Hey everyone, welcome back to My Weird Prompts. I am Corn, and I am sitting here in our living room in Jerusalem with my brother. It is a bit of a chilly afternoon outside, but we have got the heaters going and a fresh pot of coffee on the table.
Herman Poppleberry at your service. It is a beautiful day to talk about some serious hardware, Corn. I have been looking forward to this one all week because it touches on that perfect intersection of physics, semiconductor manufacturing, and artificial intelligence.
It really does. And today is prompt from Daniel is actually a bit of a reality check for anyone who spends a lot of time thinking about audio gear. He has been doing some deep testing on speech-to-text accuracy, specifically looking at which microphones give the best results when running audio through Whisper. For those who do not know, Whisper is the open-source speech recognition model from OpenAI that really changed the game a couple of years ago.
This is so classic Daniel. He did not just take the marketing at face value. He actually went out and calculated the word error rate—or W-E-R—for a bunch of different setups. He recorded himself reading paragraphs about the history of coffee, which I appreciate, and then ran those recordings through the same Whisper model to see which one produced the fewest mistakes. And the results are, well, they are kind of embarrassing for some very expensive equipment manufacturers.
Right. The big takeaway was that his built-in OnePlus phone microphone actually outperformed a dedicated Jabra headset and a professional-grade lavalier mic. He always assumed, like most of us do, that phone mics are kind of low-quality, tiny components that are "good enough" for a call but not for "real" work. But his data says otherwise. It makes you wonder if the muffled, garbage sound we often hear on phone calls is actually the fault of the microphone at all.
It is almost certainly not the microphone. That is the big misconception right there. We tend to conflate the quality of the hardware with the quality of the signal we hear on the other end of a cellular connection, but those are two completely different things. When you are on a standard mobile call, that audio is being crushed by compression and limited by the bandwidth of the cellular network. But the actual raw capture happening on the device? That is a different story entirely.
So let us pull on that thread. If the hardware is actually high quality, what are we looking at? What kind of microphones are they actually putting inside these tiny glass rectangles we carry around?
Well, we have moved far beyond the old electret condenser microphones you might remember from decades ago. Modern smartphones use what are called M-E-M-S microphones. That stands for Micro-Electro-Mechanical Systems. Imagine a tiny silicon chip where they have literally etched a diaphragm and a backplate directly into the silicon using the same lithography techniques they use for processors.
So it is a mechanical device, but it is manufactured using the same processes we use to make computer chips?
Exactly. It is a masterpiece of semiconductor engineering. These things are incredibly small—we are talking about a few millimeters square—but they are remarkably consistent. Companies like Knowles and STMicroelectronics produce these by the billions. Because they are etched into silicon, you do not have the same manufacturing variances you get with traditional microphones where a human might be assembling a capsule or tensioning a diaphragm. Every single one coming off the line is nearly identical in terms of frequency response and sensitivity.
And is that why they are so good for speech-to-text? Because they are consistent?
That is part of it. But they also have a very high signal-to-noise ratio. In the early days of M-E-M-S, they were pretty noisy, but the technology has matured incredibly fast. Today, a top-tier M-E-M-S mic can have a signal-to-noise ratio of sixty-five or even seventy-two decibels. For something that fits on the head of a pin, that is phenomenal. They also have a very flat frequency response, meaning they do not "color" the sound as much as a studio mic might. They just capture what is there.
Okay, but Daniel mentioned that he was holding the phone about six inches from his mouth. That seems like a big factor. With a headset, the mic is right there at your lip. Why would the phone still win?
This is where it gets into the physics of how headsets are designed versus how phones are designed. A lot of those office-style headsets, like the Jabras Daniel mentioned, are designed with very aggressive noise cancellation built into the hardware or the firmware. They are trying to strip out the sound of your coworkers, the air conditioner, or the coffee shop background. But that processing often introduces artifacts. It can clip the ends of words, create a "swishing" sound, or change the timbre of your voice.
So the "cleanness" of the audio for a human listener might actually be "messiness" for an artificial intelligence model like Whisper?
Precisely. Whisper and other modern speech-to-text models are trained on a vast amount of diverse data—everything from high-quality podcasts to grainy YouTube videos. They are actually quite good at ignoring background noise if the primary signal—your voice—is rich and undistorted. When a headset tries to "help" by filtering out noise, it might actually be removing the very frequencies the model uses to distinguish between a "p" and a "b" sound, or the subtle sibilance of an "s." The phone mic, especially when recording in a relatively quiet room, is giving you a much flatter, more natural representation of the voice. It is not trying to be "smart" in a way that interferes with the AI's own "smartness."
