NAMM 2026: Sounding out the inputs with klang.io’s Sebastian Murgul

We’ll be publishing coverage from the 2026 NAMM Show in Anaheim, California. The show has events running from January 20-24, 2026, and floor exhibits active from January 22 through January 24. It’s a huge exhibition, so we’ll focus on what we do best: covering the field of music notation software and related technology. Follow all of our NAMM 2026 coverage at Scoring Notes.

---

---

At the 2026 NAMM Show, I sat down with Sebastian Murgul, co-founder and CEO of klang.io, to talk about a category that sits just adjacent to music notation — and yet increasingly intersects with it in practical, unavoidable ways: music transcription.

On Scoring Notes, we spend most of our time discussing tools that help musicians write music. klang.io approaches the problem from the opposite direction, starting with audio and working backward toward notation through AI-based transcription. That difference in starting point raises a distinct set of questions — not just technical ones, but musical and philosophical ones as well.

For many musicians, transcription has always lived in a space between art and approximation. Performances are expressive, messy, and context-dependent; notation, by contrast, asks for clarity, intent, and structure. Bridging that gap is not simply a matter of better signal processing, but of musical judgment: deciding what matters, what should be simplified, and what a human will need to shape by hand.

In this conversation, Sebastian explains what klang.io’s tools are designed to do — and just as importantly, what they are not. We talk about why AI-based transcription has reached a point of practical usefulness now, where the hardest musical problems still lie, and how klang.io thinks about accuracy as something musicians can trust and build on, rather than a promise of perfection.

We also discuss interoperability with notation software via formats like MIDI and MusicXML, real-world use cases that have surprised him, and the broader anxieties musicians understandably have around automation and AI. Throughout, the focus remains on transcription as a starting point, not an endpoint — a way to accelerate human work rather than replace it.

As with all Scoring Notes interviews, the goal is not to predict the future in grand terms, but to better understand how emerging tools fit into real workflows today — and what values guide their development.

A transcript of the conversation follows, edited for clarity.

---

Scoring Notes interview with Sebastian Murgul at NAMM 2026

Philip Rothman:
Hey everyone, it’s Philip with the Scoring Notes website and podcast, and I’m back once again at the 2026 NAMM Show. And now I am having the great pleasure of speaking with someone who is not necessarily a music notation software developer, but someone who is working a lot with the types of tools that people that use music notation software will use — or perhaps are already using — and that would be transcription tools.

It’s my great pleasure to welcome to Scoring Notes Sebastian Murgul. He’s the co-founder of klang.io, and they are an AI transcription tool — and a lot more that we will talk about, I’m sure. But first of all, Sebastian, it’s great to see you. Great to meet you here at NAMM this year. Welcome.

Sebastian Murgul:
Thank you for the nice heartwarming intro, Philip. It’s great to be here. It’s great to talk to you a little bit more about music transcription, what we are doing — and especially in terms of notation. And I’m really looking forward to our talk. I’m really excited.

Philip: Me too. And the first thing, I just want to make sure: what is the correct way to say the name? Did I say it right? Is that preferred — klang.io?

Sebastian: It’s the German way — that’s the most accurate way you can do it. Internationally we say klang.io, because klang — K-L-A-N-G — is easier to spell. And that’s also our domain, where you can find us on the internet. So yes, you’re happy to say klang.io.

Philip: Okay, got it. Excellent. Of course, klang meaning sound, and io meaning input/output.

Sebastian: Input/output.

Philip: There you go.

Sebastian: Basically what we’re doing is processing audio and doing some stuff with it.

Philip: Yeah, yeah — there you go.

So now that we know what it is, how to pronounce it, and where to find it, let’s set the context about why we’re doing transcription at this moment in time.

As I just said, we talk primarily about music notation software on Scoring Notes, but much less about transcription technology — partly because it hasn’t really been that great until fairly recently. So from your perspective, why is now the moment where we see AI-based music transcription becoming viable?

Sebastian: The idea of music transcription has always been around. Since the 1960s we had the first algorithms that were able to do some spectral analysis — things like the fast Fourier transform. In the 1970s, we had the first official research paper that defined the task of automatic music transcription, which kicked off researchers all over the world trying to figure out how to find musical details from audio and how to notate them.

This was all before MIDI. MIDI came in the 1980s, and then with the rise of DAW software, there was a strong motivation to convert audio into a MIDI representation so people could work with it digitally.

