<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Noormal Wardak on Medium]]></title>
        <description><![CDATA[Stories by Noormal Wardak on Medium]]></description>
        <link>https://medium.com/@noormalwardak?source=rss-57e66c720c38------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*p46P50cqtfEoeF7g3xj8aQ.png</url>
            <title>Stories by Noormal Wardak on Medium</title>
            <link>https://medium.com/@noormalwardak?source=rss-57e66c720c38------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 07:26:35 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@noormalwardak/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Hearing the Words You Never Say, and Controlling Your Devices With Them]]></title>
            <link>https://medium.com/@noormalwardak/hearing-the-words-you-never-say-and-controlling-your-devices-with-them-90b87d390dc9?source=rss-57e66c720c38------2</link>
            <guid isPermaLink="false">https://medium.com/p/90b87d390dc9</guid>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[electronics]]></category>
            <category><![CDATA[wearables]]></category>
            <category><![CDATA[brain-computer-interface]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Noormal Wardak]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 18:34:09 GMT</pubDate>
            <atom:updated>2026-10-22T19:16:50.942Z</atom:updated>
            <content:encoded><![CDATA[<h3>Speaking Without Sound, Part 2: building a headset that lets you control your laptop and phone without making a sound.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KYbkpfnehLAabH9OCkV_lw.png" /></figure><h3>Imagine this</h3><p>Imagine telling your laptop to open your notes, texting a friend, or asking your phone a question, without moving your lips or making a sound. You just say the words silently, inside your mouth and throat, and the device hears them. It answers through a tiny speaker that sits on the bone behind your ear, so only you hear the reply. Or it turns your silent words into a spoken voice on a call, while the people sitting next to you hear nothing at all.</p><p>You could work in a library, on a plane, or in the middle of a meeting, and talk to your computer the whole time. <strong>No keyboard, no screen, no one overhearing.</strong></p><p>This isn’t mind reading. <strong>It works because even when you say a word silently, your brain still sends tiny electrical signals to the muscles of your chin, jaw and throat,</strong> and electrodes on the skin can pick them up. MIT’s AlterEgo headset showed this in 2018. Two years later, researchers at UC Berkeley turned silently mouthed words into a synthesised voice.</p><p>But a lab demo isn’t something you’d wear every day. For that, it has to keep working tomorrow, after you take it off and put it back on. That’s the question this build is designed to answer, starting small: ten digits, four electrodes, and one amplifier I’m building from scratch.</p><h3>From article to workbench</h3><p>In <a href="https://medium.com/@noormalwardak/speaking-without-sound-how-silent-speech-technology-works-and-what-it-could-do-next-168c3e576e32">Part 1</a>, I followed one silent word from the brain to the speech muscles to a neural network, and explained how MIT’s AlterEgo headset reads the faint electricity your face makes when you say words to yourself. I ended on a question no paper fully answers: <strong>does it still work tomorrow, after you take the electrodes off and put them back on?</strong></p><p>Then I hit a wall. To answer it I needed hardware, and research-grade biosignal boards cost money I didn’t have. So I asked a TKS director how other students get one. His answer changed the whole plan: don’t buy it yet. Build the amplifier yourself, from an amplifier chip, a breadboard and a few electrodes, and understand the signal from first principles. Earn the fancy board later.</p><p>That’s what this article is. Part 2 is the build: how a homemade muscle-signal amplifier works, what I’m buying and where, how the software turns voltage into words, and exactly how I’ll run the two-day test.</p><p>There are no results here yet, on purpose. <strong>I’m writing down the plan and my predictions before I collect a single recording, so I can’t quietly move the goalposts later.</strong> Part 3 will show whether I was right.</p><h3>The short version</h3><p>I’m building a four-channel muscle-signal amplifier from scratch, sticking its electrodes on my chin, under my jaw, on my throat and on my cheek, and recording myself silently saying the digits 0 to 9. I’ll record on day one, take everything off, put it back on day two, and record again. <strong>The build exists to produce four numbers: how accurate it is for mouthed and silent speech, on the same day and on the next day, plus how much a short recalibration wins back.</strong></p><p><strong>How this article is organised</strong></p><ol><li><strong>Part 1: The question.</strong> What the best recent study found, and the gap I’m measuring.</li><li><strong>Part 2: The hardware.</strong> How a homemade amplifier pulls a signal smaller than mains hum out of the noise, what I’m buying in the UAE, and where the electrodes go.</li><li><strong>Part 3: The software.</strong> From raw voltage to a predicted digit.</li><li><strong>Part 4: The experiment.</strong> The two-day test, my timeline, and what could go wrong.</li></ol><h3>Part 1: The question</h3><p>The clearest numbers on the day-two problem come from <a href="https://arxiv.org/html/2603.02847v2">SilentWear</a>, a 2026 study of a low-power silent speech wearable. Four people wore 14 EMG channels and silently said 8 commands like “up”, “left” and “stop”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*heroRaUQicZ5nnuD6___6A.png" /><figcaption><em>SilentWear’s neckband: 27 dry electrodes sewn into fabric, read as 14 channels by a board small enough to run the neural network itself. Mine sits at the other end of the scale: four gel electrodes and a breadboard. Figure reproduced from </em><a href="https://arxiv.org/abs/2603.02847"><em>Spacone et al., SilentWear</em></a><em>, arXiv 2026, Fig. 1.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LfOaXAcpa2vwzVMiPsGitA.png" /><figcaption><em>“Sessions mixed” means the training and test recordings were pooled from every session. Chart by the author; data from </em><a href="https://arxiv.org/abs/2603.02847"><em>SilentWear</em></a><em>, 2026.</em></figcaption></figure><p>Two things jumped out at me. First, <strong>putting the electrodes back on cost almost 20 points of accuracy</strong>, even in a lab, even with 14 channels. Second, a short recalibration won most of it back. That’s hopeful, but nobody wants to spend ten minutes retraining a device every morning.</p><p>And one more: <strong>silent speech fell further than speech said out loud</strong>, which only dropped from 84.8% to 71.1%. The weaker the movement, the more a small shift in the electrodes seems to hurt. That’s the same pattern I’m predicting for silent versus mouthed speech.</p><p>The original AlterEgo team knew this. Their <a href="https://dl.acm.org/doi/10.1145/3172944.3172977">2018 paper</a> says it’s “imperative for the electrodes to not shift position”. So my question from Part 1 stays exactly the same:</p><p><strong>How much does silent speech accuracy drop when articulation gets weaker, and when the electrodes are taken off and put back on another day?</strong></p><p><a href="https://medium.com/@noormalwardak/speaking-without-sound-how-silent-speech-technology-works-and-what-it-could-do-next-168c3e576e32?postPublishedType=repub">My hypotheses are in Part 1.</a> Here’s how I’ll test them.</p><h3>Part 2: The hardware</h3><h3>The problem: the signal is smaller than the noise</h3><p>When a muscle fires, the voltage that reaches your skin is tiny: <a href="https://arxiv.org/pdf/2101.04658">“only a few microvolts to millivolts”</a>, with most of its energy between about 20 and 150 Hz. Meanwhile every wall socket in the room is radiating 50 Hz hum (Dubai’s grid runs at 50 Hz), and your body picks it up like an antenna. At the skin, the hum can easily be bigger than the muscle signal you want.</p><p>So the whole circuit is one idea, repeated: <strong>keep the muscle signal, throw away everything else.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WoiPZGZ_0NDN2gCpXbGV7Q.png" /><figcaption><em>Figure 1. From a twitch to a number.