That makes so much sense. It is the difference between a raw photo and one that has been heavily filtered by a cheap app. The AI wants the raw data. Now, Daniel also asked about pickup patterns. Usually, when we talk about studio mics, we talk about cardioid or omnidirectional. How does that work when the mic is just a tiny hole in the bottom of a phone?
Most individual M-E-M-S mics are omnidirectional by nature. They pick up sound from all around. However—and this is the "magic" part of modern phones—they almost never use just one microphone. Most modern smartphones have at least three, and often four, microphones scattered around the body. There is usually one at the bottom for your mouth, one at the top for noise cancellation and speakerphone mode, and one near the camera lens on the back to capture audio for video.
So how does the phone use all those different points of capture?
It is called beamforming and noise decorrelation. When you have multiple microphones, the phone's processor—like the Snapdragon eight Gen four or the Apple A-nineteen—can compare the timing of the sound waves hitting each one. If a sound hits the bottom mic a fraction of a millisecond before it hits the top mic, the processor knows that sound is coming from below. It can then use math—essentially destructive interference—to digitally "steer" the sensitivity toward your mouth and cancel out sounds coming from other directions.
So it is creating a virtual directional microphone out of several omnidirectional ones. That is brilliant. Does that mean the phone is actually changing its "pattern" depending on how you are holding it?
Yes, absolutely. If you are holding the phone to your ear, it uses one profile. If you put it on speakerphone on a table, it switches to a different array configuration to try and capture everyone around the table equally. And if you are recording a video, it might use the mics on the back to focus on whatever the camera is pointed at. This is why Daniel saw such good results. When he is holding it six inches away, the phone is likely using its beamforming algorithms to create a very clean "cone" of sensitivity right where his mouth is, while using the other mics to identify and subtract the ambient room noise.
I wonder if some manufacturers are better at this than others. Daniel mentioned his OnePlus performed well. Are there brands that specifically market their internal mics for high-quality recording?
There definitely used to be more of a focus on it in the marketing. Do you remember the LG V-series? Like the V-twenty and V-thirty?
Oh yeah, those were the "audiophile" phones. They had the high-end digital-to-analog converters for headphones.
Exactly. But they also had incredible microphone arrays. LG actually marketed them for recording loud concerts without distortion. They used microphones that could handle very high sound pressure levels—up to one hundred and thirty-five decibels—which is basically like standing next to a jet engine. Most phone mics would just clip and produce a wall of static, but those LG phones stayed crystal clear.
It is a shame LG got out of the phone business. Who has picked up that mantle in twenty-twenty-six?
Sony is still very serious about it with their Xperia line, particularly the Xperia one Mark seven. They include dedicated recording apps that let you bypass a lot of the standard Android processing and record in twenty-four-bit, ninety-six-kilohertz audio. Apple has also been making a big deal about "studio-quality" mics since the iPhone sixteen Pro. They use a four-mic array that they claim has a much lower noise floor than previous models. In fact, for the iPhone seventeen Pro that just came out, they have integrated a new "Voice Isolation" mode that uses the neural engine to do real-time separation of the voice from the background, and it is surprisingly transparent.
It is interesting that we are seeing this shift. For a long time, the microphone was just a utility for calls. Now that we are all using voice memos, TikTok, and speech-to-text, it has become a primary feature. But let us go back to Daniel's point about the "transport layer." He mentioned that the muffled sound on calls might be due to compression. Can we break that down? Why does my voice sound like a robot from nineteen ninety-five when I call you, but perfectly clear when I send you a voice note on Telegram?
It comes down to the codec—the algorithm used to shrink the audio data so it can travel over the network. Standard cellular calls for a long time used something called narrow-band audio. It limited the frequency range to between three hundred and thirty-four hundred hertz. For context, human hearing goes up to twenty thousand hertz. So you were losing all the high-end "air" and the low-end "warmth." It was designed specifically to make speech intelligible while using as little data as possible, but it sounds terrible.
It is like looking at the world through a tiny, blurry porthole.
Exactly. Now we have Vo-L-T-E, or Voice over L-T-E, and Vo-five-G, which use wideband codecs like G-point-seven-two-two-point-two, often called H-D Voice. Even better is E-V-S, or Enhanced Voice Services, which can handle up to twenty kilohertz—the full range of human hearing. But even with H-D Voice, the network might decide to throttle your bitrate if the signal is weak. When you record a voice note or use an app like Whisper on a local file, you are bypassing the cellular network's limitations entirely. You are getting the full range of what those M-E-M-S mics can capture, which is often surprisingly close to twenty thousand hertz.
