<figure class="kg-card kg-embed-card"><iframe width="200" height="113" src="https://www.youtube.com/embed/ww13MkJrieY?feature=oembed&cc_load_policy=1" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen="" title="Cámara hears"></iframe></figure><p>Cámara is a robotic creature built on a commercial PTZ camera and a Raspberry Pi 5. When someone says <strong>“oye cámara”</strong> (“hey camera”) —its wake word—, it understands it, knows which side the voice came from and turns there.</p><p>Its ear is the microphone of a PS3 Eye, a row of microphones <strong>20 mm</strong> apart. The voice reaches the nearest microphone slightly earlier, and from that difference Cámara knows which side it is being spoken to from. Then <a href="https://alphacephei.com/vosk/?ref=localhost" rel="noreferrer">Vosk</a> —a speech recognizer running on the Pi, offline— understands “oye cámara” and waits <strong>4 seconds</strong> for a command, within a closed vocabulary of <strong>62 words</strong>: “sígueme” (follow me), “foto” (photo), “baila” (dance)…</p>
<!--kg-card-begin: html-->
<figure class="kg-card luf-model"><script type="module" src="https://luf.tech/assets/js/model-viewer.min.js"></script><model-viewer src="https://luf.tech/assets/model/camara-oido.glb?v=1" alt="3D model of Cámara's base with the PS3 Eye and its microphones highlighted in pink, behind the Sense HAT" poster="https://luf.tech/content/images/2026/10/ps3eye-v3-7.png" camera-controls touch-action="pan-y" auto-rotate environment-image="neutral" tone-mapping="neutral" exposure="0.75" shadow-intensity="0.5" camera-orbit="-80deg 93deg auto" style="width:100%;height:460px;background:transparent"></model-viewer><figcaption>Cámara's ear: the PS3 Eye and its microphones, in pink, mounted on the base next to the Raspberry Pi. Rotate it with your mouse or finger.</figcaption></figure>
<!--kg-card-end: html-->
<!--kg-card-begin: html-->
<div class="luf-mods"><div class="luf-mod"><h3>Listen</h3><p>A single program opens the microphone and hands out audio and direction.</p></div><div class="luf-mod-arrow" aria-hidden="true"><svg viewBox="0 0 28 16"><path d="M0 8h25M19 2l6 6-6 6" fill="none" stroke="currentColor" stroke-width="2"/></svg></div><div class="luf-mod"><h3>Locate</h3><p>Which side the voice came from: left, front or right.</p></div><div class="luf-mod-arrow" aria-hidden="true"><svg viewBox="0 0 28 16"><path d="M0 8h25M19 2l6 6-6 6" fill="none" stroke="currentColor" stroke-width="2"/></svg></div><div class="luf-mod"><h3>Understand</h3><p>Recognize “oye cámara” offline.</p></div></div>
<!--kg-card-end: html-->
<h2 id="listen-one-microphone-one-owner">Listen: one microphone, one owner</h2><p>The PS3 Eye delivers 4 channels at 16 kHz. A single program opens them, <a href="https://github.com/introlab/odas?ref=localhost" rel="noreferrer">ODAS</a>, which works out where the sound comes from and hands out two things: the direction and the audio already focused on it. Nobody else touches the microphone.</p><p>That “focused” audio comes from a technique called <strong>beamforming</strong>. Since the voice reaches each microphone with a small offset, ODAS delays each channel just enough for the voice coming from the chosen direction to line up, and then adds them. The voice adds to itself and gets stronger; noise from other directions stays misaligned and fades. It is like turning an ear toward whoever is speaking, without moving anything.</p><figure class="kg-card kg-image-card kg-card-hascaption"><img src="https://luf.tech/content/images/2026/10/beamforming_en.png" class="kg-image" alt="Beamforming diagram: three channels with voice and noise offset, the delay that lines up the voice, and the sum where the voice gets stronger and the noise fades" loading="lazy"><figcaption>Diagram: the offsets are exaggerated so they can be seen.</figcaption></figure>
<!--kg-card-begin: html-->
<div class="luf-concl"><div><span class="luf-concl-k">Takeaway</span><h4>One physical resource, one owner</h4><p>If two programs open the same device, failures come and go and cannot be reproduced.</p></div></div>
