https://neural-processing-lab.github.io/OVMI/

1<!doctype html>
2<html lang="en">
3<head>
4  <meta charset="utf-8">
5  <meta name="viewport" content="width=device-width, initial-scale=1">
6  <title>OVMI — A Common Measure of Communication for Speech Brain–Computer Interfaces</title>
7  <meta name="description" content="OVMI measures lexical information transfer from speech brain–computer interfaces relative to an explicit reference communication distribution.">
8  <meta name="theme-color" content="#f7f7f4">
9  <link rel="canonical" href="https://neural-processing-lab.github.io/OVMI/">
10  <link rel="icon" href="./assets/favicon.svg?v=2" type="image/svg+xml">
11  <link rel="stylesheet" href="./styles.css">
12  <meta property="og:title" content="A Common Measure of Communication for Speech Brain–Computer Interfaces">
13  <meta property="og:description" content="An interactive comparison of published speech-BCI systems under different reference communication distributions.">
14  <meta property="og:type" content="website">
15  <meta property="og:url" content="https://neural-processing-lab.github.io/OVMI/">
16  <meta name="twitter:card" content="summary">
17</head>
18<body>
19  <a class="skip-link" href="#main-content">Skip to content</a>
20
21  <header class="site-header">
22    <div class="site-shell header-inner">
23      <a class="wordmark" href="#top" aria-label="OVMI home">OVMI</a>
24      <nav aria-label="Primary navigation">
25        <a href="#leaderboard">Explorer</a>
26        <a href="#interpretation">How to interpret OVMI</a>
27        <a href="#why">Why OVMI?</a>
28        <a href="#use-ovmi">Use OVMI</a>
29      </nav>
30    </div>
31  </header>
32
33  <main id="main-content">
34    <section class="hero site-shell" id="top" aria-labelledby="hero-title">
35      <header class="paper-masthead">
36        <p class="eyebrow">Research paper · interactive explorer</p>
37        <h1 id="hero-title">A Common Measure of Communication for<br>Speech Brain–Computer Interfaces</h1>
38        <p class="paper-subtitle">An interactive comparison of published speech-BCI systems under different reference communication distributions.</p>
39        <p class="paper-authors" aria-label="Authors">
40          <span>Dulhan Jayalath</span>
41          <span>Benjamin Ballyk</span>
42          <span>Oiwi Parker Jones</span>
43        </p>
44        <p class="paper-affiliation"><a href="https://neural-processing-lab.github.io/" rel="noopener noreferrer">PNPL🍍, University of Oxford</a></p>
45        <div class="hero-actions" aria-label="Project links">
46          <a class="button button-primary" data-config-link="paper" href="https://arxiv.org/abs/2609.02887" rel="noopener noreferrer">
47            <svg class="button-icon" viewBox="0 0 20 20" aria-hidden="true"><path d="M5.5 2.5h6l3 3v12h-9z"/><path d="M11.5 2.5v3h3M8 9h4M8 12h4M8 15h3"/></svg>
48            <span>Paper</span>
49          </a>
50          <a class="button" data-config-link="package" href="https://github.com/neural-processing-lab/OVMI">
51            <svg class="button-icon" viewBox="0 0 20 20" aria-hidden="true"><rect x="2.5" y="3.5" width="15" height="13" rx="1"/><path d="m6.5 8 2 2-2 2M10.5 13h3"/></svg>
52            <span>Code + package</span>
53          </a>
54          <a class="button" data-config-link="blog" href="https://neural-processing-lab.github.io/blog/what-can-speech-brain-computer-interfaces-communicate/" rel="noopener noreferrer">
55            <svg class="button-icon" viewBox="0 0 20 20" aria-hidden="true"><path d="M4 3.5h12v13H4zM7 7h6M7 10h6M7 13h4"/></svg>
56            <span>Blog</span>
57          </a>
58          <a class="button" href="#leaderboard">
59            <svg class="button-icon" viewBox="0 0 20 20" aria-hidden="true"><path d="M3.5 16.5h13M5.5 14V9.5h2V14zM9 14V5.5h2V14zM12.5 14V7.5h2V14z"/></svg>
60            <span>Explorer</span>
61          </a>
62        </div>
63      </header>
64
65    </section>
66
67    <section class="section section-rule site-shell" id="leaderboard" aria-labelledby="leaderboard-title">
68      <p class="benchmark-intro">Speech-BCI studies use different vocabularies, datasets, and communication domains, making reported scores hard to compare. Open-Vocabulary Mutual Information (OVMI) evaluates each system against an explicit, common reference communication distribution, so otherwise heterogeneous systems can be placed on the same communication scale. Choose a communication target below to compare published systems.</p>
69      <div class="section-heading split-heading">
70        <div>
71          <p class="section-index">01 / Explorer</p>
72          <h2 id="leaderboard-title">OVMI Explorer</h2>
73        </div>
74      </div>
75
76      <div class="reference-panel">
77        <fieldset class="reference-selector" id="reference-selector">
78          <legend>Reference communication distribution</legend>
79          <label><input type="radio" name="reference" value="subtlex" checked> <span>Broad speech<br><strong>SUBTLEX-UK</strong></span></label>
80          <label><input type="radio" name="reference" value="conversation"> <span>Conversation<br><strong>Switchboard</strong></span></label>
81          <label><input type="radio" name="reference" value="ucv"> <span>AAC / clinical<br><strong>UCV</strong></span></label>
82          <label><input type="radio" name="reference" value="narrative"> <span>Narrative<br><strong>Sherlock</strong></span></label>
83        </fieldset>
84        <div class="reference-readout" aria-live="polite">