But those approaches were very limited, because they relied entirely on classical signal processing. I’m an electrical engineer myself. I studied in Germany at KIT, and we learned all the basics — Fourier transforms, filters, and other signal-processing techniques that were used back then to create MIDI representations from audio.

But it wasn’t really possible to get useful results until the early 2000s, when machine-learning approaches became practical. And of course it depends on the complexity of the audio.

For example, if you just have a piano melody with one note at a time, that’s a fairly simple problem — and that was already possible in the 1980s. But real recordings — real music performed by real humans, in a hall, with reverb, background noise, tempo variation — all of that completely confused those classical approaches.

What changed was not only machine learning, but deep learning, which really started to rise in the 2010s, driven mainly by computer vision — for example in autonomous driving. Music transcription developers started using those technologies, not on images, but on spectrograms — images of audio.

That’s when everything started to accelerate and get faster and better. And now we’ve reached a point where we can not only transcribe a simple piano melody, but a full band arrangement, using the computing power you have at home. I’ve been doing research in this field for 11 years now, and I’ve seen the rise of deep learning. At the beginning it was very limited, but now we’re getting very usable results.

So now is the time.

Philip: Now is the time.

That’s a great overview of the history of automated music transcription. I have about a thousand questions just based on what you said, but I want to keep the focus on your product — otherwise we’ll go down this rabbit hole, and perhaps we will at another point.

But one thing that you did mention that particularly resonated with me is the fact that you can do these sorts of things in an academic vacuum, so to speak. Like, you can say: “Oh yes, I can analyze in a perfect controlled environment, with no extra noise, on monophonic instruments, and played perfectly, with no variance of timing or tempo, or mistakes, or anything like that.”

Then perhaps that is the ideal environment, and that’s where we’ve seen the music transcription tools be okay. And also OMR, which is another topic entirely.

Sebastian: Yeah.

Philip: But when you introduce the variability, we’re talking about OMR. Of course there’s splotches on the page, there’s this, there’s that.

So anyway, I think that is a great way of thinking about it, and it makes perfect sense why—until relatively recently, with deep learning and the acceleration in computing—we’re able to…

Sebastian: Maybe one short, minor thing to add, because what’s the main difference between the approaches that we used back then and now? Back then we tried to teach the computer by giving instructions on how to process the audio—and it’s so complex that we can’t formulate that in code. And now we can use a few million examples, show that to an algorithm, and it learns the connection between audio and sheet music, which is the key.

Philip: Yeah. Yeah. Before, you were totally relying on whatever you loaded into the system. For better or for worse, that’s what it had. It would run its checks against.

So let’s talk about your product a little bit more specifically. What would be your way of describing what your tools do—and just as importantly, what they don’t do?

Sebastian: Yeah—that’s important.

What we are doing is building AI-based apps for automatic transcription of audio into sheet music—but not only sheet music. We can output different abstraction levels. For example: MIDI—quantized and unquantized.

And for instruments which have some special needs—for example the guitar—we’re not only interested in finding out the pitches of the tones that are played, but we’re also interested in the way they are played on the fretboard. So speaking of the string/fret positioning: we can also transcribe a tab representation. And that can also be exported as Guitar Pro.

And basically the idea is: we are using AI to create a quick transcription with good results. So it’s not 100% perfect—of course. That’s always the case where the human transcriber can stand out, and it’ll always be like that, because adding this additional human layer solves the last 5%, which are really the hardest.

What we instead give is a good transcription that you can work with—that you can use to understand more about the music you’re creating, or understand other people’s music better, and be able to recreate it on your musical instrument.

So that’s what we’re doing, and we’re doing that for multiple different instruments—for example piano, of course, guitar, vocals, drums—which is also very interesting—brass instruments, string instruments in a solo setting, but also in a multi-instrument setting with multiple instruments playing at the same time—for example in a classical way, with up to 13 classical instruments that can play at the same time.

And we split all these tracks and create parts for each instrument in one big, beautiful score.

Philip: Okay.

Sebastian: And our attempt is not to create just a MIDI and just to display the MIDI in sheet music with all of these inaccuracies that you get—but to create readable and playable sheet music, which is very important.

Philip: So talking about readable and playable sheet music—when a performer’s interpretation diverges significantly from where that clean notation might suggest it would otherwise be realized—how do you think about that when you are separating… Oh yeah. I know, as a human—or as a musician—what that musical intent is, and then separating the musical intent from the performance nuance when you’re training your models.