<!--kg-card-end: html-->
<h2 id="locate-the-side-not-the-angle">Locate: the side, not the angle</h2><p>The voice reaches one microphone slightly before another, and that shift is the direction. With <strong>three microphones</strong> in a row, the row spans <strong>40 mm</strong>: enough to tell <strong>left from right</strong>, not to give a fine angle. With the voice in front it reads 0 in 20 out of 20 samples; to the sides the reading scatters <strong>±15°</strong>.</p><figure class="kg-card kg-image-card kg-card-hascaption"><img src="https://luf.tech/content/images/2026/10/localizar_en.png" class="kg-image" alt="A: the 3 microphones in a row and a voice arriving at an angle. B: the same wave reaches each microphone tens of microseconds apart" loading="lazy"><figcaption>Between one microphone and the next there are tens of microseconds, less than one sample at 16 kHz. Enough to know the side.</figcaption></figure><p>Direction is read <strong>backwards</strong>: by the time Cámara recognizes “oye cámara”, the phrase is over. It looks at the <strong>2.0 to 0.2 s</strong> before and compares them with the <strong>3.5 s</strong> before that. The winner is the side that <strong>grew</strong>, not the loudest: that way the noise already in the room does not win.</p>
<!--kg-card-begin: html-->
<div class="luf-data"><table><caption>How much a side has to grow to win, over 13 real calls with the side labeled</caption><thead><tr><th>criterion</th><th class="num">correct</th><th class="num">turns to the wrong side</th></tr></thead><tbody><tr><td>by 15° sector, threshold 0.10</td><td class="num">7/13</td><td class="num">0</td></tr><tr class="pick"><td>by side, threshold 0.10</td><td class="num">9/13</td><td class="num">0</td></tr><tr><td>by side, threshold 0.06 to 0.08</td><td class="num">10/13</td><td class="num">1</td></tr></tbody></table><p class="note">Asking for less wins one call, but adds a turn to the wrong side. When it is not sure, Cámara does not turn: it searches.</p></div>
<!--kg-card-end: html-->
<!--kg-card-begin: html-->
<div class="luf-concl"><div><span class="luf-concl-k">Takeaway</span><h4>Ask the sensor for what its geometry gives</h4><p>A short aperture separates sides; demanding angles from it only adds noise to the decision.</p></div><div><span class="luf-concl-k">Takeaway</span><h4>Getting the side wrong is worse than not knowing</h4><p>The threshold is chosen by counting the errors that cost most, not the hits.</p></div></div>
<!--kg-card-end: html-->
<h2 id="understand-few-words-offline">Understand: few words, offline</h2><p>The recognizer is <strong>Vosk</strong>, which runs on the Pi and only knows a closed list: the wake word and 62 command words. It does not hear raw audio: it hears what ODAS builds by pointing where it thinks the voice is. Each word has to pass a confidence of <strong>0.60</strong>, and after “oye cámara” the command has to arrive within <strong>4 s</strong>.</p><p>Leaving Vosk only “oye cámara” saves <strong>25 %</strong> CPU, but breaks saying “oye cámara sígueme” in one go. So Vosk keeps hearing every word, and what gets filtered is what to do with them.</p>
<!--kg-card-begin: html-->
<div class="luf-concl"><div><span class="luf-concl-k">Takeaway</span><h4>What is heard at home stays at home</h4><p>A robot that listens all day in a home should not send that audio to a server. Recognizing speech on the robot itself, offline, makes privacy a property of the design rather than a promise.</p></div><div><span class="luf-concl-k">Takeaway</span><h4>A closed vocabulary, for now</h4><p>With few words, a small recognizer gets far more right and does not invent orders nobody gave. It is just enough to obey commands; conversation will require opening it up.</p></div></div>
<!--kg-card-end: html-->