85          <span>Selected entropy</span>
86          <output id="reference-entropy">9.77 bits</output>
87          <p id="reference-description">Broad spoken-English frequency norm derived from British film and television subtitles.</p>
88        </div>
89      </div>
90
91      <p class="explanation">A vocabulary covers different amounts of language under different distributions. Select a reference to update the scores.</p>
92      <p class="example-callout"><strong>Example.</strong> The 50-word Willett +LM system carries <span id="willett-broad">6.4%</span> of the lexical information in broad spoken English and <span id="willett-ucv">40.4%</span> under the AAC reference.</p>
93
94      <section class="benchmark-overview" aria-labelledby="benchmark-overview-title">
95        <div class="benchmark-overview-heading">
96          <div>
97            <p class="section-index">At a glance</p>
98            <h3 id="benchmark-overview-title">Published systems on the selected communication scale</h3>
99          </div>
100          <p>The vertical position is OVMI / H(p); the horizontal position is publication year. Each panel uses its own y-axis scale.</p>
101        </div>
102        <div class="scatter-legend" aria-label="Study group legend">
103          <span><i class="legend-swatch invasive-swatch"></i>Attempted speech / invasive</span>
104          <span><i class="legend-swatch noninvasive-swatch"></i>Perceived speech / non-invasive</span>
105        </div>
106        <div id="benchmark-scatter" class="benchmark-scatter" aria-live="polite"></div>
107      </section>
108
109      <div class="table-controls">
110        <fieldset class="group-toggle" id="group-filter">
111          <legend>Study group</legend>
112          <label><input type="radio" name="group" value="attempted_invasive" checked>Invasive</label>
113          <label><input type="radio" name="group" value="perceived_noninvasive">Non-invasive</label>
114        </fieldset>
115        <label for="sort-select">Sort within groups
116          <select id="sort-select">
117            <option value="ovmi-desc">OVMI, high to low</option>
118            <option value="percent-desc">OVMI / H(p), high to low</option>
119            <option value="year-desc">Year, newest first</option>
120            <option value="year-asc">Year, oldest first</option>
121            <option value="vocabulary-desc">Vocabulary size, high to low</option>
122          </select>
123        </label>
124      </div>
125
126      <div class="table-scroll" tabindex="0" aria-label="Scrollable system comparison">
127        <table class="leaderboard-table">
128          <caption>Speech-BCI operating points evaluated using the selected communication reference.</caption>
129          <thead>
130            <tr>
131              <th scope="col">System</th>
132              <th scope="col">Year</th>
133              <th scope="col">Setting / modality</th>
134              <th scope="col">Vocabulary</th>
135              <th scope="col">Reported metric</th>
136              <th scope="col">Comparison <span id="bar-metric-label" class="unit">OVMI bits</span></th>
137              <th scope="col">OVMI <span class="unit">bits</span></th>
138              <th scope="col">OVMI / H(p)</th>
139              <th scope="col">Uncertainty</th>
140              <th scope="col">Source</th>
141            </tr>
142          </thead>
143          <tbody id="leaderboard-body">
144            <tr><td colspan="10" class="loading-cell">Loading comparison data…</td></tr>
145          </tbody>
146        </table>
147      </div>
148      <noscript><p class="error-note">JavaScript is required for reference switching and sorting. The generated data remain available in <a href="./data/leaderboard.json">the comparison data file</a>.</p></noscript>
149
150      <div class="leaderboard-notes">
151        <aside id="methodology">
152          <h3>How are scores calculated?</h3>
153          <p>Historical results are estimated from reported accuracy or WER where full decoder channels are unavailable. <a data-config-link="paper" href="https://arxiv.org/abs/2609.02887" rel="noopener noreferrer">See the paper</a> for assumptions and uncertainty estimation.</p>
154        </aside>
155        <aside id="add-result">
156          <h3>Add a result</h3>
157          <p>Submit a paper citation, decoder vocabulary, reference distribution, and reported metric through a GitHub issue or pull request.</p>
158          <a class="text-link" data-config-link="addResult" href="https://github.com/neural-processing-lab/OVMI/issues/new?template=add-ovmi-result.yml">Submit a result →</a>
159        </aside>
160      </div>
161    </section>
162
163    <section class="section section-rule site-shell interpretation-note" id="interpretation" aria-labelledby="interpretation-title">
164      <p class="section-index">02 / Interpretation</p>
165      <h2 id="interpretation-title">How to interpret the comparison</h2>
166      <p>OVMI puts systems on a common communication scale, but it does not make their underlying experiments equivalent. Studies differ in task, speech type, participants, recording modality, data quantity, and language-model use. A higher OVMI therefore means that a reported system conveys more lexical information relative to the selected reference distribution; it does not mean that its task or recording paradigm is more useful, practical, or clinically relevant. For example, strong performance on perceived speech may reflect good lexical decoding without corresponding to a practical communication interface.</p>
167    </section>
168
169    <section class="section section-rule site-shell why-grid" id="why" aria-labelledby="why-title">
170      <div>
171        <p class="section-index">03 / Definition</p>
172        <h2 id="why-title">Why OVMI?</h2>
173        <p>Speech-BCI studies use different vocabularies and communication domains, making their reported scores difficult to compare. OVMI evaluates each system against a common reference communication distribution.</p>
174        <p>OVMI multiplies the speech BCI's vocabulary coverage of the reference distribution by its in-vocabulary information transfer.</p>
175        <div class="formula" role="math" aria-label="OVMI equals lexical coverage times in-vocabulary information">