</em> <strong>Key insight:</strong> the hum lands on both electrodes almost equally, so an amplifier that only boosts their difference deletes most of it in one step. Diagram by the author; the waveforms are illustrative sketches, not recordings.</figcaption></figure><h3>The chip that does the magic</h3><p>Step 2 is an instrumentation amplifier, the <a href="https://www.ti.com/product/INA128">INA128</a> from Texas Instruments. It’s an 8-pin chip that amplifies the <em>difference</em> between its two inputs and ignores what they have in common. Its common-mode rejection is at least 120 dB, which means a signal that hits both electrodes equally comes out about a million times weaker than a signal that differs between them. One resistor sets the gain, and it runs from ±2.25 V to ±18 V, so two 9 V batteries can power it.</p><p>One thing surprised me: <strong>the first stage has to keep its gain low, around 10.</strong> Electrodes on skin carry a small, steady voltage offset of their own. Amplify that by 1,000 straight away and the output slams into the battery voltage and flatlines. So I amplify a little, filter out the steady part with the high-pass filter in step 3, and only then boost the signal by about 100 in step 4.</p><h3>Why I’m recording a wider band than I need</h3><p>Here’s a detail I only found by reading the papers side by side. The two systems I’m comparing myself to disagree about which frequencies matter:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RB43PH4GRNgCCWq15roj0g.png" /><figcaption><em>Chart by the author; data from </em><a href="https://doi.org/10.1145/3172944.3172977"><em>AlterEgo</em></a><em>, </em><a href="https://arxiv.org/html/2603.02847v2"><em>SilentWear</em></a><em> and the </em><a href="https://arxiv.org/pdf/2101.04658"><em>Zheng et al. review</em></a><em>.</em></figcaption></figure><p>AlterEgo threw away almost everything above 50 Hz, while classic muscle recording keeps exactly that range. So my hardware records wide, from about 1 to 450 Hz at 1,000 samples a second, and <strong>I’ll test both bands in software on the same recordings.</strong> That’s a free mini experiment: does the slow part or the fast part of the signal carry the silent words?</p><p>Four electrode sites means four copies of steps 2 to 6, one per channel, all feeding the same Arduino.</p><h3>What I’m buying, and where</h3><p>The whole build costs roughly AED 300 to 400 by my estimate, a fraction of a research board. <strong>Nothing on this list is exotic, which is the point: anyone could repeat this.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UtYZh-eVDb71sb_tWi-qsg.png" /><figcaption><em>The insurance board is the </em><a href="https://www.tindie.com/products/upsidedownlabs/bioamp-exg-pill-sensor-for-ecg-emg-eog-or-eeg/"><em>BioAmp EXG Pill</em></a><em> on Tindie. List by the author; prices change, so check before buying.</em></figcaption></figure><p>The last item is my insurance. When my homemade channel shows nothing, I’ll plug the same electrodes into a board I know works. If it shows a signal, the bug is in my circuit. If it doesn’t, the problem is my skin contact. That one test saves days of guessing.</p><p>For soldering, tools and help, I’ll use <a href="https://www.ha.ae/5/programs/fab-lab-uae">Fab Lab UAE</a>, run by the Hamdan Foundation in Dubai.</p><h3>Where the electrodes go</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tC6M6wFp9lZNbwPLU1VrJA.png" /><figcaption><em>Figure 2. Where the electrodes go on v0, the build I’m doing now.</em> <strong>Key insight:</strong> the cheek is a control, a spot I don’t expect to matter much, so if it carries as much information as the chin, something in my setup is wrong. 3D render by the author in Blender</figcaption></figure><p>The four sites come straight from Part 1: the chin and under the jaw ranked near the top in MIT’s search for the best spots, and the throat sits over the voice box muscles. <strong>The reference goes behind the ear, on bone, because bone has almost no muscle of its own,</strong> so it gives the amplifier a quiet baseline to compare every channel against. Berkeley’s dataset put its reference behind the ear too.</p><p>Each channel gets its own wire colour, matching the figure, so a swapped lead shows up in the placement photo before it ruins a day of data.</p><h3>Part 3: The software</h3><p>The hardware ends with a stream of numbers. The software’s job is to turn that stream into a digit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*huRyLvnE4DlrTd55S9NKXA.png" /><figcaption><em>An oscilloscope draws voltage over time. I don’t have one, so a live plot on my laptop does the same job.</em></figcaption></figure><h3>Step 0: practise on someone else’s data</h3><p>Before my own circuit works, I’ll train on real recordings. In 2020, David Gaddy and Dan Klein at UC Berkeley <a href="https://aclanthology.org/2020.emnlp-main.445/">released a dataset</a> of nearly 20 hours of facial EMG from one speaker, recorded on 8 channels at 1,000 samples a second, with electrodes on the cheeks, chin, jaw and throat. It includes both silent and spoken speech. <strong>If my code can’t learn from clean lab data, it definitely won’t learn from my breadboard.</strong></p><h3>Step 1: record with labels</h3><p>A small Python script shows a random digit on screen, beeps, and saves the next 1.5 seconds from all four channels with that digit as the label. Random order matters: if I said all the zeros first and all the nines last, a model could learn the <em>time</em> instead of the word, because electrodes drift as the gel warms up.</p><h3>Step 2: clean the signal</h3><ol><li>Notch out 50 Hz and its multiples (100, 150 Hz and so on), the leftover hum.</li><li>Make two copies of every recording: one band-passed like AlterEgo (1.3 to 50 Hz), one like classic muscle recording (20 to 450 Hz).</li></ol><h3>Step 3: turn each recording into features</h3><p>I’ll start dumb and get smarter:</p><ol><li><strong>Simple:</strong> chop each recording into short windows and measure the signal strength (RMS) on each channel. That gives a small table of numbers per word. A basic classifier like logistic regression learns from it.</li><li><strong>AlterEgo’s way:</strong> MFCC features, the same trick speech recognition uses for audio, fed into a small 1D convolutional neural network.</li></ol><p>Starting simple tells me whether the fancy model is actually adding anything.</p><h3>Step 4: score it honestly</h3><p>With 10 digits, random guessing scores 10%, so that’s the floor. I’ll report accuracy plus a confusion matrix, a grid that shows <em>which</em> digits get mixed up. If “five” and “nine” get confused but “zero” never does, that tells me something about which muscles each word uses.</p><h3>Part 4: The experiment</h3><h3>The two-day test</h3><p>My test borrows its shape from SilentWear: separate days, every command repeated 20 times, spoken and silent recordings alternating, all in random order.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VJ5oQoUyPuomgPL560FiIA.png" /><figcaption><em>SilentWear’s protocol: three sessions on three days, each with 10 batches alternating spoken and silent, and 20 repetitions of each of 8 commands per batch. Figure reproduced from </em><a href="https://arxiv.org/abs/2603.02847"><em>Spacone et al., SilentWear</em></a><em>, arXiv 2026, Fig. 2.</em></figcaption></figure><p>Mine is smaller: two days, ten digits, four channels, and mouthed instead of spoken out loud.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OuHWxErAp2um3Z2L0X0FLQ.png" /><figcaption><em>Figure 3. The two-day test.</em> <strong>Key insight:</strong> only two things change between cells, how strongly I articulate and where the electrodes sit, so any drop in accuracy can be blamed on one of them. Design by the author; the predictions are untested guesses.</figcaption></figure><p>Here’s exactly what happens on each day:</p><ol><li><strong>Prepare the skin the same way.</strong> Clean each spot with an alcohol wipe, wait for it to dry, apply fresh gel electrodes.</li><li><strong>Measure where every electrode landed.</strong> Distance from the tip of the chin, the angle of the jaw and the bump of the voice box, plus a photo with a ruler in frame. If day two goes badly, I’ll know how many millimetres each electrode moved.