Sebastian: Yeah, that’s a really interesting task because humans and computers are very different from each other in the perception of music.

For example, for a machine-learning transcriber—an AI transcriber—it’s easy to distinguish the exact starting point, in seconds, from a note, and the end point, and the exact pitch. And for a human that’s hard.

But the other way around: when it comes to rhythm, it’s easy for a human to abstract the rhythm and to find out what is the rhythmical intention of every note—what is the quantized length that a note should have, with all these tempo fluctuations that a performance might have.

And teaching that to an AI is really, really difficult. So you’re not only telling the AI to figure out what is played in the audio, but how it is meant. And that’s very, very difficult, and you need a lot of examples—and you need to go one step further than just describing audio that you hear.

Philip: Right. Well, it sounds like it is very, very difficult—and it’s rarely unambiguous what the result is. You could interpret the results in so many different ways, and they could all be played and they would sound more or less the same by a human player, if they were notated by a human player. You know what I mean?

So how do you think about accuracy—not just as perfection, but as something that the musicians can trust and build on? Because even if I prepare 99% of the notes correctly, that 1% is a pretty bad error rate. You know? And that’s a wrong note, and then you have to infer it from context as the player.

And I think when we’re talking about AI models with respect to spoken text—spoken transcription that then gets converted into text—if it misses a word or two or three or four or ten, or misspells something… when you read something, you can infer what it is much more easily.

The difference with music is the temporal aspect of it—because if you make that mistake and it’s in real time and there’s no way of going back and doing it again… Whereas if you’re just reading something, your brain can forge the connections.

So what I’m trying to ask is: the accuracy question again—going back to the ambiguity of it—can the user, can someone that uses klang.io, trust the results?

Sebastian: Yes. So the topic of accuracy is really, really interesting.

I’m about to wrap up my PhD in automatic transcription of acoustic guitar signals—not only to sheet music, but to tablature. Especially in tablature, we have a huge problem with ambiguity, because there are multiple ways to play the same note on the fretboard.

And of course there’s a difference in the sound—in the timbre—and there are slightly pitch deviations because you’re bending the string. But in the end, there are also some playing preferences of the performer.

And evaluating such a fingering-position system is not that easy, because of course there’s the ground truth that you can define—so it’s meant to be played like that—but is it the only way? No, most of the time not. There are multiple ways, and maybe there’s a better way that is not present in your ground truth data.

And in that case, often we see in these studies that human musicians are asked to rate the sheet music and to give a human score. Of course that’s not objective—it’s quite a subjective evaluation—but in my opinion that totally makes sense, because it is hard not only to create the transcription, but also to evaluate the transcription and to give a hard number.

And that’s also where we see in research that there are these systems that get like 99.9% F1 score or accuracy—but in the end, in a practical use case, they perform much worse. And that’s something that we’re also taking account of. We’re using human interpretation and accuracy measurements to ensure that we get good results for you.

Philip: Okay. So—you said you’re finishing up your PhD right now, but you’re also running a company. You and your co-founder and your team have an active company right now: you have users, paying customers, and all of that. First of all, I’m thoroughly impressed that you’re able to juggle all those balls and keep them all going.

At the same time, that means if you have a product that has customers, you have customer feedback. So who are your users right now? Are they educators, arrangers, composers, publishers… and what type of feedback are they giving you, and does it affect your product decisions?

Sebastian: Absolutely. Basically it’s all of the user groups that you just said—but also amateur musicians. And that is a user group that we hadn’t thought of at the beginning. At the beginning we thought: “Hey, music transcription is cool—it’s something special, something niche.” But we found out there’s a huge market evolving of musicians—hobby musicians—that want to either find out and understand what they are playing, or find out the notes of a specific version of a song—specific performance of a song—and want to learn to play exactly that. And often you don’t find the sheet music for that.

And this is actually our largest user group. And this means that our apps have to be that simple to use.

So we have a lot of content that explains the basics of music theory that you might need in order to get the last few steps. But still, we added some tools—for example, we have a sheet music editor in our software, and it has some easy-to-fix tools for, for example, fixing pickup-bar detection failures, or if there’s a wrong splitting between the left and right hand for the piano use case. And that’s something that we had to make easily fixable, because the user isn’t a professional in many cases.

