Beyond Four Instruments: AudioStrip's Quest for Perfect Source Separation
Basil Woods and George Spooner Share Insights on the Challenges and Innovations in High-Quality Stem Separation Beyond Four Stems.
Every seasoned music producer has faced that heart-sinking moment: a lost project file, inaccessible stems, or a sample that just won't sit right in the mix. Enter source separation, a groundbreaking technology that's solving these age-old challenges.
The ability to deconstruct an audio master into its individual stems is a foundational element of Music AI. This technology is reshaping music production, unleashing new creative potential and offering innovative solutions to longstanding challenges. It's with great excitement that we journey into a conversation with the team at AudioStrip, pioneers at the forefront of high-quality stem separation technology. Their mission? To push beyond the conventional four-stem limit and unlock new areas of audio manipulation.
For those new to the company, here's the backstory of AudioStrip and its cutting-edge technology.
“We specialise in high-quality music-stem-separation technology. Our ambition is to achieve perfect, automated stem separation for an arbitrary number of instruments. We are currently trusted by over 40,000 monthly users, including multi-platinum producers, award-winning DJs, and BAFTA member composers. On the B2B side, we are integrated into a variety of external services such as music licensing platforms, gen music AI services and audio-plugins.”
Thanks for taking the time out to talk today, let's start off with a bit of background about yourself. What got you started in music AI?
BASIL: I studied mathematics at Cambridge University, where I earned both a bachelor's and a master's degree. Since graduating, I have used my mathematics background to pursue a career in AI. I have worked as an AI scientist and engineer, among other roles.
I’ve always loved making music and just love music in general. In my free time, I was thinking, what's a way of combining my skills in AI science and engineering with my passion for music, knowing I'm not a very good musician, but I'm a good engineer. AudioStrip came from an idea I had as a hobby music producer and hobby DJ, finding that there weren't any easily accessible ways to generate STEMs. So, it kind of grew out of that problem I had. That’s my background and how I got started.
GEORGE: Basil and I met at Cognizant, a global systems integrator, while we were working on a government AI project. On my side, I was dealing with stakeholder management, speaking to the senior leadership team, and making content.
Consulting is a lot like B2B sales, where you have to convince people it's a good idea to keep you on and maintain engagement. I was very much on that side of things, but it became apparent to me that Basil was very good technically. I think we both appreciated each other's strengths while we were working on that.
Fast forward a few years, and Basil approached me when I was working with Boston Consulting Group. In any startup, there is an inflection point where, as a founder, you realise you need to start delegating stuff. Basil reached out to me and said he needed some help, and the relationship grew from there.
There are numerous instances of startups successfully collaborating with academia. AudioStrip has a partnership with the world-leading C4DM at Queen Mary University of London. Could you elaborate on how this collaboration was initiated and how it has enhanced your research approach and team?
BASIL: From the start, we've really been interested in improving the quality of separation with AudioStrip and seeing how we can go beyond what already exists out there. From early on, I've been speaking to academics and researchers, including quite a few researchers at Queen Mary’s, as a way to trade notes and get advice on potential ways we could improve our algorithms or processes.
Through networking and speaking to a bunch of academics, I eventually spoke to a lecturer at C4DM at Queen Mary, and he seemed to really like what we're doing at AudioStrip. Organically, a relationship formed as we kept talking, and eventually, it grew into a partnership where we've been working on projects together. Most recently, the partnership between us and Queen Mary’s was awarded an Innovate UK grant.
It’s been an organic relationship that I've built purely out of my interest in improving the technology as much as possible and trying to speak to as many experts as possible. It’s been a very good relationship, giving us access to top talent from C4DM. We’ve been able to trade ideas, and there has been great knowledge transfer between their team and ours. They have helped us structure our R&D efforts with their experience in the academic and research world. I’m really enjoying working with them.
GEORGE: Our collaboration with C4DM is going to supercharge our functionality and maintain our position as one of the best STEM isolations tools on the market.
In a recent discussion with Music AI companies, the emphasis was on the critical importance of high-quality source separation when working with existing catalogues for adaptive and generative music. Given your specialisation in this field, could you share insights into why achieving high-quality source separation poses challenges and offer strategies to overcome them?