</li><li><strong>Record 20 of each digit, mouthed, then 20 of each digit, silent.</strong> That’s 200 recordings per mode per day, in random order, with short breaks so my jaw doesn’t get tired.</li><li><strong>Same time of day.</strong> Sweat and skin change through the day, and I only want the electrodes to change.</li><li><strong>Day two:</strong> after day one everything comes off, and on day two I repeat steps 1 to 4 from scratch.</li></ol><p><strong>Every one of my guesses is in Figure 3, written down before I have any data.</strong> If I’m wrong, that’s a result too.</p><h3>Three extra questions, for free</h3><p>The same 800 recordings can answer more than the four headline numbers:</p><ol><li><strong>Recalibration.</strong> Add 5 day-two recordings per digit to the training data, then test on the other 15. SilentWear needed about 10 minutes of fresh data to win back most of the loss. How little can I get away with?</li><li><strong>Which electrode matters most.</strong> Retrain the model with one channel removed at a time. The channel whose removal hurts most is carrying the most information. My guess, from Part 1: the chin and under the jaw.</li><li><strong>Slow or fast signals.</strong> Run everything twice, once with AlterEgo’s 1.3 to 50 Hz band and once with the classic 20 to 450 Hz band.</li></ol><p>All the raw recordings and placement photos will be published openly with Part 3, so anyone can check my numbers or beat them.</p><h3>The timeline</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BadaV4GIVbAQGyiEw0guIg.png" /><figcaption><em>Figure 4. Six weeks, one milestone each.</em> <strong>Key insight:</strong> every week ends with something I can show, so if the plan is failing I find out in week 2, not week 6. Plan by the author.</figcaption></figure><p>Week 3 is my favourite. With more gain, the same amplifier can pick up alpha waves, the rhythm of roughly 8 to 12 cycles a second that grows when you close your eyes. <strong>That makes the build a real brain-computer interface before it’s a silent speech one,</strong> and proves the electronics work on a much smaller signal than the face will give me.</p><h3>Safety, and what could go wrong</h3><p><strong>One rule I’m not breaking: electrodes on my face never share a circuit with mains power.</strong> The board runs on two 9 V batteries, and the laptop stays unplugged from the wall while I record. This is a learning project, not a medical device.</p><p>What could go wrong, and my plan for each:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6k6MrqlWj6fESYBnoUIosQ.png" /><figcaption><em>Plan by the author.</em></figcaption></figure><h3>What comes next: Mini AlterEgo v1</h3><p>Figure 2 is v0, the build I’m actually doing: stick-on electrodes and a breadboard. It answers the question, but nobody would ever wear it <strong>If the two-day test works, and if I can get funding, this is the next build:</strong> the same four channels on the same electrode sites, moved into a headset that clips behind the ear.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LFhrzsrORr10JCDJGo6nZQ.png" /><figcaption><em>Mini AlterEgo v1, a concept for the next build. 3D render by the author in Blender; head scan: Lee Perry-Smith by Infinite-Realities, CC BY 3.0.</em></figcaption></figure><p>Two upgrades matter most. <strong>A ring around each electrode glows when skin contact is good,</strong> so on day two I’d know straight away whether a bad result is the model’s fault or the electrodes’. And <strong>a bone conduction speaker in front of the ear lets the device answer back</strong> without anyone else hearing it, which is the “talk to your laptop in silence” idea from the start of this article.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BX-zr9vMjsAdPUr3ZEM0Fw.png" /><figcaption><em>What’s inside v1, pulled apart. Parts shown at different scales. 3D concept by the author.</em></figcaption></figure><p>None of this exists yet. Whether it gets built depends on what Part 3 shows, and on funding, which brings me to the last section.</p><h3>About me</h3><p>Hey, I’m Noormal Wardak, a student in Dubai and an innovator at TKS, heading toward neuroscience at university. I also build voice AI at Convoscale, and I put what it earns back into iPads and internet data so girls in Afghanistan can keep studying.</p><p>If you’ve built an EMG amplifier, worked on silent speech, or just think my plan has a hole in it, I’d genuinely love to hear from you. Tell me before week 5, while I can still fix it.</p><p><strong>And if you can help fund this, that would be amazing.</strong> I’m paying for this build myself, so every bit goes straight into parts. Sponsoring electrodes and chips, lending equipment, or funding a research-grade board so I can test more people than just me would take this project much further, much faster. Message me here on Medium or on LinkedIn, and I’ll share every result with you first or email me!! would love to hear your thoughts!!</p><p>noormal.wardakak@gmail.com <br>https://www.linkedin.com/in/noormal-wardak/</p><p><em>Clap if you want to see whether my guesses survive. Part 3, the results, is next.</em></p><h3>Sources</h3><p><strong>Research</strong></p><ul><li>Kapur, Kapur and Maes, “AlterEgo: A Personalized Wearable Silent Speech Interface”, IUI 2018: <a href="https://doi.org/10.1145/3172944.3172977">doi.org/10.1145/3172944.3172977</a></li><li>Spacone et al., “SilentWear: an Ultra-Low Power Wearable System for EMG-based Silent Speech Recognition”, 2026: <a href="https://arxiv.org/html/2603.02847v2">arxiv.org/html/2603.02847v2</a></li><li>Gaddy and Klein, “Digital Voicing of Silent Speech”, EMNLP 2020: <a href="https://aclanthology.org/2020.emnlp-main.445/">aclanthology.org/2020.emnlp-main.445</a></li><li>Zheng, Crouch and Eggleston, “Surface Electromyography as a Natural Human-Machine Interface: A Review”: <a href="https://arxiv.org/pdf/2101.04658">arxiv.org/pdf/2101.04658</a></li></ul><p><strong>Hardware and places</strong></p><ul><li>Texas Instruments, INA128 product page: <a href="https://www.ti.com/product/INA128">ti.com/product/INA128</a></li><li>Upside Down Labs, BioAmp EXG Pill on Tindie: <a href="https://www.tindie.com/products/upsidedownlabs/bioamp-exg-pill-sensor-for-ecg-emg-eog-or-eeg/">tindie.com</a></li><li>Fab Lab UAE, Hamdan Foundation: <a href="https://www.ha.ae/5/programs/fab-lab-uae">ha.ae</a></li></ul><p><strong>Images</strong></p><ul><li>SilentWear figures reproduced from Spacone et al., arXiv 2026 (Figs. 1 and 2)</li><li>3D renders by the author in Blender. Head scan: Lee Perry-Smith by Infinite-Realities, CC BY 3.0.</li></ul><p><em>This is a personal learning project, not a medical device.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=90b87d390dc9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Speaking Without Sound: How Silent Speech Technology Works, and What It Could Do Next]]></title>
            <link>https://medium.com/@noormalwardak/speaking-without-sound-how-silent-speech-technology-works-and-what-it-could-do-next-168c3e576e32?source=rss-57e66c720c38------2</link>
            <guid isPermaLink="false">https://medium.com/p/168c3e576e32</guid>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[brain-computer-interface]]></category>
            <category><![CDATA[neurotechnology]]></category>
            <category><![CDATA[wearables]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Noormal Wardak]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 19:56:26 GMT</pubDate>
