly shifted, given more or less noise, or made louder or softer. The model learns what does not change from one voice to another.
We keep aside voices the model has never heard, to measure honestly whether it will recognise the next person.
A voice-controlled machine must first know how not to obey.
Knowing when to abstain
In service, a voice command model must be able to say it is not sure. A word outside its vocabulary, or spoken in noise, must trigger nothing. We set a confidence level below which it does nothing, and every decision is logged.
Key points
- Brief, similar words require listening to the order of sounds.
- Few recordings are offset by realistic variations.
- In service, abstaining when unsure is a feature, not a flaw.