BASIL: I think the biggest problem in source separation at the moment is going beyond four stems at high quality. A lot of the solutions and research in the academic world have been focusing on four stems: vocals , bass, drums, and other. Although businesses and some academics are starting to explore how we can separate more instruments, I don't think anyone has convincingly achieved high-quality separation of more than four instruments yet.
What we're trying to do is explore how we can get models and algorithms to work on more than four stems. Without giving too much away, some areas we think might have promise are how we define and process our data before feeding it into a model and also exploring new architectures.
How are people currently utilising AudioStrip, and could you provide any examples of interesting or unexpected use cases?
GEORGE: There was one in particular where they gave us something very old, very grainy. You can imagine it was probably ripped off a vinyl. And lo and behold, the vocal came out very clean minus all the backing track. It sounded very peculiar and a bit scary because you've never heard that original clear voice before. I think it's safe to say that we've got very good tech. Basil has made something absolutely fantastic, and people are using it for things like that, both in licensing and beyond what we initially thought people would use it for. There are use cases coming out of the woodwork all the time. We get approached by different organisations for all kinds of projects.
BASIL: One very interesting use case that blew my mind was a couple of individuals who subscribed to our services that are completely deaf. They lost their hearing, and they say they're using our tech to regain an understanding of music, essentially relying on vibrations. If they can separate out the drums or the guitar, they feel they can connect more deeply with the music. They understand, "Oh, this is what the drum beat is like." "Oh, this is the guitar."
I found this really interesting because we wouldn't typically think that deaf people could enjoy music through our technology.
An example of AudioStrip in action:
Master: Medicate Me By The Night is licensed under a Creative Commons License.
Separated Instrumental
Separated Vocal
Are there any accepted "truths" about AI's role in music creation that you disagree with based on your discussions with musicians? What's your contrasting viewpoint and evidence?
BASIL: Yeah, so obviously in the music space, a lot of people seem to be anti-AI or take a lot of convincing as to how it can be used in this space. It is a valuable tool for musicians, not in the way that they may think. Lots of people are talking about AI generating music as good as your favourite band, DJ, or producer, and therefore musicians no longer have a job.
I think actually the more interesting way that AI is making an impact in the music space is using it to take away tedious or difficult tasks, or lowering the barriers of entry for some people who just want to get into music but may find some things challenging. Like source separation for cleaning damaged audio or retrieving stems that were thought to be lost. Also, auto mastering, where many people don't really understand mastering or what it entails and would rather not do it.
From my conversations with musicians, many of them encounter creative blocks where they struggle to come up with ideas. It might be interesting to use AI to generate ideas that musicians can use as a starting point for creating something. I've found that many really creative musicians enjoy the challenge of placing constraints on themselves to see what they can create.
So I think there could be something really interesting there if AI can be utilised in that way—providing an interesting constraint and exploring how to create something unique.
GEORGE: Just to add to what Basil said, in the industry—I'm primarily working in consulting. In the industry that I work in, it's a very good starting point, like what Basil was saying. So often, I will either take an approach I've got, run it through AI for a refined version, or have it critique my work. Sometimes, it's like, "Hi, I don't know anything about this," So let's say, in a music track, you want to add a new instrument or a synth that you haven't used before. I might just say, "Add a synth over the top and see what happens." It gives you a good starting point to then refine. Ultimately, it's like the world of work, right? There will be purists in the music industry, don't get me wrong, but I do think that we're moving towards a world where if you aren't using it, you will get left behind. You're pitting yourself against people who are enabling themselves using this technology.
How do you see the future of AI in the music industry evolving in the next 5-10 years? Where do you envision AudioStrip in that future?
BASIL: AI will be more commonplace for most musicians in the future. As I kind of said before, it'll be used in their workflow to speed up the process, take away tasks that they don't want to do or don't know how to do. A lot of people are saying people will stop making music in the future if AI keeps continuing, but I think actually more people will be making music.
Just like I mentioned before, AI will be a tool to help musicians create music rather than just using AI to generate tracks that we're all listening to all day. I don't think AI music is going to become like popular music. That's my opinion.
And then in terms of AudioStrip, I guess we're focusing quite hard on stem separation. We're aiming for getting perfect studio quality separation on an arbitrary number of instruments. Again, addressing that problem where people generally want more than just vocals, bass, drums and others. It'd be nice to have, say, lead vocals, backing vocals, lead guitar, backing guitar, acoustic guitar, and so on and so forth. So, kind of removing the constraints so that it never becomes a problem.
