Speech recognition

Also called Voice recognition, Voice control, Dictation, ASR, Speech-to-text

Talking to a computer, and the uncomfortable truth that it works worst for many of the people who need it most.

What it is

Speech recognition turns spoken words into text or commands. Dictation types what you say; voice control operates the device ("open messages", "scroll down", "tap send"). Both are built into every modern phone, tablet and computer, and both have got dramatically better in the last decade for people with typical speech.

For people with atypical speech, the improvement has been slower and patchier, and this entry is mostly about that gap.

Who it can help

People with clear speech and limited hand use: spinal injury, multiple sclerosis, arthritis, upper-limb difference, dyspraxia affecting handwriting and typing, fatigue conditions. For them it is the fastest access method there is, and voice control of a phone can replace a switch entirely.

It is unreliable for dysarthria, for speech after stroke, for many people with cerebral palsy, Parkinson's or motor neurone disease, for Deaf speakers, for people who stammer, and for strong regional accents. The systems were trained mostly on typical voices, and a person whose speech is 60 per cent intelligible to a stranger will get roughly that from a machine, or worse. Research projects are collecting atypical speech to fix this; it is not fixed yet.

How it works

Sound is sliced into tiny frames, matched against a model of what speech sounds look like, and decoded into the most probable words given a language model. Most recognisers send the audio to a server to do this, which needs a connection and raises privacy questions; on-device recognition exists on recent phones and is worth turning on.

Voice control adds a command grammar: numbered labels on everything tappable, "show numbers", "tap 7". Some systems can learn a small set of a person's own words or sounds and map them to commands, which is a real option for someone whose speech is consistent even if it is not standard: a reliable "ah" can be a switch.

Getting started

Try it honestly, for a week, with the built-in dictation and voice control on the person's own device, in the rooms they use. Count what it gets right in ten sentences. If it is above about nine, build on it. If it is below about six, stop, and do not let anyone tell the person to "speak more clearly"; the fault is the software's.

For someone with consistent but non-standard speech, look at custom vocabulary or sound-to-command features in the device's accessibility settings, and at using a handful of reliable vocalisations as switches for switch scanning.

Where to get help

AbilityNet has plain guides to dictation and voice control on every platform and can advise on work set-ups; Access to Work can fund speech-recognition software and training for a job. For people whose speech is not recognised, a speech and language therapist and an AAC assessment is the route, and the ACE Centre covers voice as an access method alongside the others.