<h2 id="problems-we-ran-into">Problems we ran into</h2><h3 id="a-dead-channel-lost-half-of-the-hotwords">A dead channel lost half of the hotwords</h3><p>Cámara would stop recognizing “oye cámara” now and then, without a single error. The cause: one of the four microphones was dead. Three channels read <strong>−49 to −47 dBFS</strong> and resembled each other (correlation <strong>0.79–0.87</strong>); the fourth read <strong>−79 dBFS</strong> and resembled none of them (<strong>~0.00</strong>): background noise, not a microphone.</p><figure class="kg-card kg-image-card kg-card-hascaption"><img src="https://luf.tech/content/images/2026/10/canal-muerto_en.png" class="kg-image" alt="Real capture of the PS3 Eye's 4 channels: three hear the room and channel 3 stays flat at −79 dBFS" loading="lazy"><figcaption>Channel 3 does not fail: it delivers data, but only hears its own noise.</figcaption></figure><p>A dead channel does not make ODAS fail: it makes it compute <strong>another</strong> direction, and the audio it passes to Vosk points at the wrong place. We tested it with 8 “oye cámara” recorded once and run through all three paths:</p>
<!--kg-card-begin: html-->
<div class="luf-data"><table><caption>The same audio, three paths</caption><thead><tr><th>path</th><th class="num">hotwords out of 8</th></tr></thead><tbody><tr><td>raw audio, one channel</td><td class="num">8/8</td></tr><tr><td>ODAS with all 4 mics</td><td class="num">4/8</td></tr><tr class="pick"><td>ODAS with 3 mics</td><td class="num">8/8</td></tr></tbody></table><p class="note">The fix was one line of configuration: dropping the dead channel. The cost: with 3 mics direction gives the side, not the angle.</p></div>
<!--kg-card-end: html-->
<!--kg-card-begin: html-->
<div class="luf-concl"><div><span class="luf-concl-k">Takeaway</span><h4>A dead channel does not fail, it lies</h4><p>Each channel's health has to be watched: the system keeps running, on false data.</p></div><div><span class="luf-concl-k">Takeaway</span><h4>Compare on the same stimulus</h4><p>Recording once and running the same audio through every variant makes the test repeatable.</p></div></div>
<!--kg-card-end: html-->
<h3 id="the-fan-the-closest-noise">The fan, the closest noise</h3><p>The Pi's fan sits behind the microphones, about <strong>40 mm</strong> away. The Pi drives it on its own at five speeds, from 0 (off) to 4 (full), according to its temperature. With the room silent, the direction Cámara reads drifts toward the fan as soon as it reaches speed 2: from there on, its ear points at the fan and not at whoever is talking.</p>
<!--kg-card-begin: html-->
<div class="luf-data"><table><caption>The fan against hearing, in silence</caption><thead><tr><th>speed</th><th class="num">RPM</th><th class="num">direction read</th></tr></thead><tbody><tr><td>0</td><td class="num">0</td><td class="num">+3°</td></tr><tr class="pick"><td>1</td><td class="num">3200</td><td class="num">+5°</td></tr><tr><td>2</td><td class="num">5500</td><td class="num">+26°</td></tr><tr><td>3</td><td class="num">7500</td><td class="num">+28°</td></tr><tr><td>4</td><td class="num">9200</td><td class="num">+27°</td></tr></tbody></table><p class="note">At speed 1 it does not interfere. From 2 on, background noise rises from −52 to −40 dBFS, with tones at 1.8–2.06 kHz: right where the voice is.</p></div>
<!--kg-card-end: html-->
<h4 id="why-the-fan-speeds-up">Why the fan speeds up</h4><p>To follow a person, Cámara looks for faces in every frame of the camera's video. It is a chain: <strong>more video frames per second → more CPU → more heat → faster fan → more noise</strong>. Looking for faces (YuNet) takes <strong>31 of the ~34 ms</strong> available per frame. While following someone, the Pi reached <strong>64–73 °C</strong>, with the fan at speed 3.</p><h4 id="what-we-tried">What we tried</h4><p><strong>Filtering the noise.</strong> It did not work. Ignoring the fan's direction moves the peak to the edge of what is ignored: a few centimeters away, it does not sound from a point but from an area. Removing its tones did not work either, because it is also a hiss that covers every frequency. Subtracting its footprint recovers the order of the sides, but not the angle.</p><p><strong>Giving it air.</strong> The hot air from the heatsink stayed inside the enclosure, and the fan kept cooling the Pi with that same hot air. We designed and printed two parts: an extractor —a <strong>30×30 mm</strong> hood over the fan with a ~25 mm tube that draws fresh air from the lid window— and an outlet that blows the hot air out through the Sense HAT side, away from the microphones.</p>