176          <span>I<sub>OVMI</sub>(S)</span>
177          <span class="formula-equals">=</span>
178          <span>C(S)</span>
179          <span class="formula-times">×</span>
180          <span>I(X;Y | X∈S)</span>
181        </div>
182        <dl class="definitions">
183          <div><dt>S</dt><dd>supported decoder vocabulary</dd></div>
184          <div><dt>p</dt><dd>reference communication distribution</dd></div>
185          <div><dt>C(S)</dt><dd>probability that a word drawn from p lies in S</dd></div>
186          <div class="definition-long"><dt>I(X;Y | X∈S)</dt><dd>information transferred about the intended word, given that it lies in S</dd></div>
187        </dl>
188      </div>
189      <aside class="symbol-example">
190        <h3>Why the target matters</h3>
191        <p>The useful vocabulary depends on the application. A spelling system needs broad lexical coverage; a command or AAC system may focus on a smaller set of task-specific words.</p>
192        <p>The same decoding accuracy can therefore support different amounts of communication, depending on which words the user needs.</p>
193      </aside>
194    </section>
195
196    <section class="section section-rule site-shell use-grid" id="use-ovmi" aria-labelledby="use-title">
197      <div>
198        <p class="section-index">04 / Reproduce</p>
199        <h2 id="use-title">Use OVMI</h2>
200        <p>Install from GitHub:</p>
201        <pre><code>pip install git+https://github.com/neural-processing-lab/OVMI.git</code></pre>
202        <p>Pass a reference distribution, supported vocabulary, and macro accuracy:</p>
203        <pre><code>from ovmi import ovmi
204
205reference = {"yes": 120, "no": 80, "water": 20}
206vocabulary = ["yes", "no"]
207
208score = ovmi(reference, vocabulary, accuracy=0.70)
209print(score)</code></pre>
210        <p><a class="text-link" data-config-link="repository" href="https://github.com/neural-processing-lab/OVMI">Package documentation and source →</a></p>
211      </div>
212
213      <div class="citation-panel" id="citation">
214        <div>
215          <p class="section-index">Citation</p>
216          <h3>Cite this paper</h3>
217          <p>Use the following BibTeX entry to cite the arXiv preprint.</p>
218        </div>
219        <pre><code>@article{jayalath2026ovmi,
220  title={A Common Measure of Communication for Speech Brain-Computer Interfaces},
221  author={Jayalath, Dulhan and Ballyk, Benjamin and Parker Jones, Oiwi},
222  journal={arXiv preprint arXiv:2609.02887},
223  year={2026}
224}</code></pre>
225      </div>
226    </section>
227
228    <section class="section section-rule site-shell project-links" id="project-links" aria-labelledby="links-title">
229      <p class="section-index">05 / Project</p>
230      <h2 id="links-title">Paper and project links</h2>
231      <div class="link-status-grid">
232        <p id="paper-status"><strong>Paper.</strong> <a href="https://arxiv.org/abs/2609.02887" rel="noopener noreferrer">Read the arXiv preprint.</a></p>
233        <p id="blog-status"><strong>Blog.</strong> <a data-config-link="blog" href="https://neural-processing-lab.github.io/blog/what-can-speech-brain-computer-interfaces-communicate/" rel="noopener noreferrer">Read the blog post.</a></p>
234      </div>
235    </section>
236  </main>
237
238  <footer class="site-footer">
239    <div class="site-shell footer-inner">
240      <p><strong>A Common Measure of Communication for Speech Brain–Computer Interfaces</strong><br>Dulhan Jayalath, Benjamin Ballyk, and Oiwi Parker Jones</p>
241      <p><a data-config-link="repository" href="https://github.com/neural-processing-lab/OVMI">Source and data</a> · MIT License</p>
242    </div>
243  </footer>
244
245  

245<script src="./config.js"></script>

245
246  

246<script src="./app.js"></script>

246
247</body>
248</html>