So the hardware is the Ferrari, but the cellular network is a dirt road with a forty-mile-per-hour speed limit.
That is a perfect analogy. When Daniel records his coffee history paragraphs, he is letting the Ferrari actually hit its top speed because he is saving the file locally and then feeding it to Whisper. The "muffled" sound we associate with phones is a network problem, not a microphone problem.
You know, this makes me think about the "gooseneck" microphones Daniel mentioned. He said they are under-loved. I have seen them in those high-end conference rooms or at podiums. They look very professional, but he found they did not necessarily beat the phone. Why would a dedicated, foot-long microphone lose to a tiny chip?
It is often about the environment and the electronics behind the mic. A gooseneck is usually a traditional condenser mic. It needs a good preamp and a good analog-to-digital converter to really shine. If you are plugging a cheap gooseneck into a basic computer sound card, you are introducing all kinds of electrical noise from the computer's internals—the fans, the power supply, the motherboard. The phone, on the other hand, is a tightly integrated system. The path from the M-E-M-S mic to the processor is incredibly short and well-shielded.
Plus, the phone has all that dedicated silicon for digital signal processing. A standard desktop computer is trying to do a million things at once. The phone's audio subsystem is specialized for this.
Exactly. And let us not forget the software side. The "drivers" for a phone microphone are tuned specifically for that exact hardware by hundreds of engineers at the manufacturer. When you plug a random U-S-B mic or a gooseneck into a P-C, you are using generic drivers that are designed to work "well enough" with everything, but perfectly with nothing. The phone is a closed ecosystem where every component knows exactly how to talk to every other component.
That is a great point. I want to talk about the "built-in" aspect of this. Daniel found that his phone beat his Jabra headset. Do you think that is a universal truth now? Should people stop buying headsets for dictation?
Well, there is still the comfort factor. Holding a phone six inches from your face for eight hours a day is going to give you some serious "gorilla arm." And if you are in a noisy office, the headset's physical proximity to your mouth still gives it a massive advantage in terms of raw signal-to-noise ratio. But if you are in a quiet home office? Honestly, the phone or a high-quality "puck" style conference mic might actually be the superior choice for speech-to-text accuracy.
It is funny you mention the conference "puck" mics. Daniel mentioned a twenty-dollar one from AliExpress called the E-M-E-E-T, or something similar, that performed surprisingly well.
Yeah, those conference pucks are basically just big housings for a circular array of M-E-M-S microphones. They are using the same tech as the phone, but because they have more space, they can spread the mics further apart—maybe four or five inches apart instead of just the width of a phone. That actually makes the beamforming math even more effective. A wider "base" for your microphone array allows for much more precise directional targeting. It is like having a wider set of eyes—it gives you better depth perception, but for sound.
Precisely. It is called spatial filtering. By having mics on opposite sides of the puck, the device can very accurately nullify sound coming from the sides while focusing on the person speaking from above. It is basically a specialized version of what your phone is doing, but optimized for a table-top environment.
And because those pucks are often U-S-B devices, they have their own built-in sound card, which bypasses the noisy internals of your computer. So you get the benefit of M-E-M-S consistency, array-based noise cancellation, and a clean digital signal. For twenty dollars, that is a lot of engineering.
I am curious about the future of this. If phones are already this good in twenty-twenty-six, where do we go from here? Are we going to see phones with ten microphones? Or is the improvement going to be all on the A-I side?
I think we will see a bit of both. We are starting to see "on-device" A-I processing that is specifically tuned to the microphone hardware. Instead of just sending a raw audio stream to Whisper, the phone might use a small, local neural network to "clean" the audio in a way that is specifically optimized for a larger model. It is like a specialized pre-processor.
Like a translator for the translator.
Exactly. And on the hardware side, there is research into "optical" M-E-M-S microphones. Instead of measuring the change in capacitance on a silicon diaphragm, they use a tiny laser to measure the vibrations. That would effectively eliminate all electrical noise from the capture process. You could have a noise floor that is basically zero. We are also seeing "V-P-U" or Voice Pick-Up units that use bone conduction. They sit against your skin and pick up the vibrations of your jawbone, which is completely immune to background noise. Some high-end earbuds are already using these to supplement the M-E-M-S mics.
A laser-mic in my pocket and bone conduction in my ears. We really are living in the future, Herman.
We really are. But I think the most important takeaway for Daniel, and for all of us, is to stop underestimating the engineering in these devices. We tend to think "bigger is better" with microphones because we see professional singers using these massive, heavy tubes in studios. But for the specific task of "turning human speech into text," the goals are different. You do not need "warmth" or "character" or that "vintage tube sound." You need accuracy, consistency, and a clean signal. And silicon is very, very good at that.