<!--kg-card-begin: html-->
<figure class="kg-card luf-model"><script type="module" src="https://luf.tech/assets/js/model-viewer.min.js"></script><model-viewer src="https://luf.tech/assets/model/camara-aire.glb?v=4" alt="Assembled 3D model: the base, the Raspberry Pi 5, the Sense HAT, the PS3 Eye and the lid in gray; the extractor over the fan and the air outlet in pink" poster="https://luf.tech/content/images/2026/10/ensamblado-v3-7.png" camera-controls touch-action="pan-y" auto-rotate environment-image="neutral" tone-mapping="neutral" exposure="0.75" shadow-intensity="0.5" camera-orbit="35deg 65deg auto" style="width:100%;height:460px;background:transparent"></model-viewer><figcaption>Rotate it with your mouse or finger. In pink, the two new parts: the extractor brings fresh air in from the lid and the outlet blows the hot air away from the microphones.</figcaption></figure>
<!--kg-card-end: html-->
<div class="kg-card kg-file-card"><a class="kg-file-card-container" href="https://luf.tech/content/files/2026/10/extractor-salida-aire.zip" title="Download" download=""><div class="kg-file-card-contents"><div class="kg-file-card-title">Air extractor and outlet (STL)</div><div class="kg-file-card-caption">For the Raspberry Pi 5 with Active Cooler: the extractor and the two-part outlet, ready to print.</div><div class="kg-file-card-metadata"><div class="kg-file-card-filename">extractor-salida-aire.zip</div><div class="kg-file-card-filesize">167 KB</div></div></div><div class="kg-file-card-icon"><svg viewBox="0 0 24 24"><defs><style>.a{fill:none;stroke:currentColor;stroke-linecap:round;stroke-linejoin:round;stroke-width:1.5px;}</style></defs><title>download-circle</title><polyline class="a" points="8.25 14.25 12 18 15.75 14.25"></polyline><line class="a" x1="12" y1="6.75" x2="12" y2="18"></line><circle class="a" cx="12" cy="12" r="11.25"></circle></svg></div></a></div>
<!--kg-card-begin: html-->
<div class="luf-data"><table><caption>With and without airflow, with the robot awake</caption><thead><tr><th></th><th class="num">no airflow (closed box)</th><th class="num">with extractor and outlet</th></tr></thead><tbody><tr><td>Pi temperature</td><td class="num">81.5–83.7 °C</td><td class="num">64–73 °C</td></tr><tr><td>fan</td><td class="num">speed 4 (~9900 rpm)</td><td class="num">speed 3 (~7900 rpm)</td></tr><tr><td>Pi slows itself down from heat</td><td class="num">yes</td><td class="num">no</td></tr></tbody></table><p class="note">Some <strong>10–18 °C</strong> lower, but it is not a clean comparison: between one measurement and the other the software also changed (fewer threads for face detection, resting pace). What is certain is that the Pi stopped slowing itself down.</p></div>
<!--kg-card-end: html-->
<p><strong>Fewer frames per second.</strong> With the fan held at speed 1, each lower level cools the Pi by about 4 °C. It works, but following someone with few frames means following them worse.</p>
<!--kg-card-begin: html-->
<div class="luf-data"><table><caption>Frames-per-second cap against temperature, with the fan held at 1 and a person in frame</caption><thead><tr><th>fps cap</th><th class="num">temperature (last minute)</th><th class="num">range</th></tr></thead><tbody><tr><td>8</td><td class="num">68.8 °C</td><td class="num">65.6–72.7 °C</td></tr><tr><td>4</td><td class="num">65.2 °C</td><td class="num">63.4–67.8 °C</td></tr><tr class="pick"><td>2</td><td class="num">61.3 °C</td><td class="num">59.5–62.8 °C</td></tr></tbody></table><p class="note">Each level ran 4 minutes and carries heat over from the previous one, so the numbers err on the high side. None of them made the Pi slow itself down.</p></div>
<!--kg-card-end: html-->
<h4 id="how-we-solved-it">How we solved it</h4><p>Airflow helped, but the noise is born in the computation. The solution was for tracking not to work when there is no need, and for hearing to know when not to trust itself:</p>