            <atom:updated>2026-10-22T19:14:40.803Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*34_E7paMTnhPg4KOP4ryng.png" /><figcaption><em>Cover. Design by the author on a historical anatomical engraving of the neck muscles (public domain). The orange dots mark three of the electrode regions used by </em><a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf"><em>Kapur et al. 2018</em></a><em>.</em></figcaption></figure><p><strong>The paper in 60 seconds</strong></p><blockquote><em>In 2018, Arnav Kapur, Shreyas Kapur and Pattie Maes at the MIT Media Lab published </em><a href="https://doi.org/10.1145/3172944.3172977"><em>AlterEgo: A Personalized Wearable Silent Speech Interface</em></a><em> at the IUI 2018 conference in Tokyo. Their idea: a wearable that lets you talk to a computer, an AI assistant or another person </em>without making a sound and without any visible movement<em>. When you intend to speak and say words to yourself, tiny neuromuscular signals fire in your face and neck. The headset picks them up, rebuilds the words, and answers you privately through bone conduction headphones. The authors call it </em><strong><em>intelligence augmentation</em></strong><em>: computing that feels like an extension of your own thinking, not a separate device. In their user study it reached a </em><strong><em>92% median word accuracy</em></strong><em>.</em></blockquote><p>That’s the promise. This article is about how it actually works, how far it really is from that promise, and what it could do next.</p><p>Read this sentence slowly, in your head.</p><p>Now read it again, but this time pay attention to your throat and your chin.</p><p>If you felt something, a tiny tug under your jaw or a flicker near your voice box, you might not be imagining it. Scientists call it <strong>subvocalization</strong>: <a href="https://www.sciencedirect.com/science/article/abs/pii/S0892199721000448">low-grade activity in your speech muscles while you read or think in words</a>. It’s faint, and it varies a lot from person to person. But when you <em>deliberately</em> say a word to yourself, the signal gets stronger. Your brain sends a weak “speak” command down to your speech muscles. They barely move. You can’t see it in a mirror. But they fire, and every time a muscle fires, it makes electricity.</p><p>In 2018, a team at the <a href="https://www.media.mit.edu/publications/alterego-IUI/">MIT Media Lab</a> built a headset that listens to exactly that electricity. They called it <strong>AlterEgo</strong>. People wore it, silently read numbers to themselves without moving their lips, and the system figured out which number they were “saying” with a <a href="https://www.media.mit.edu/publications/alterego-IUI/">92% median word accuracy</a>. In 2025 the project <a href="https://www.axios.com/2025/09/08/startup-query-ai-silent-speech-neural-interface">spun out of MIT</a> as a Boston startup, and its launch video, where the founder texts, takes notes and talks to an AI without making a sound, went viral.</p><p>I went into this thinking AlterEgo reads minds. It doesn’t. The founder himself says <a href="https://www.axios.com/2025/09/08/startup-query-ai-silent-speech-neural-interface">the hardware has no access to brain data</a>. What it actually does turned out to be way more interesting, and it’s the question this whole article is about:</p><p><strong>Why would anyone want this?</strong> Voice assistants already exist, so why go silent? The MIT paper names <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">four problems with talking to machines out loud</a>. There’s no <strong>privacy</strong>, because everyone around you hears what you say. They’re always <strong>listening</strong>, waiting for a wake word. They’re <strong>impersonal</strong>, so anyone can talk to your device. And they demand <strong>attention</strong>, because you have to be near the device and speak up. A silent interface fixes all four at once, if it works.</p><p><strong>How do you turn the electricity of muscles that barely move into words?</strong></p><p>To answer it, we have to follow a single word on its journey: from the moment your brain decides to say it, down through your nerves, into your face, out through your skin, and into a neural network.</p><p><strong>The short version.</strong> AlterEgo reads the faint electrical signals of your speech muscles, not your brain. It works well in the lab, on 10 words, for one person in one sitting. The big open problem is whether it still works tomorrow, after you take it off and put it back on. That problem, <strong>cross-session robustness</strong>, runs through this whole article, and it’s the one I’m going to test myself.</p><p><strong>How this article is organised</strong></p><ul><li><strong>Part 1: How it works.</strong> The speech dial, then one word’s journey from brain to muscle to skin to neural network.</li><li><strong>Part 2: How well it works.</strong> The results, the competition, and the five unsolved problems, each with what might fix it.</li><li><strong>Part 3: What I think, and the question I’ll answer. </strong>My hot takes, and the question my build will answer.</li><li><strong>Part 4: What comes next.</strong> What silent speech could do, who it could help, and the data the field is missing.</li></ul><p>Let’s go.</p><h3>Part 1: How it works</h3><h3>Speech isn’t on or off. It’s a dial.</h3><p>We think of speech as binary: you’re either talking or you’re not. In reality it’s a spectrum, and the further you slide down it, the weaker the signal your body gives off.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PaKYox04AIJPvHyOiIoYfg.png" /><figcaption><em>Figure 1. The speech dial. Key insight: the further down the dial you go, the weaker the muscle signal. AlterEgo works on the subvocal notch, which is why it needs no surgery, and also why its signal is so faint. Diagram by the author, based on the definition of internal vocalization in </em><a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf"><em>Kapur et al. 2018</em></a><em>.</em></figcaption></figure><p><strong>Spoken.</strong> Full voice. Air pushes through your vocal cords, your mouth shapes it, and a microphone can catch it from across the room.</p><p><strong>Whispered.</strong> Same mouth shapes, no vocal cord vibration. Still audible up close.</p><p><strong>Mouthed.</strong> Lips and jaw move, but no sound comes out. A camera can lip read this.</p><p><strong>Subvocal.</strong> This is the one. You say the word to yourself on purpose, the way you do when you read in your head, but your lips stay shut and nothing visibly moves. The MIT team calls it <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">internal vocalization</a>: the inner voice you notice while reading, minus any deliberate lip or muscle movement. The muscles still receive the command. They just fire at a tiny fraction of normal strength.</p><p><strong>Imagined.</strong> Pure thought, words you think without any intention to say them. As far as we know, nothing reaches the muscles here, so the only way to read it is from inside the brain.</p><p>AlterEgo lives on the fourth notch. That single fact explains almost everything about it: why it doesn’t need surgery, why it can’t read your private thoughts, and why it’s so hard to build. You’re trying to catch a whisper of electricity from muscles that are deliberately trying not to move.</p><h3>Step 1: A word leaves your brain</h3><p>Speaking is one of the most complex motor actions a human does. The <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">MIT paper</a> describes it as one of the most intricate, coordinated movements we perform, and when you trace a single word you see why.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*teKwyThepAjEbePb0EU2_A.png" /><figcaption><em>Figure 2. One word’s path from thought to muscle. Key insight: speech muscles get their commands through cranial nerves from the brainstem, not through the spinal cord. Diagram by the author, based on the </em><a href="https://www.ncbi.nlm.nih.gov/sites/books/NBK555891/"><em>corticobulbar tract</em></a><em> (StatPearls, NCBI) and </em><a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf"><em>Kapur et al. 2018</em></a><em>.</em></figcaption></figure><p><strong>1. Planning.</strong> Broca’s area, in the left frontal lobe, turns an idea into a language plan: which word, which sounds, in what order.</p><p><strong>2. Sequencing.</strong> The supplementary motor area turns that plan into a movement program, like choreography for your mouth.</p><p><strong>3. Commanding.</strong> The ventral sensorimotor cortex, a strip of brain running along the side of your head, fires the actual commands. It decides which muscles activate and how fast.</p><p><strong>4. Relaying.</strong> Those commands travel down a bundle of fibres called the corticobulbar tract to the brainstem. Here’s a detail I had to double check: your arm and leg muscles get their orders through the spinal cord, but speech muscles are wired straight from motor centres in the brainstem, through cranial nerves.</p><p><strong>5. Delivering.