But on the other hand, we have a lot of teachers using that in education—for creating examples, exercises for the students. We have music producers, composers that want to write down their own stuff. And we also have transcribers that are doing music transcription as their daily business, using us to create the first transcription that they can then finalize and sell it to their clients.

Philip: Wow. So you have a user base of amateur musicians that perhaps… you know, music that doesn’t exist otherwise, and they’re trying to get it into a notated format.

But there are also a lot of musicians out there that see this as something that could integrate in their workflow, including people in our world that are publishers that do create published sheet music and are very much not amateur musicians.

So: klang.io exports MIDI and also MusicXML.

Sebastian: Exactly. And even LilyPond for the total nerds.

Philip: Oh, wow. Okay. LilyPond makes an appearance at NAMM, folks. Take note of that one.

So we have those formats that basically slot directly into music notation software. How important is that to your overall vision, and how closely are you following the developments in notation platforms—the kind of ones that we cover on Scoring Notes?

Sebastian: It’s very important, because the transcriptions that we create—they shouldn’t be just dead wood. Just engraved paper. They are digital scores and they should be used to create something new—to do something with it.

And that’s why it’s important to be compatible with all these major notation tools—Sibelius, Finale, Dorico, you name it.

Philip: It’s very interesting—even tools that, we were talking with Piascore, and their interactive sheet music product brings in MusicXML. But then, just thinking as we’re talking right now: somebody has audio, they need to transcribe it, and then they want to be able to play it back and see if they’re playing it back accurately. That is a possible workflow.

Sebastian: Exactly. Exactly. Yeah. It’s very interesting. It’s a very creative task for humans to solve.

Philip: Yeah. Well, we were talking about OMR a few moments ago, and talking about the publishers that use these tools, but there’s still a human element.

So: AI transcription not as a replacement for human work, but as a starting point. How would you say klang.io fits into that workflow?

Sebastian: That’s exactly the idea. Hearing out the notes from an audio track is very tedious, and it takes even professionals a lot of time.

On the other hand, making the real artwork—the engraved sheet music—is something which requires a human touch in the end. And I envision klang to be exactly in that spot: to do the first transcription, audio to sheet music. And then the final product—the beautifully engraved sheet music—is done by the human transcriber, or notator.

So you get sheet music that you can even sell.

Philip: Right. Yep.

So we’ve described a few different use cases. Is there anything in the real world so far that has surprised you about how people are using klang.io, or what they’re expecting from it?

Sebastian: Yeah—we are around since 2018, so we’ve seen a lot of interesting stuff.

For example, we had a composer from LA reaching out to us a few years ago, and he had like 3,000 recordings of himself playing the piano. They were taking over 30 years or so, and he wanted to finally create an album from that.

So he reached out and asked: “Hey, can you do that to MIDI?” We said: “Hey, okay—it’s a pain to upload 3,000 files to our software. Just send us a Dropbox link and we do that for you, and you get the MIDI.” And then he received it, was happy, created his new album with it. And that’s really cool.

But I actually heard a very cool use case on my first day at NAMM: a young man came to me—18 years old—and he told me that he’s using our Piano2Notes app for three years now. He was playing around with the piano, and then he got interested in figuring out what he’s playing, notation-wise. And he tried to do some more stuff, experiment with his music, and somehow got into jazz piano, and now he’s releasing his first jazz… And that was such a heartwarming story, actually.

Philip: Yeah. That’s amazing.

Well, that actually leads me to my next question, because you just described the human element of all of this, and I think there’s some understandable concern out there about AI eroding creative control among musicians—or livelihoods. What’s your response to those concerns, from the perspective of someone that’s actually building those tools?

Sebastian: First of all, we draw a line when it comes to the creative element. We say: we don’t want to build any tool that replaces human creativity. We want to enable humans to be more creative.

So our goal is to reduce the tedious work—all this stuff that, of course, it’s very important to learn transcribing by ear; it really helps your musical career. But when you’ve done that multiple times already, it might not be necessary to do it every time yourself. And that’s where we come into play and help you save a lot of time.

Also save a lot of money when you compare it to sending it out to human transcribers.

But still—I think I have another take. If you think of AI replacement, it happens kind of everywhere. Not only in the music industry—also software development, where programmers get replaced by AI. And I think it shouldn’t be like that. The human should learn to work with these tools and find a way to integrate that in their workflow, and to get even better in what they’re doing. Because it just doesn’t make sense to do it themselves if there’s a tool that does it.