There'll never be a problem where someone says, "Oh, crap, I've lost the stems," or "I don't know if we've still got the stems" or something like that. Ideally a place where any song can be separated perfectly and reduce headaches for people managing all these files and losing all these files.
GEORGE: To add to that, it's the paradigm of music. Some music has a very high barrier to entry, right? Playing the violin, for example, sounds rough until you're very skilled at it. I see a parallel with AI in the music industry—you know, personally, I play guitar and use several apps that allow me to play along. They can identify the notes I'm playing and provide guidance like "Try getting there" or "Here's the correct hand shape."
And also, I'm very into metalcore and that kind of stuff. Which is a lot of very complicated music. It’d be so helpful to have access to technology that helps people transcribe by ear, isolating all the music out. I think that it's not just a case of helping people make music; it's also helping people learn music, right? That sort of thing.
Can you share anything about the next thing you are shipping or working on?
BASIL: We're continuing to work on high-quality stem separation on a growing number of instruments. Our main priority at the moment is working with Queen Mary's to figure out how to achieve high quality for certain difficult instruments. I've mentioned things like, you know, many offerings provide piano, synth, or wind instruments, but the quality isn't very good.
There's something inherent about how these instruments relate to each other. So, in general, we're focusing on these challenging instruments to release something within the next year with a broader range of instruments. But yeah, without giving away too much, I think that's a good summary of where our tech's aimed for at least the next year.
GEORGE: On the business side, as I mentioned, we have people approaching us all the time who say, "I wouldn't have thought of that as a use case." Recently, there's been a lot of interest in the technology for licensing purposes - where licensors (or their client) need to be able to modify existing songs to fit their needs. They seek separation capabilities to enhance their products and streamline production turnaround times.
You can also imagine scenarios like, "That's my riff, you're stealing my riff," where companies want the ability to search or analyse songs that sound similar.
What advice would you give to aspiring entrepreneurs looking to enter the AI or music tech space?
GEORGE: So I think one thing, having seen this from the AudioStrip perspective, and having also worked in ventures for a little bit, for anybody that's a tech entrepreneur watching Silicon Valley is one of my pieces of advice. Because in that show, they cover every kind of thing that could possibly go wrong. Sometimes what could go wrong does go wrong, and it turns into a bit of a minefield.
If you have a great idea, there will be challenges along the way, and you're going to have to go shoulder to shoulder with people in the industry who might have an inferior product, but earn more money than you do. So, there's all this kind of stuff going on. My advice is, it's not for the faint-hearted; you will make mistakes, you will run into horrible problems that don't seem to have an immediate answer. And definitely watch Silicon Valley because it will give you a crash course in the personalities and the problems that you'll run into.
BASIL: Yeah, when George and I are talking, he usually references something like, 'Oh, this guy is like that character from Silicon Valley.'
GEORGE: It’s true!
BASIL: Yeah, I guess apart from that, one key to success, which has made AudioStrip quite successful so far, is finding an area you're interested in and focusing deeply on it—maybe one niche area—rather than spreading yourself thin across several areas. From everyone we've spoken to, they say we've developed a solution better than everyone else's. So I'm sure it's much easier this way, rather than trying to do 50 different things and being average, which may not attract clients.
So yeah, focus on something you enjoy, work hard at it, and strive to excel.
Is there anything else you'd like to share about AudioStrip or your personal journey that we haven’t covered?
BASIL: I'll just add, your question before we kind of dipped into why music source separation is hard. But if anyone's interested to dig deeply into why it's hard in terms of spectrally what's going on and how audio data is stored, we have a blog on that. So maybe people can click on that link if they're really interested.
https://www.audiostrip.com/blog/why-stem-separation-is-hard
GEORGE: Yeah, the thing about the journey is, Basil reached out to me and just asked for some help. As a result, we began working together. What I'm trying to say is, you'd be surprised how willing people are to help. It's worth reaching out to your network because you may have met someone along your career path who knows a lot more than you do about a specific problem. There's a third musketeer who helps us out with AudioStrip. He's been immensely helpful throughout our journey, and he's someone Basil and I have known for a long time.
Browse the Music AI Archive to find AI Tools for your Music
Find out more about AudioStrip
AudioStrip Links: Website | Instagram | LinkedIn | Twitter/X