<!--kg-card-begin: html-->
<div class="luf-facts"><div class="luf-fact"><span class="luf-fact-v">87 %</span><span class="luf-fact-t"><strong>Look for faces only when something changed.</strong> Before looking, it compares the frame with the previous one; if nothing moved, it reuses what it already found. It skips 87 % of searches with someone still and <strong>94 %</strong> with the room empty.</span></div><div class="luf-fact"><span class="luf-fact-v">4.3 %</span><span class="luf-fact-t"><strong>Of one CPU core</strong> is what looking for faces came to cost, down from <strong>25 %</strong>. At rest, the Pi dropped to <strong>57.3 °C</strong> with the fan at speed 1, the one that does not interfere.</span></div><div class="luf-fact"><span class="luf-fact-v">70 °C</span><span class="luf-fact-t"><strong>Fan later.</strong> We raised the temperature at which the fan goes to speed 2: from 60 to <strong>70 °C</strong>. The Pi is still far from the 80 °C at which it slows itself down.</span></div><div class="luf-fact"><span class="luf-fact-v">reposo</span><span class="luf-fact-t"><strong>Resting pace.</strong> If the person is still or nobody is there, tracking slows down, and a small movement does not speed it back up.</span></div><div class="luf-fact"><span class="luf-fact-v">≥ 2</span><span class="luf-fact-t"><strong>Fan speed from which hearing is not trusted.</strong> There Cámara does not trust the voice's direction and does not turn toward the fan.</span></div></div>
<!--kg-card-end: html-->
<!--kg-card-begin: html-->
<div class="luf-concl"><div><span class="luf-concl-k">Takeaway</span><h4>The hardest noise is the robot's own</h4><p>It is centimeters away, changes with what the robot is doing and does not filter out like room noise.</p></div><div><span class="luf-concl-k">Takeaway</span><h4>Cutting useless compute is filtering at the source</h4><p>Not processing a frame that did not change lowers heat, fan and noise all at once.</p></div></div>
<!--kg-card-end: html-->
<h2 id="how-we-measured-it">How we measured it</h2><p>Everything on the Pi and in the real room. Recordings are made once and reprocessed with each variant, so the comparison runs on the same audio.</p>
<!--kg-card-begin: html-->
<div class="luf-data"><table><caption>The tests</caption><thead><tr><th>test</th><th>tool</th><th>stimulus</th><th>what it measures</th></tr></thead><tbody><tr><td>Clock and level per channel</td><td><code>luf-002 plant mic</code></td><td>known tone</td><td>Hz, dBFS, correlation</td></tr><tr><td>Hotword A/B</td><td><code>hotword-ab.py</code></td><td>8 recorded “oye cámara”</td><td>hotwords out of 8</td></tr><tr><td>Fan vs direction</td><td><code>doa-raw.sh</code></td><td>silence and voice, fan commanded 0–4</td><td>direction, dBFS</td></tr><tr><td>Hotword side</td><td><code>doa-criterio.py</code></td><td>13 real labeled calls</td><td>hits, wrong-side turns</td></tr><tr><td>Temperature by fps</td><td><code>termica-fps.sh</code></td><td>person in frame, fan held at 1</td><td>°C</td></tr></tbody></table></div>
<!--kg-card-end: html-->
<h2 id="what-carries-over-to-other-robots">What carries over to other robots</h2>
<!--kg-card-begin: html-->
<div class="luf-concl-general"><div><span class="n">1</span><div><h4>Computation also makes noise and heat</h4><p>Processing only when something changes lowers CPU, temperature and fan noise all at once.</p></div></div><div><span class="n">2</span><div><h4>Check each channel on its own</h4><p>In a multi-channel sensor, look at each channel's level and how alike they are: a dead one throws no error, it changes the result.</p></div></div><div><span class="n">3</span><div><h4>Record once, compare many times</h4><p>Running the same audio through every variant makes the difference come from the method, not the take.</p></div></div></div>
<!--kg-card-end: html-->
How Cámara hears
Cámara recognizes “oye cámara” offline, knows which side the voice came from and turns there.
·
8 min read