</strong> Four cranial nerves do most of the work for speech:</p><ul><li><strong>Trigeminal (V)</strong> moves the jaw and part of the floor of the mouth</li><li><strong>Facial (VII)</strong> moves the lips, chin and the thin platysma muscle on your neck</li><li><strong>Vagus (X)</strong> controls the voice box</li><li><strong>Hypoglossal (XII)</strong> moves the tongue</li></ul><p><strong>6. Firing.</strong> At the end of each nerve is a neuromuscular junction. The nerve releases a chemical messenger called acetylcholine, which opens channels in the muscle fibre and lets sodium rush in. That flood of charged particles sends an electrical wave, an action potential, rippling along the fibre.</p><p>One motor neuron controls a whole group of muscle fibres at once, a team called a <strong>motor unit</strong>. When it fires, they all fire together, and their combined electrical wave is strong enough to spread through the surrounding tissue and reach your skin.</p><p>That’s the signal AlterEgo catches. Not your thoughts. The electrical echo of a command your brain already decided to send.</p><p>And here’s what makes subvocal speech so hard: when you barely articulate, your brain recruits only a few small motor units and fires them gently. The echo is tiny.</p><h3>Step 2: Catching electricity through skin</h3><p>Recording muscle electricity from the surface of the skin is called <strong>surface electromyography</strong>, or sEMG. Doctors use it to check nerve damage, and it’s the same principle Meta uses in its <a href="https://www.uploadvr.com/alterego-silent-speech-wearable-announcement/">neural wristband</a> to read finger movements. AlterEgo just points it at your face.</p><p>The problem is that by the time a subvocal signal reaches your skin, it’s been smeared through fat and tissue and shrunk to a few millionths of a volt. Meanwhile your body and your room are screaming much louder signals right on top of it: your heartbeat, your eye blinks, every swallow, and the electrical hum of every power line around you.</p><p>So the first engineering question is simple: <strong>where do you even put the electrodes?</strong></p><p>The MIT team didn’t guess. They started with a <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">30-point grid</a> covering the chin, cheeks, jaw, throat and under the eyes, recorded people silently saying words, and ranked every spot by how well its signal could tell words apart. Seven made the cut.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RnB11D8qbx42RVdAHoJMIg.png" /><figcaption><em>Figure 3. Where AlterEgo listens. Key insight: the chin and the voice box beat the lips. Rankings from Table 1 of </em><a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf"><em>Kapur et al. 2018</em></a><em>, drawn by the author on a public domain anatomical engraving.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yRSkqi06y9_DRVaullmHRQ.png" /></figure><p>The winner surprised me. I expected the lips. Instead, the <strong>chin</strong> ranked first, with the voice box second and third. The paper doesn’t explain why. It only <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">ranked each site by how well its signal told words apart</a>.</p><p><strong>My hypothesis (not something the paper tested):</strong> when you subvocalize, you clamp your lips on purpose, so the lips are actively trying not to move. The deeper muscles under the chin and in the throat still get recruited, because they help position the tongue and larynx, and you’re not consciously holding them still.</p><p>A few other details that matter if you ever try to build this:</p><ul><li><strong>Electrodes.</strong> Gold plated silver electrodes with a conductive area just 1.45 mm wide, used with conductive paste, gave better signals than 4 mm dry electrodes. Paste lowers the resistance between skin and metal.</li><li><strong>A reference point.</strong> One extra electrode sits on the earlobe or wrist, somewhere with no speech muscles, so the system can subtract everything the face and the reference have in common. What’s left is mostly the speech signal.</li><li><strong>Sampling.</strong> The signal was sampled 250 times per second and amplified 24 times using an <a href="https://openbci.com/">OpenBCI</a> board, an open source biosensing kit anyone can buy. It then travels over Bluetooth to a phone, which passes it to the model.</li><li><strong>Mains hum.</strong> The team cancelled 60 Hz interference, the frequency of the US power grid. Fun fact for anyone building this in Dubai: the UAE grid runs at 50 Hz, so the filter has to move.</li><li><strong>The frame.</strong> The headset was <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">3D printed in resin</a> and wraps around the back of the head, with brass rods holding each electrode arm steady. The arms are modular, so they can be moved for different faces. The whole design is about one thing: stopping the electrodes from moving.</li></ul><h3>Step 3: Turning noise into words</h3><p>Seven wires of messy electricity come off the face. Here’s how the <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">2018 system</a> turns them into a word.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sRTePbAKl0agxP8_vAjzPQ.png" /><figcaption><em>Figure 4. The AlterEgo pipeline. Key insight: the muscle signal is treated like audio, using the same MFCC features as speech recognition. Diagram by the author, based on the methods in </em><a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf"><em>Kapur et al. 2018</em></a><em>.</em></figcaption></figure><p><strong>Clean it.</strong> First, a band pass filter keeps only frequencies between 1.3 and 50 Hz. The low cut removes slow drift, the high cut removes fast movement artifacts. A separate notch filter knocks out the power line hum. Then a technique called Independent Component Analysis pulls the mixed signals apart into separate sources, like isolating one voice in a crowded room, so movement noise can be thrown away. Any sudden spike more than 30 microvolts above baseline gets deleted as a glitch.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RPz2Mk0U2G2KPl17rCCARg.png" /><figcaption><em>Figure 5. Finding a word in the noise. Key insight: after filtering removes drift and power line hum, the word shows up as a burst you can’t see in the raw signal. This is a simulation made by the author, not a real recording. The filter settings follow </em><a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf"><em>Kapur et al. 2018</em></a><em>.</em></figcaption></figure><p><strong>Describe it.</strong> This is my favourite part. To turn the cleaned signal into something a neural network can learn from, the team used <strong>mel frequency cepstral coefficients</strong>, or MFCCs. If that sounds familiar, it’s because MFCCs are the classic way to describe <em>audio</em> in speech recognition. The signal is chopped into 25 millisecond windows, each window’s frequencies are measured, and the result is squashed into a compact fingerprint. In other words, they treated the electricity of your chin muscles as if it were sound. A word’s muscle pattern gets the same treatment as a word’s audio.</p><p><strong>Classify it.</strong> The fingerprints feed a 1D convolutional neural network. Each layer slides 400 small filters across the signal over time, looking for short patterns, the electrical equivalent of a “t” or an “oo”. Two of those layers, then a fully connected layer, then an output that votes on which word it heard. Half the connections are randomly switched off during training (a trick called dropout) so the network learns general patterns instead of memorising one recording.</p><p><strong>Narrow it down.</strong> Here’s a clever bit of design. AlterEgo never tries to recognise every English word at once. You first silently say a trigger word, like an app name, and that switches the system into a small vocabulary. Say “arithmetic” and it only listens for the digits 0 to 9 plus <em>add, subtract, multiply, divide</em> and <em>percent</em>. Fewer options means fewer mistakes.</p><p><strong>Reply.</strong> Once a word is recognised, the computer does the task and answers out loud, but through <strong>bone conduction</strong> headphones that vibrate your skull instead of covering your ears. You hear the answer, nobody around you does, and your ears stay open to the world. Input is silent, output is private. That’s what makes it a closed loop.</p><p><strong>How well did it work?