A Common Measure of Communication for Speech Brain-Computer Interfaces
The paper proposes OVMI, a shared information-theoretic scale for judging how much speech a BCI can actually communicate.
The authors argue that current speech BCI scores are hard to compare because systems differ in datasets, recordings, speech types, and vocabularies. OVMI measures decoder output against a reference distribution of words a user may want to say, so limited-vocabulary systems are not credited as if they cover all intended speech. They show that accuracy and WER can overstate real communication when measured only on supported words. The paper also reports that OVMI-guided vocabulary selection improved accuracy by up to 16.3% across three speech domains. HF Daily Papers' note
The authors argue that current speech BCI scores are hard to compare because systems differ in datasets, recordings, speech types, and vocabularies. OVMI measures decoder output against a reference distribution of words a user may want to say, so limited-vocabulary systems are not credited as if they cover all intended speech. They show that accuracy and WER can overstate real communication when measured only on supported words. The paper also reports that OVMI-guided vocabulary selection improved accuracy by up to 16.3% across three speech domains. HF Daily Papers' note
score 4