Sign up for SAPC2: The Second Speech Accessibility Project Challenge

Registration is now open for SAPC2, the Speech Accessibility Project's second challenge competition.

Contribute to large, significant improvements in automatic speech recognition for people with disabilities. The project's first challenge resulted in a word error rate reduction from 18% (best open-source system before the competition) to 8% (best score reported by the competition winner in February 2025) on the speech of people with Parkinson's disease and other disorders (346 people with Parkinson's, 89 with ALS, 32 with cerebral palsy, 41 with Down syndrome, 16 with stroke). SAPC2 significantly expands the scope of the competition in two ways:

1. Greater diversity of speech patterns. The training and dev corpora distributed to all participants will include the speech of 346 people with PD, 196 with ALS, 195 with CP, 165 with DS, and 97 with stroke. The sequestered test corpora will include the speech of 13 w with PD, 66 with ALS, 58 with CP, 32 with DS, and 24 with stroke.

2. Streaming ASR. A cash prize will be given to every team that places a system on the Pareto frontier of system latency and system accuracy. Exactly one non-streaming system will win, and at least one streaming system will win.

Metrics:

System accuracy: Recent work ("Advocating Character Error Rate for Multilingual ASR Evaluation," Thennal et al., NAACL 2025) suggests that character error rate correlates more closely than word error rate with user perceptions of system accuracy, therefore systems will be scored based on a weighted average of character error rate and word error rate.

System latency of streaming ASRs will be computed on CPU. Latency will be the fusion of two or more measures including time to first token and time to last token. Non-streaming ASR will be assigned a latency of infinity.  

Prizes: A total prize of U.S. $10,000 will be divided equally among all teams with a system on the Pareto frontier of accuracy and latency, as measured using the sequestered test2 set. Exactly one non-streaming ASR will win, and at least one streaming ASR will win.

Scoring: Competitors will submit trained model parameters and inference code up to a maximum number of permitted submissions. Results on test1 will be released within three days of submission. Results on test2 will be released after the close of competition.

Competition period: The submission portal will open March 1, 2026, and will close April 30, 2026.

Publication: Teams submitting to the competition will be invited to present at a competition workshop, scheduled coincident with a major conference TBA.

How to participate: Check for news and updates on the competition website. Use  this form to register.

Have questions? Contact us.

Data use agreement