</strong> First the team built their training data: 3 volunteers recorded about <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">31 hours of silently spoken words</a>, spread across different sessions so the model would learn to cope with session changes. Then they tested 10 new people, aged 19 to 31, who’d never used the device. Each silently read 750 random digits. Averaged over every run, the system got 92.01% right, with each person’s median landing somewhere between about 89% and 95%. It took about 0.43 seconds from the end of a word to the answer, fast enough to feel like a conversation.</p><p><strong>The catch hiding in that 92%.</strong> To be fair, the 31 hour training set was recorded across sessions on purpose. But the headline test split each new person’s recordings <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">80/20 at random</a>, so the test words came from the same sitting as the training words, with the electrodes in exactly the same place on the same skin. It’s the easiest possible test. Keep it in mind, because it’s the thread that runs through the rest of this article.</p><h3>Part 2: How well it works, and what’s unsolved</h3><h3>Who else is chasing silent speech</h3><p>AlterEgo isn’t alone. Silent speech is a race with two very different strategies: listen to the <strong>muscles</strong> from outside the body, or listen to the <strong>brain</strong> from inside the skull.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u63V0II1jb1LLajTiihu5A.png" /><figcaption><em>Figure 6. The silent speech map. Key insight: implants get huge vocabularies, wearables get tiny ones, and nobody has a large vocabulary without surgery yet. Positions are approximate, drawn by the author from the published results in the table below.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WZRtZLl9eDCtFXp6renkEA.png" /></figure><p>Two of these deserve a closer look.</p><p><strong>Berkeley went the other way on vocabulary.</strong> Instead of a tiny word list, David Gaddy and Dan Klein tried to turn silently <em>mouthed</em> sentences straight into synthesized audio. Their trick: record people saying the same sentences both out loud and silently, then transfer the audio from the spoken version onto the silent one so the model has something to learn from. It won <a href="https://syncedreview.com/2020/11/18/emnlp-2020-best-paper-award-goes-to-uc-berkeley-team/">Best Paper at EMNLP 2020</a>. Even better for people like me, they released <a href="https://zenodo.org/records/4064409">nearly 20 hours of facial EMG data</a> for free, so anyone can train on it.</p><p><strong>Stanford skipped the muscles entirely.</strong> Using electrode arrays implanted in the motor cortex of four people with ALS or stroke, the team found that <a href="https://neurosciencenews.com/bci-inner-speech-decoding-29574/">attempted and imagined speech activate overlapping regions</a> of the brain, and decoded imagined sentences in real time. Participants actually <a href="https://www.ninds.nih.gov/news-events/news/press-releases/decoding-inner-speech-brain-signals">preferred imagined speech</a> because it took less physical effort. But it raised an obvious problem: if a device can hear your inner voice, what stops it hearing things you didn’t mean to share? Their answer was a mental password. The decoder stays locked until the user imagines a rare phrase like “chitty chitty bang bang”, which it recognised <a href="https://hayadan.com/inner-speech-bci-real-time-cell-2025-http">more than 98% of the time</a>.</p><p>Put those side by side and you see the tradeoff that defines this whole field. <strong>Go inside the skull and you get a huge vocabulary, but you need brain surgery. Stay outside and anyone can wear it, but the signal is so faint that you’re stuck with small vocabularies or visible mouthing.</strong> AlterEgo is betting it can close that gap from the outside.</p><h3>Hype vs reality: five unsolved problems, and what might fix them</h3><p>The launch video makes it look finished. The research says otherwise. Here are the five problems I think matter most, starting with the biggest. For each one I’ve added what might fix it.</p><p><strong>1. Today’s model may not work tomorrow. This is the big one.</strong> Every time you put the headset back on, the electrodes land a few millimetres differently, your skin is a bit sweatier or drier, and the signal shifts. The MIT team knew this. Their paper says it’s <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">“imperative for the electrodes to not shift position”</a>, and lists multi-user, longitudinal and real-world testing as future work. Newer research shows how big the drop can be. In 2026 the <a href="https://arxiv.org/html/2603.02847v1">SilentWear</a> team tested an EMG wearable on 4 people silently saying 8 commands. Trained and tested on a mix of all its sessions, it got 77.5% right. In a new session, after the electrodes were repositioned, it fell to 59.3%. Researchers still call cross-session robustness <a href="https://arxiv.org/pdf/2603.10324">a core problem</a> for EMG silent speech.</p><p><em>What might fix it:</em> recalibrating with a little fresh data every time you put it on (SilentWear’s incremental fine tuning won back more than 10% accuracy), training on many sessions instead of one, and hardware that lands in the same spot every time, like a moulded headset instead of stick-on electrodes.</p><p><strong>2. The vocabulary is tiny.</strong> The 92% result came from recognising 10 digits, and the whole 2018 system worked by switching between small word lists. Recognising 10 words is a classification problem. Recognising any sentence in English is a completely different problem, and no one has shown non invasive <em>subvocal</em> speech doing it reliably.</p><p><em>What might fix it:</em> bigger datasets and smarter training. Berkeley’s trick of borrowing audio from spoken recordings (more below) helped mouthed speech, and a <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC5851476/">2017 study</a> reached a 2,500 word vocabulary, but with mouthed words, not subvocal ones.</p><p><strong>3. Everyone needs their own model.</strong> Your face, your muscle sizes and your way of speaking are different from mine. The 2018 system was trained per person, and building one model that works for everyone out of the box is still on the <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">to do list</a>.</p><p><em>What might fix it:</em> pretraining one big model on many people, then fine tuning it on a few minutes of your own data. It’s the same idea that let voice assistants learn new accents. This one is my suggestion, not something the AlterEgo team has tested.</p><p><strong>4. Some sounds hide.</strong> Sounds made by the tip of the tongue behind your teeth, like “t”, “d” and “n”, happen deep inside the mouth, far from any surface electrode. NASA’s earlier system <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">couldn’t recognise them well</a>. Throat and under jaw electrodes help, but there’s a physical limit to what you can read through skin.</p><p><em>What might fix it:</em> more sensors in a smaller space. A 2022 study using <a href="https://www.sciencedirect.com/science/article/abs/pii/S1746809422007522">high density surface EMG</a>, grids of many tiny electrodes, reported 96.37% phrase recognition. A language model can also guess a hidden sound from the words around it.</p><p><strong>5. “Near telepathic” is marketing.</strong> The startup calls its device the <a href="https://thedebrief.org/mit-linked-startup-unveils-near-telepathic-wearable-device-for-silent-communication/">world’s first near-telepathic wearable</a>. But the hardware reads muscles, not neurons, and experts have <a href="https://www.uploadvr.com/alterego-silent-speech-wearable-announcement/">criticised the company</a> for presenting itself as a brain computer interface. That matters, because overselling a technology is the fastest way to lose people’s trust in it.</p><p><em>What might fix it:</em> publishing real accuracy numbers, across days and across people, and calling it what it is: a muscle interface.</p><p>And here’s the opportunity hiding inside all five: <strong>every one of them is a data and design problem, not a physics problem.</strong> The signal is real. What’s missing is enough recordings, from enough people, over enough days.