AlterEgo continuous silent speech (Kapur et al., 2018)
Measured by Wadkins · MIT MEng thesis (2019)
Inputs
The measured or assumed values behind the calculations, each with its source.
- N = 20
- 20-word vocabulary in the 200-sentence continuous silent-speech dataset
- P = 0.893
- 1 - 10.7% word error rate for CNN with CTC plus language model on the same 20-word, 200-sentence dataset
- rate = 102.4 word/min
- Average speech rate for the same 20-word, 200-sentence dataset; the thesis reports 62% of samples at at least 100 WPM
- ITR_reported = 345.8 bits/min
- Author's reported Wolpaw bit-rate for the CNN+CTC model on the 20-word, 200-sentence dataset (10.7% WER, 102.4 WPM, N=20): 345.8 bits/min ≈ 5.76 bits/s. The compute below reproduces this figure.
Strictest ITR
Each scoring method is an upper bound on the channel, so the headline is the strictest (smallest) one for this entry. Use the score selector on the home page to view any single method across entries.
-
Bits per selection (Wolpaw formula)
B = log2(N) + P*log2(P) + (1-P)*log2((1-P)/(N-1)) = log2(20) + 0.893*log2(0.893) + 0.107*log2(0.107/19) = 3.377 bits / selection
Term 1 is the information if every choice were correct; terms 2-3 subtract the bits lost to the error rate, assumed spread evenly over the other N-1 targets.
-
Selections per second
T = 0.58594 s/selection -> 1 / 0.58594 = 1.707 selections/s
-
Information transfer rate
ITR = B * selections/s = 3.377 * 1.707 = 5.763 bits/s
What counts as a bit depends on the action space. The number of distinguishable actions and how likely each one is are design choices of the task, not the sensing hardware. The same modality can present a fixed set of targets, a set pruned per step by a grammar or language model, or a continuous control space. Each of these changes how many actions are live and how the probability mass is spread, and therefore the information per selection. Read the action space below before comparing headline numbers across entries.
Action space
What the user can produce at each step, and how those options are distributed.
- Structure
- Fixed set of targets
- Size
- 20 distinguishable actions
- Prior
- Non-uniform: some actions more likely than others
- Notes
- The reference task is Wadkins' 20-word, 200-sentence continuous silent-speech dataset, decoded from facial sEMG with CNN+CTC and a simple language model. This avoids mixing the IUI 2018 command-task accuracy with the follow-on thesis rate: N, WER, rate, and the reported bitrate all come from the same task. The action space is a 20-word closed vocabulary, and the sentence set is constrained, so the headline bits/s should not be read as open-vocabulary language throughput. Vocabulary size mainly changes the per-token comparison metric: log2(20) is only about 4.3 bits/word, while natural English content is not proportional to dictionary size.
Comparability The strictest bound here is the Wolpaw selection rate: log₂(N) over the real action set, discounted by accuracy. This fits because the output is a closed command set, not free English, so no language predictor applies. Comparable to the other command-and-control entries; against the text entries it reads high per selection, since a few large choices carry more raw bits than the ~1 bit/character of natural language.
Other score types
Bounds the atlas keeps out of the default strictest headline: as-reported figures, alternate task conditions, or raw-channel ceilings that shouldn't win the headline by default. Each still carries a score type, so the home-page selector ranks this entry on it when you choose that type. Read its derivation before comparing across entries.
-
Achieved-bitrate credit per net-correct word
N = 20 → log2(N − 1) = log2(19) = 4.25 bits per net-correct selection (field-standard achieved bitrate, e.g. Webgrid; Nuyujukian 2015, which introduced the metric, used log2(N)).
-
Net-correct word rate
net-correct = 2P − 1 = 2(0.893) − 1 = 0.786 of words. At 102.4 word/min (0.586 s/word) → 0.786 × 102.4 / 60 = 1.34 correct/s.
A word error commits the wrong word rather than timing out, so incorrect = 1 − P. Same N (20), word accuracy (89.3%, i.e. 10.7% WER) and word rate (102.4 wpm) as the entry's Wolpaw calc, which is the ranked figure here because the output is a 20-word closed vocabulary. Netting each wrong word against a correct one (2P − 1) lands just under the 5.76 bits/s Wolpaw ITR. Like it, this should not be read as open-vocabulary language throughput.
-
Achieved bitrate
4.25 bits × 1.34 correct/s = 5.70 bits/s.
Source
- Authors
- Wadkins
- Publication
- MIT MEng thesis, 2019