TongueBoard Intraoral Silent Speech (Li et al., 2019)

Measured by Li, Wu & Starner · Augmented Human (AH) 2019 (2019)

Electropalatography Silent speech 2019

Inputs

The measured or assumed values behind the calculations, each with its source.

N = 17: Choices in the user study: 15 non-vocalized words plus 2 tongue gestures, operating a calculator application.
P = 0.971: Recognition accuracy, held across stationary (desktop) and mobile (walking) contexts.
rate = 2.18 bits/s: Authors' reported information transfer rate (≈2.18 bits/s), held across stationary and mobile contexts: a real measured online rate. Equivalent to 3.78 bits/decision at ~0.58 decisions/s (≈1.73 s/decision).

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.

Strictest Wolpaw Author-reported · reproduced

Wolpaw bitrate over the 17-choice calculator set (authors' reported ITR)

Command-and-control selection: non-language output, so log2(N) over the real action set is the realized measure

2.19 bits/s

Bits per selection (Wolpaw formula)
```
B = log2(N) + P*log2(P) + (1-P)*log2((1-P)/(N-1))
  = log2(17) + 0.971*log2(0.971) + 0.029*log2(0.029/16)
  = 3.782 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 = 1.73 s/selection  ->  1 / 1.73 = 0.578 selections/s

Information transfer rate

ITR = B * selections/s = 3.782 * 0.578 = 2.186 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: 17 distinguishable actions
Prior: Uniform: all actions assumed equally likely
Notes: An intraoral retainer with capacitive touch sensors on the palate tracks absolute tongue position to recognize non-vocalized (silently mouthed) input. The user study operated a calculator with 15 non-vocalized words plus 2 tongue gestures (17 choices) at 97.1% accuracy, both stationary and while walking. Because the output is a closed command-and-control set rather than free English text, the realized measure is the selection bitrate: log2(N) over the real action set, discounted by accuracy (Wolpaw). This is exactly the non-language carve-out the methodology applies to command interfaces, and what the authors themselves report. Same lab lineage as the SilentSpeller electropalatography speller; folded under the silent-speech color and tagged Electropalatography for the capacitive-palate sensing. Unlike the offline EMG recognizers, this rate is measured online.

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.

Nuyujukian Recomputed

Nuyujukian achieved bitrate over the 17-choice calculator set

Achieved-bitrate view of the command set, shown for comparison

2.18 bits/s

Achieved-bitrate credit per net-correct selection

N = 17 → log2(N − 1) = log2(16) = 4.0 bits per net-correct selection (field-standard achieved bitrate, e.g. Webgrid; Nuyujukian 2015, which introduced the metric, used log2(N)).

Net-correct selection rate
```
net-correct = 2P − 1 = 2(0.971) − 1 = 0.942 of selections. At 1.73 s/decision → 0.942 / 1.73 = 0.545 correct/s.
```
A misrecognition commits the wrong command rather than timing out, so incorrect = 1 − P. Same N (17), recognition accuracy (97.1%) and decision interval (1.73 s) as the entry's Wolpaw calc, which is the ranked figure here because the output is a closed command set rather than English text. At this high accuracy the achieved view coincides with that Wolpaw ITR (~2.18 bits/s).

Achieved bitrate

4.0 bits × 0.545 correct/s = 2.18 bits/s.

Source

Authors: Li, Wu & Starner
Publication: Augmented Human (AH) 2019, 2019
Paper: 10.1145/3311823.3311831
Reference: Open-access PDF