Musiio provides AI analytics, tagging, and search tools for some of the world's largest music catalogs, including Sony Music, Hipgnosis, Amanotes, Epidemic Sound, and Blanco Y Negro.
Hazel Savage, a rock-loving guitarist who went on to become co-founder and CEO, spent 15 years in the music industry working for some of the world's largest music brands – from stacking shelves at HMV to running teams in music listening and cutting-edge enterprises. It is recommended that Hazel understand the needs of the industry, from musicians to large multinational corporations.
You've been in the music industry for over 15 years, what makes you love music so much and why do you want to get involved in the music industry?
My parents were very rocky. They're huge music fans, so I was always surrounded by music when I was growing up. Then, on my 13th birthday, I got a guitar. I still play and am passionate about live performances. So when I figure out what I'm going to do, it makes sense to focus on what I'm committed to almost all of my time.
I ended up doing a lot of related things. I play in a band. I manage the band. I run club nights. I was handing out flyers for other people's club nights, making a list of guests, and before I knew it, it turned into a profession, albeit with certainly techian tendencies.
Can you share the origin story behind Musiio?
My first job after graduating from college was stacking shelves at HMV (British Record Stores). So, you could say I've been aware of the problem of music classification since then. Fast forward a few years (via Shazam, Pandora, and Universal) and I work on a UGC music platform, uploading thousands of tracks a day. I work with a playlist who has to manually upload the best music to the playlist. He listened to hundreds of tracks a day. Sometimes, he has enough content to fit the playlist. Some days he didn't. I'm starting to wonder if there's a way to automatically find the best tracks for a given scene. This way, he can use his skills as a music expert to curating, rather than just acting as a filter for bad music.
In 2018, I met my co-founder Aron Pettersson, Musiio, in Singapore through entrepreneurial incubator Entrepreneur First. Aron is an artificial intelligence genius. When we discussed ways we could collaborate, we realized that we could use Aron's AI skills to solve the problem of music-based filtering, auto-tagging, or searching for music with genres, moods, BPM, etc., or fingerprint-based searches. Aron spent an afternoon building a prototype of the algorithm, which we set up as a free music archive. We went out to lunch and let it process the data. When we returned, we were amazed at the accuracy of the results. We can't possibly hope for a more successful proof of concept. From there, we've done massive optimizations to the algorithm. We have a music team to help teach artificial intelligence and do quality checks, we have released products for labeling,
What different types of machine learning algorithms are used?
We've built our own proprietary algorithms, which we think is our secret! My co-founder, Aron, has worked in molecular biology, neuroscience, physics, and even game development for over a decade and has been at the forefront of machine learning. He leads our AI team. We also leverage available technologies such as TensorFlow, Kubernetes, and Google Cloud Services to achieve scalability and deliver our products at scale, marking 5,000,000 tracks per day at our maximum number! We also spent a lot of time and effort to streamline our workflow in JIRA; It's not just about which tools you use, but also about the efficiency with which you work with teams of developers and music experts. The combination of the two teams of AI and Music is the second part of our secret.
What are the challenges behind building a music search engine?
Speed and accuracy are the biggest challenges to searching. It has to be fast because people are using it in real time. This is different from tagging because users typically make multiple search queries, but tagging occurs only once.
There are several things you can do to speed up your search. You can display only tracks that share the same label as the seed track, but at the expense of accuracy. For example, an audio-only reference search in a catalog of 200 million tracks can take a long time, so you need to constantly balance speed and accuracy and look for solutions. It's tricky, and some of it is hard-won knowledge, but what I can share is that we convert audio files into spectrograms, highly detailed fingerprints of audio files, and when we do audio reference searches, the algorithm analyzes up to 1,500 data points – far beyond what can be done with text tags alone. It also has musical characteristics that are difficult to describe, such as vocal quality, ambience, and ambience. We also allow users to define filters so their searches can be faster and more focused.
Another challenge is how to manage dependencies. Most people don't go past the first page of the results, so we spent a lot of time on that.
What issues does Musiio solve for b2b customers?
We offer a music catalog for anyone. We've built scalable technology, whether you're a musician who doesn't have time to tag music and wants to focus on creating, or a streaming service with hundreds of millions of tracks.
We help record labels organize their data for better catalog navigation, we help sync companies (putting music into video/TV and movies) to find hidden gems, and we help streaming services build better playlists. The problem with all these companies is that processing audio manually by listening to each track is labor-intensive and difficult to do accurately over a sustained period of time. I recorded the goal of 1000 tracks as an experiment. It took two weeks and it wasn't fun at all. Our AI can mark millions of tracks every day with 90-99% accuracy.
With our Musiio search product, we allow B2B customers to provide audio reference search as a feature. If a video producer is looking for a music placement, they'll first understand what customers expect from genre, sentiment, BPM, and then search on the site of their choice.
Musiio shortens this process with the partner who installed our search by allowing the same video producer to search the entire database in a matter of seconds using the "reference track". Our AI will scan the reference track and return the nearest audio match.
Musiio recently launched an NFT Song Slicer product, can you describe what it is?
NFT Song Slicer is a prototype designed to help artists get more value from their music. It uses an AI-driven process to find the hooks needed in the track (up to three per song) and provides timecode so that artists can cast those song parts as NFTs. It can also do this automatically for the entire catalog, making it easier for labels and artists with a large number of post-catalogs to quickly create new digital collection assets.
What are the potential use cases for this type of Song Slicer product?
For catalog owners or artists with a large number of post-catalogs, NFT Song Slicer can choose the most valuable part from millions of songs per day. For example, a record label can convert these song clips to NFTs and sell them as limited-edition digital goods.
With the streaming revolution, it's hard for fans to get a dollar out of the pockets of artists they like. We see NFT Song Slicer as a way for fans to support their favorite artists and for fans to own digital collectibles. Rights holders can also price each slice differently. For example, a chorus may cost more than a verse.
And, since NFT Song Slicer can identify the most valuable parts of a track, we see that this technology can provide value predictions for NFTs and even entire music catalogs.
What is your vision for the future of Musiio?
I said Musiio is a third of a billion-dollar company. To build that company, you need three parts. The first is legitimate access to large amounts of data, or "pipelines." The second part is technology. That's who we are, and we're very good at what we do. The third and final part is labeling: a way to monetize what you discover, search, or discover. Musiio has been working towards this long-term goal.
Do you think AI will be able to create and compose music in the near future?
I'm pretty blunt about saying I'm not an avid fan of ARTIFICIAL intelligence. It's an interesting academic experiment, and there are systems that can do it, but I just don't think it's necessary. What makes Musiio so great is that no one wants to tag thousands of songs every day. It's not fun, and you don't need someone to do it efficiently or quickly. But music production? I'm not sure. There are not a few people who want to make music.
Even so, I think it will take at least 5 to 10 years for the AI-generated music to sound good. I heard some AI-generated piano pieces the other day, and it's hard to say whether it was written by AI or someone who wasn't very accomplished. I don't believe that ai-powered performance can be indistinguishable from accomplished human players.
Why are you doing this? Much of the music is interesting around the legends of the artists, their characters, their styles, and their messages. It's not just about music.
Is there anything else you'd like to share about Musiio?
I'm thrilled that Musiio has just finished fourth in Fast Company's 10 Most Innovative Music Companies of 2022. Our team and technology have grown from the seed of an idea to gain international recognition alongside huge industry names like Hipgnosis and SoundCloud. It's a tribute to the blood, sweat and tears our team put into our industry-leading products. We're excited to be at the forefront of the intersection of music and technology. Knowing that there are use cases that we didn't even think of makes me very excited about the future.