Philip: Look—I see this as a continuum. It’s accelerated now, and this is a transformative moment, but it’s still a continuum of tools being developed and workflows being changed. Certain occupations being replaced or displaced or changing. The work of a whole room of people can be done by a computer now. That’s the nature of technological progress.

So: are there aspects of music notation or transcription that you believe should never be fully automated?

Sebastian: That’s a very interesting question. Lemme think about it.

I would say it totally goes to the last final step. I already mentioned that sheet music is not only recorded notation, but also an artwork itself. So you’re not only telling the music that should be played—you’re also telling a story. And I think doing this step—converting the raw sheet music to something beautiful, art—that’s the human touch. That’s always a thing that we cannot and should not replace by AI.

Philip: Right. I think that will reassure a lot of people.

If you can look down the road: you started in 2018—so that was eight years ago. Let’s look down the road another eight years. What is your hope that musicians will think klang.io made easier—or even possible—that wasn’t the case before?

Sebastian: We’ll totally go for multi transcription, because that’s the complex stuff that I really try to solve. I really love to solve that problem—not only in a kind of limited way, by just saying it has to be up to 13 instruments, or it can only be one violin at once.

Solving symphonic or orchestral transcription would be… wow. I think that’s the holy grail of music transcription. And if AI would be able to do that one time, it would be amazing. I think then we have finally made it.

Philip: Well—so that leads me to a question. In order to transcribe something accurately, don’t you need to know how to create it accurately?

Sebastian: Yeah. Okay. Exactly. We have to learn a lot from composers, and from the transcribers.

Philip: So if that’s the case, do you think your product—or some other product that does transcription—will also be able to create a symphonic piece? No, I’m serious.

Sebastian: No—yeah.

Philip: You know what I mean? Because if I transcribe as a human… it stands to reason I would also know how to create that music. Maybe some transcribers aren’t amazing creative musicians with original ideas, but they know the mechanics, they know the tools. If you ask a transcriber: “Create something else that is similar in style and form and timbre and texture and instrumentation to what you just transcribed”—someone could do that.

So it stands to reason that an AI could do that as well.

Sebastian: I would say absolutely. And it’ll be able to do that.

Philip: Yeah.

Sebastian: But humans will say it sounds boring.

Philip: Which is an interesting philosophical discussion, because a lot of the generative AI that has been rolled out pretty recently—you could say, maybe subjectively, but still—is appealing. Whether it’s artwork or music or whatever: is it boring? Is there an unexpected human element…?

Sebastian: That’s the big philosophical question. If it comes with something really, really new that you wouldn’t be able to expect.

So technically, when we look back to techno music, for example—which of course is algorithmically generated, somehow controlled by a machine—still it can have some element of surprise. That can happen.

Music generation too, with AI.

But I think it’ll be hard to steer them to say: “Okay, create me like 10 happy accidents in my five-minute track.” I’m not sure if that works.

Philip: Okay. All right. Fair enough.

It offers a lot of food for thought in terms of where this is going and how your product’s going to be developed. I think it’s fascinating, and I think before too long—if people aren’t using it already—they’ll be using these tools in conjunction with music notation software, as we talked about earlier.

But for the moment: Sebastian Murgul—this is your first time at NAMM, is it?

Sebastian: It’s my first time. Yeah. And I absolutely love it. It’s not only my first time at NAMM, it’s also my first time in the US.

Philip: Oh, great.

Sebastian: It’s amazing. I met so many friendly people. It’s so cool. Everyone is so open. It’s just something that you’re not used to as a German. And I saw a lot of nice music stuff here at NAMM—so crazy.

I myself play the guitar, but I’m also a synthesizer nerd, and there’s so much stuff for that.

Philip: Absolutely. That’s terrific. Welcome to NAMM. Welcome to the country. We hope it’s your first visit and not your last.

Sebastian: For sure.

Philip: We’ll stay in touch and cover your developments over the time to come. But until then: Sebastian Murgul from klang.io—thank you for taking the time to talk to us on Scoring Notes, and best of luck.

Sebastian: Thanks for having me.

Philip: My pleasure. Take care.

Sebastian: Yep.

---

Keep coming back to Scoring Notes for more coverage from the 2026 NAMM Show.