</p><h3>Part 3: What I think, and the question I’ll answer My hot takes</h3><p><strong>Hot take 1: the muscle is a consent button, but it’s a dimmer, not an on/off switch.</strong></p><p>Stanford had to invent a mental password because an implant in the motor cortex can pick up inner speech you never meant to share. AlterEgo gets a lot of that protection for free. It’s trained on <em>deliberate</em> internal vocalization, and the faint subvocalization of everyday thinking is much weaker and far less consistent. A <a href="https://www.sciencedirect.com/science/article/abs/pii/S0892199721000448">2023 study of 45 people</a> measured voice box muscle activity during silent tasks and found it went up in some people, down in others, and was negligible for most muscles. So the signal AlterEgo needs mostly doesn’t exist until you choose to send it. Nobody’s marketing this, but I think it’s the most important design property the technology has.</p><p>“Mostly” is the catch. Remember the tug you might have felt in your throat at the start of this article? That’s the leak. The more sensitive these devices get, the further down the speech dial they reach, and the closer they get to that leak. My prediction: within five years, the argument that decides whether people wear these won’t be accuracy. It’ll be privacy.</p><p><strong>Hot take 2: the best first users aren’t people talking to AI on the train. They’re people who lost their voice box.</strong></p><p>The launch pitch leans on ALS. But in ALS, the motor neurons that drive the speech muscles degenerate over time, so the very signal AlterEgo reads can fade as the disease progresses. That’s exactly where brain implants like Stanford’s win, because they read the cortex upstream of the damage.</p><p>Now think about someone who’s had a <strong>laryngectomy</strong>, their voice box removed, usually because of cancer. Their brain is fine. Their nerves are fine. Their lips, jaw and tongue still work. The only thing missing is the sound source. That’s the perfect match for muscle based silent speech.</p><p>This isn’t a tiny group. In the US alone, surgeons performed an estimated <a href="https://digitalcommons.unl.edu/cgi/viewcontent.cgi?article=1182&amp;context=publichealthresources">3,000 to 5,000 total laryngectomies a year</a> between 1997 and 2008. And it has been tested. In 2017, a team led by Geoffrey Meltzner recorded <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC5851476/">8 people at least six months after total laryngectomy</a>, using 8 EMG sensors on the face and neck while they mouthed phrases from a 2,500 word vocabulary. The average word error rate was 10.3%, and 13.6% with only 4 sensors. That’s mouthed speech, not subvocal, but it proves the muscle signal is still there and still readable without a voice box.</p><p>There’s a twist, though. Two of AlterEgo’s top three electrode sites sit on the voice box, which these patients no longer have. Meltzner’s team put their neck sensors <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC5851476/">under the chin and along the front of the neck</a> instead.</p><p><strong>My hypothesis (untested):</strong> a laryngectomy version of AlterEgo should drop the voice box sites and lean on the chin and under the jaw, which were AlterEgo’s first and fourth best sites. That’s a question I’d love to ask a head and neck surgeon.</p><h3>The question my build will answer</h3><p>I don’t want to just rebuild AlterEgo. I want my build to answer one question:</p><p><strong>How much does silent speech accuracy drop when articulation gets weaker, and when the electrodes are taken off and put back on another day?</strong></p><p><strong>My hypotheses:</strong></p><ol><li>Accuracy will drop as articulation shrinks: mouthed words will be much easier to recognise than subvocal ones.</li><li>A model trained on day one and tested on day two, after I remove and replace the electrodes, will lose far more accuracy than one tested in the same session. That gap will be bigger for subvocal speech than for mouthed speech.</li><li>A short recalibration on day two, a few recordings per word, will win back most of that loss.</li><li>The chin and under jaw electrodes will carry the most useful information, matching MIT’s ranking. My extra guess: they’ll also be the most stable from day to day.</li></ol><blockquote><em>How I’m building it, from the circuit to the two-day experiment, gets its own article: Hearing the Words You Never Say.</em></blockquote><p>This article covers English only. Whether silent speech works equally well across languages, especially ones that use sounds made deep in the throat, is a question I want to come back to later.</p><h3>Part 4: What comes next</h3><h3>What silent speech could already do</h3><p>The 2018 team didn’t stop at digits. They built <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">small apps on top of the headset</a>, each opened by silently saying a trigger word:</p><ul><li><strong>Arithmetic.</strong> Silently say “2581 times 698 divide 2 add 13” and hear the answer in your ear.</li><li><strong>Smart home.</strong> “Light on”, “fan off”, with no sound and no movement.</li><li><strong>World clock and calendar.</strong> Ask the time in a city, or step through your meetings.</li><li><strong>Chess and Go with an AI.</strong> Silently describe the board, and the AI whispers back the next move.</li><li><strong>Silent replies.</strong> Answer someone with “call you later” in the middle of a meeting.</li></ul><h3>Where it could go</h3><p>Those are small. The big possibilities open up when the vocabulary grows. These are my predictions, not results:</p><ul><li><strong>A voice for people who lost theirs.</strong> People after a laryngectomy still move the muscles of speech but can’t make sound. As Hot take 2 showed, muscle based silent speech has already been tested with them, with promising results.</li><li><strong>Talking where you can’t talk.</strong> A library, a hospital ward at night, a loud factory floor, or a meeting where you need to ask your AI something without interrupting.</li><li><strong>An AI that feels like thinking.</strong> If it ever gets fast and accurate enough, asking an AI a question could feel less like using a device and more like remembering something. That’s the paper’s intelligence augmentation idea, and it’s the one that excites me most, and worries me most.</li></ul><h3>The missing ingredient is data</h3><p>The paper’s own first item of future work is to <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">collect more data</a> for a model that works for many people out of the box. Every unsolved problem in Part 2 comes back to the same thing: not enough recordings, from enough people, over enough days.</p><p>So here’s my small contribution. Every recording from my build, mouthed and subvocal, day one and day two, with photos of where the electrodes sat, will be published openly, the same way Berkeley shared theirs. It starts with one person: me. If it works, the next step is a small cross-session dataset from volunteers, with proper consent, because nobody has enough of this kind of data yet.</p><h3>About me</h3><p>Hey, I’m Noormal Wardak, a student in Dubai and an innovator at <a href="https://tks.world/">TKS</a>. I’m fascinated by the point where the brain meets the outside world, and I’m heading toward neuroscience at university. Outside of this, I build voice AI at <a href="https://convoscale.app/">Convoscale</a>, and I put what it earns back into iPads and internet data so girls can keep studying in Afghanistan.