Atypical speech
Voices the recogniser cannot decode: pathological speech, strong accents, irregular diction.
The problem
Speech recognisers are trained on standard speech. Anyone with a speech pathology, a strong accent or irregular diction repeats themselves three times, gets frustrated and ends up typing or giving up.
We work so that the machine learns to understand each person the way they speak, and we measure systems by how they treat those who have it hardest, not by their average. Sample text.
Research challenges
- 01
Scarce and diverse data
There are few recordings of atypical speech, and every person speaks differently.
- 02
Adapting to each speaker
Systems that learn from one person with only a few samples.
- 03
Evaluating for real
Measuring accuracy with the people who have it hardest, not with the average.
Results and impact
What already works outside the lab.
-50% retries with atypical speech
00 users
03 systems in production
Transfer case
Voice-based public services
A public service that answers the phone to everyone who calls, strong accents included.
The line's output
In progress
Jun 2025 – May 2027
Voice-based public services for a regional government
ACRÓNIMO · ID CONT-2025-00
Funded by: Partner organisation
Sep 2023 – Aug 2027
Full title of the research project
ACRÓNIMO · ID 101000000
Funded by: European Commission · Horizon Europe
Previous
Do you work with atypical speech?
Hospitals, associations, public services or companies: tell us about your case and we will find a way to work together.
Collaborate with DELTA