NeuroGenesysHow fast does your BCI communicate?
Information transfer rate (ITR) is the standard speed measure for brain-computer interfaces: how many bits a user can send per minute, from the number of targets, the accuracy and the time each selection takes. Enter yours, see bits per minute and the equivalent typing speed, and check that your accuracy is really above chance.
To size the raw signal behind it, use the neural data rate calculator.
Free · runs in your browser · updated
1 · Your BCI
2 · Information transfer rate
Characters per minute if every bit went into text. 4.70bits is one of 26 equally likely letters (no language model); about 1 bit per letter is Shannon's 1951 estimate for English read with full context (0.6–1.3 bits), roughly what a strong language model can exploit.
3 · Published results for comparison
- P300 speller (EEG)≈5 letters/min (≈0.5 bits/s)The classic 6×6 matrix speller of Farwell & Donchin (1988), as summarised by Chen et al. 2015. Chen et al., PNAS 2015
- SSVEP speller (EEG)up to 5.32 bits/s (≈319 bits/min), ≈60 characters/min40-target joint frequency-phase coded speller, cued spelling, peak online performance. Chen et al., PNAS 2015
- Handwriting BCI (intracortical)90 characters/minAttempted handwriting decoded to text by one participant with paralysis; copy-typing plateau. Willett et al., Nature 2021
- Speech BCI (intracortical)62 words/minAttempted speech decoded at 9.1% word error on a 50-word vocabulary and 23.8% on 125,000 words; one participant with ALS. Willett et al., Nature 2023
- Speech BCI (intracortical)≈32 words/min, 97.5% accuracySelf-paced conversation with a 125,000-word vocabulary, sustained over 8.4 months; one participant with ALS. Card et al., NEJM 2024
Speech and handwriting BCIs report words or characters per minute with a language model in the loop, so their rates are not Wolpaw ITRs and can't be compared bit for bit. The intracortical results each come from one or a few participants.
4 · What the formula assumes
- Every selection takes the same time, and the pause between selections counts. Leaving it out is the most common way ITRs get inflated.
- All targets are equally likely and errors are spread evenly over the wrong ones. Real confusions cluster on neighbours, so the true information can differ.
- Errors cost nothing. In a speller, each error also costs a backspace and a retype; practical rates are lower.
- Accuracy must be measured online, on enough selections, and significantly above chance. Yuan et al. (J. Neural Eng. 2013) found ITR frequently misreported for these reasons.
Questions
How do you calculate BCI information transfer rate?
With Wolpaw's formula: bits per selection B = log2 N + P·log2 P + (1 − P)·log2((1 − P)/(N − 1)), where N is the number of targets and P the accuracy. Multiply by selections per minute, counting the pause between selections, to get bits per minute.
What is a good ITR for a BCI?
Classic P300 spellers manage roughly 5 letters per minute (about 30 bits per minute). The fastest published EEG spellers, using SSVEP, reached about 5.3 bits per second (around 320 bits per minute) in cued spelling. Intracortical speech BCIs now decode 32–62 words per minute, though those rates come with language models and aren't Wolpaw ITRs.
Why is my ITR zero?
Accuracy at or below chance (1/N) carries no information, and Wolpaw's convention sets the ITR to zero there. Check too that your accuracy is significantly above chance for the number of selections you tested; with few trials, even 70% on a 2-class task can be luck.
Should the pause between selections count?
Yes. ITR is a rate, so gaze shifts, feedback and rest periods between selections are part of the time per selection. Leaving them out is the most common way published ITRs are inflated.
What are the limits of the Wolpaw ITR?
It assumes equal selection times, equally likely targets and errors spread evenly over wrong targets, and it ignores the cost of correcting errors. Yuan et al. (2013) found ITR is frequently misreported in BCI papers, so report the accuracy, timing and number of trials alongside it.