</p><p>If you work on silent speech, EMG or speech neuroprostheses, I’d genuinely love to talk. Find me on <a href="https://www.linkedin.com/in/noormal-wardak/?isSelfProfile=true">https://www.linkedin.com/in/noormal-wardak/</a> or at (noormal.wardakak@gmail.com)</p><p><a href="https://medium.com/@noormalwardak/hearing-the-words-you-never-say-and-controlling-your-devices-with-them-90b87d390dc9"><em>Clap if you learned something. The build is up now: Hearing the Words You Never Say</em></a></p><h3>Sources</h3><ul><li>Kapur, Kapur and Maes, <a href="https://andymatuschak.org/files/papers/Kapur%20et%20al%20-%202018%20-%20AlterEgo.pdf">AlterEgo: A Personalized Wearable Silent Speech Interface</a>, IUI 2018 (<a href="https://www.media.mit.edu/publications/alterego-IUI/">MIT Media Lab page</a>)</li><li>Axios, <a href="https://www.axios.com/2025/09/08/startup-query-ai-silent-speech-neural-interface">Startup lets you query AI with silent speech</a>, September 2025</li><li>UploadVR, <a href="https://www.uploadvr.com/alterego-silent-speech-wearable-announcement/">Alterego’s silent speech could be the answer to dictating text in public</a>, September 2025</li><li>The Debrief, <a href="https://thedebrief.org/mit-linked-startup-unveils-near-telepathic-wearable-device-for-silent-communication/">MIT-linked startup unveils near-telepathic wearable</a>, October 2025</li><li>Helou et al., <a href="https://www.sciencedirect.com/science/article/abs/pii/S0892199721000448">Intrinsic laryngeal muscle activity during subvocalization</a>, Journal of Voice, 2023</li><li>Bhardwaj and Yadala, <a href="https://www.ncbi.nlm.nih.gov/sites/books/NBK555891/">Neuroanatomy, Corticobulbar Tract</a>, StatPearls, NCBI</li><li>Gaddy and Klein, <a href="https://arxiv.org/pdf/2010.02960">Digital Voicing of Silent Speech</a>, EMNLP 2020, and the <a href="https://zenodo.org/records/4064409">Silent Speech EMG dataset</a></li><li>Synced, <a href="https://syncedreview.com/2020/11/18/emnlp-2020-best-paper-award-goes-to-uc-berkeley-team/">EMNLP 2020 Best Paper goes to UC Berkeley team</a></li><li>Kunz, Willett et al., <a href="https://www.cell.com/cell/fulltext/S0092-8674(25)00681-6">Inner speech in motor cortex and implications for speech neuroprostheses</a>, Cell 2025</li><li>Stanford Medicine, <a href="https://med.stanford.edu/news/all-news/2025/08/brain-computer-interface.html">Study of promising speech-enabling interface</a>, August 2025</li><li>NIH NINDS, <a href="https://www.ninds.nih.gov/news-events/news/press-releases/decoding-inner-speech-brain-signals">Decoding inner speech brain signals</a></li><li>Neuroscience News, <a href="https://neurosciencenews.com/bci-inner-speech-decoding-29574/">Mind-reading tech decodes inner speech with up to 74% accuracy</a></li><li>Spacone et al., <a href="https://arxiv.org/html/2603.02847v1">SilentWear: an ultra-low power wearable system for EMG-based silent speech recognition</a>, 2026</li><li>NasoVoce, <a href="https://arxiv.org/pdf/2603.10324">related work on silent speech interfaces</a>, 2026</li><li><a href="https://www.sciencedirect.com/science/article/abs/pii/S1746809422007522">Decoding silent speech from high-density surface EMG</a>, Biomedical Signal Processing and Control, 2022</li><li>Meltzner et al., <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC5851476/">Silent speech recognition as an alternative communication device for persons with laryngectomy</a>, IEEE/ACM TASLP, 2017</li><li>Maddox and Davies, <a href="https://digitalcommons.unl.edu/cgi/viewcontent.cgi?article=1182&amp;context=publichealthresources">Trends in total laryngectomy in the era of organ preservation</a></li></ul><p><em>Every figure has a caption with its source and key insight. All figures were made by the author. The cover and Figure 3 are built on historical anatomical engravings (public domain). Figure 5 is a simulation built to show the processing steps, not a real recording.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=168c3e576e32" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nobody dies of a bed shortage. They die waiting for one.]]></title>
            <link>https://medium.com/@noormalwardak/nobody-dies-of-a-bed-shortage-they-die-waiting-for-one-525e76649970?source=rss-57e66c720c38------2</link>
            <guid isPermaLink="false">https://medium.com/p/525e76649970</guid>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[afghanisthan-healthcare]]></category>
            <category><![CDATA[global-health-care]]></category>
            <category><![CDATA[public-health]]></category>
            <dc:creator><![CDATA[Noormal Wardak]]></dc:creator>
            <pubDate>Sun, 20 Sep 2026 19:24:09 GMT</pubDate>
            <atom:updated>2026-09-20T19:25:24.054Z</atom:updated>
            <content:encoded><![CDATA[<p>Three years inside Kabul’s Wazir Akbar Khan Hospital, and what 21.9 million people in need actually looks like from the corridor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vDmlVRWoe6Ua_-w7SG3mdw.jpeg" /></figure><p>There’s a corridor in Wazir Akbar Khan Hospital in Kabul where the queue for the X-ray starts before the machine is switched on. Men holding their own arm. Mothers with a child on their hip. Old men on the floor because there’s nowhere to sit. Some travelled from outside Kabul to stand in that line. Some will stand in it all day and go home without being seen.</p><p>I’ve been in that hospital for three years. My father runs it. I’m not a doctor. I’m a student, and I’ve been on rounds, scrubbed in at the table, and slept in my scrubs on a bunk in a side room because the day didn’t end when I got tired.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O_otNS31dC2ZMGt89XZWhA.jpeg" /></figure><p>I want to tell you what that corridor actually represents, because it isn’t one bad hospital.</p><h4>The scale of it</h4><p>In 2026, around 21.9 million people, roughly 45 per cent of Afghanistan’s population, need humanitarian assistance. The UN’s $1.7 billion appeal was 14 per cent funded as of mid-year.</p><p>Funding cuts in 2025 closed more than 422 health facilities, cutting off an estimated 3 million Afghans from basic health services. Those people didn’t stop being sick. They just have further to travel now, and the place they travel to is a hospital like ours.</p><p>Afghanistan has roughly 10 health workers per 10,000 people. The WHO’s recommendation is 44.5. There are about 0.4 hospital beds per 1,000 people here. The global average is 3.3.</p><p>Afghanistan has one of the highest maternal mortality rates on earth, around 638 deaths per 100,000 live births, the highest in Asia. With girls barred from education, UNICEF estimates the country could lose more than 25,000 female teachers and health workers by 2030. Which means the women who can’t get seen today will have even fewer people to see them in five years.</p><h4>What those numbers look like from the floor</h4><p>Numbers like that are easy to scroll past. Here’s the version you can see.</p><p>It looks like two patients in one bed. In some Afghan facilities, up to three people share a bed. It looks like a bed that emptied forty minutes ago still sitting empty, because the only person who knows it’s free is a nurse busy in another room. No board. No screen. No shared list. Paper and memory.</p><p>It looks like a maternity centre in Kabul with a single ultrasound machine, and entire provinces sending anyone who needs a CT scan into the capital. It looks like buildings put up forty years ago for a fraction of the people now standing in them.</p><p>Somewhere inside that, a woman who arrived at 7am is still on the floor at 4pm.</p><h4>The doctors are not the problem</h4><p>I need to say this clearly, because the easy story about Afghanistan is a story about helplessness. It’s wrong.</p><p>I’ve watched surgeons at that hospital operate with a fraction of the equipment they should have and save the person on the table anyway. The skill in that building is not in short supply. The staff are not the bottleneck. They’re the only reason the number of deaths isn’t far worse.</p><h4>Two different problems, and only one of them is ours</h4><p>Some of this needs the world. Funding, medicine, equipment, and above all letting women train as doctors and nurses again. No app fixes any of that, and I’m not going to pretend otherwise.</p><p>But some of it is smaller and closer. A hospital that can’t see its own beds is wasting capacity it already has. That’s not a medicine problem. It’s an information problem, and it’s the kind I know how to work on.</p><p>So I’m building one thing, not a system overhaul. A live bed count for one ward, updated from a nurse’s phone in under five seconds, visible at the station and to admissions. It has to work with no internet and no power, because in Kabul that’s a normal Tuesday, and it has to be in Dari and Pashto. If it’s slower than shouting down the corridor, it’s dead.</p><p>I’ll write about what happens, including if it fails.</p><p>If you’ve worked in a hospital like this, or built anything for one, tell me what I’m missing. I’d rather hear it now than after.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=525e76649970" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>