It is a classic case of the right tool for the job. A vintage vacuum tube microphone might make your singing voice sound like velvet, but it might actually make it harder for an A-I to figure out if you said "can" or "can't" because it adds harmonic distortion.
Exactly. The A-I does not care if you sound like Frank Sinatra. It just wants to see the distinct frequency peaks of your consonants. It wants a high-fidelity representation of the acoustic pressure waves, and a modern M-E-M-S mic is a precision instrument for measuring those waves.
You know, I actually tried this myself after reading Daniel's notes. I recorded a bit of a script on my old desktop mic—a big, impressive-looking thing on a boom arm—and then did the same on my phone. When I ran them through a basic transcription tool, the phone version had fewer errors in the technical terms. It was subtle, but it was there.
It is that "flat" response. Professional studio mics often have a "presence boost" in the upper-mid frequencies to help a voice cut through a musical mix. But that boost can distort the natural balance of the speech sounds. The M-E-M-S mic in your phone is designed to be as flat as possible because the engineers know they can always add "flavor" later with software if they want to. For A-I, flat is king.
So, if you are a listener out there and you have been struggling with your dictation software, the answer might not be a three-hundred-dollar headset. It might just be the device you are already holding. But Herman, there has to be a catch. What is the downside of using the phone?
Handling noise. That is the one area where phones really struggle. If you are moving your hand around on the glass or the frame while you are talking, those vibrations travel directly into the M-E-M-S chip. Because the chip is physically part of the phone's structure, it picks up every little scrape and tap. It sounds like a series of earthquakes to the microphone.
Right, Daniel mentioned he was trying to keep it as steady as possible. So, the pro-tip is: prop your phone up on a stack of books or a small tripod, keep it about six to eight inches from your mouth, and you might just have the best transcription rig in the world.
It is funny, we spend all this money on gear, and the best solution is often the one we have already paid for. It is just about knowing how to use it. And maybe turning off all the "enhancements" in your recording app so you get that raw, flat signal.
That is really the theme of so many of our discussions, isn't it? The technology is often ahead of our understanding of how to apply it. We are still using nineteen-eighties mental models for twenty-twenty-six hardware. We think a "real" microphone has to be big and heavy, but in the world of A-I, the tiny silicon chip is the heavyweight champion.
Guilty as charged. I still catch myself looking at big microphones and thinking, "Ooh, that must be better." I have to remind myself of the physics. The diaphragm in an M-E-M-S mic is so light that it can respond to changes in air pressure with incredible speed and precision. It has very little "inertia" compared to a large-diaphragm studio mic.
Well, I think we have thoroughly deconstructed the smartphone microphone. It is a silicon marvel, it uses math to "see" sound, and it is probably being held back more by your cellular provider than by its own hardware.
And if you are recording for A-I, "raw and flat" beats "processed and pretty" every single time.
That is a great mantra. Raw and flat beats processed and pretty. I like that.
I might get it on a t-shirt.
Please don't. One Poppleberry in a technical t-shirt is enough for this house.
Hey! This shirt is a classic. It has the original schematic for the five-five-five timer on it!
If you say so, Herman. If you say so. Anyway, I think that covers the bulk of Daniel's prompt. It is a fascinating look at how our assumptions about quality are often built on old information.
Absolutely. And it is a great reminder to actually test things. Daniel didn't just wonder; he measured. He calculated the word error rate. That is the gold standard for any kind of technical inquiry.
The Poppleberry Seal of Approval for Daniel's methodology.
Definitely. He even accounted for the "transport layer" by testing different recording methods. That is a high-level catch.
Well, before we wrap up, I want to say a big thank you to everyone for listening. We have been doing this for six hundred and seventy-one episodes now, and it is the curiosity of people like Daniel—and all of you—that keeps us going.
It really does. We love diving into these rabbit holes with you. If you are enjoying the show, we would really appreciate it if you could leave us a quick review on your podcast app or on Spotify. It genuinely helps other people find the show, and we love reading your feedback. Even the technical corrections! Especially the technical corrections.
Yeah, it makes a huge difference. You can find My Weird Prompts on Spotify, Apple Podcasts, or wherever you get your podcasts. You can also visit our website at myweirdprompts dot com for the full archive and our R-S-S feed.
And if you have a prompt of your own—something that has been bugging you or a weird experiment you have been running—send it our way! You can reach us at show at myweirdprompts dot com. We might just feature it in a future episode.
Thanks again for joining us in Jerusalem. This has been My Weird Prompts.
Until next time, keep testing those assumptions. Goodbye!
Goodbye!
