Table 1 Mean phoneme and word error rates (with 95% CIs) for the speech BCI across all evaluation days

From: A high-performance speech neuroprosthesis

 

Phoneme error rate, % (95% CI)

Word error rate, % (95% CI)

Online

125,000-word, vocal

19.7 (18.6, 20.9)

23.8 (21.8, 25.9)

125,000-word, silent

20.9 (19.3, 22.6)

24.7 (22.0, 27.4)

50-word, vocal

21.4 (19.6, 23.2)

9.1 (7.2, 11.2)

50-word, silent

22.1 (19.9, 24.3)

11.2 (8.3, 14.4)

Offline

125,000-word, improved LM

19.7 (18.6, 20.9)

17.4 (15.4, 19.5)

125,000-word, improved LM + proximal test set

17.0 (15.7, 18.3)

11.8 (9.8, 13.9)

  1. Phoneme error rates assess the quality of the RNN decoder’s output before a language model is applied, whereas word error rates assess the quality of the combined RNN and language model (LM) pipeline. CIs were computed with the bootstrap percentile method (resampling over trials 10,000 times). Online refers to what was decoded in real time whereas offline refers to post hoc analysis of data using an improved language model (improved LM) or different partitioning of training and testing data (proximal test set). In the proximal test set, training sentences occur much closer in time to testing sentences, mitigating the effect of within-day neural non-stationarities.
Back to article page