SoundSafe Turns Vocal Stems Into a Lyric-Based Copyright Check with LALAL.AI API: Interview

This time, as part of our interview series dedicated to the Musicathon, we’re chatting with Bryce, the mind behind SoundSafe. This project has taken our special award during the music tech hackathon in June, so we simply had to share the story behind it. 

SoundSafe is a combined tool for musicians and lawyers that uses LALAL.AI, Musixmatch, ElevenLabs, as well as other solutions to separate vocals, transcribe lyrics, and compare them against existing songs to estimate copyright risk before release.

Projects at the intersection of music, law, and copyright are rare enough on their own, and even rarer are the ones that focus on lyrics rather than melody. So we sat down with Bryce to learn more about the tool and whether there are plans to grow it into a permanent, essential player in the industry.

The idea behind SoundSafe is an industry gap  

Bryce comes from the music data and tech side. As a contributor to the Musixmatch catalog, he has seen, edited, and added many lyrics in the last few years. In Bryce’s words, the idea behind SoundSafe stems from the gap in the tools that exist: seeing first-hand how independent artists get their music pulled from streaming or hit with copyright strikes is what inspired Bryce to make this tool. 

The industry has great software for finding master recording infringements: if you copy-paste an audio file, existing software will flag it. But if someone re-sings that famous hook in their own voice, the audio waveform is new, so those standard audio scanners miss it. 

Traditional fingerprinting isn't designed to detect interpolations because the recording itself is different. Because of that technical gap, catching lyric infringement sometimes still relies on manual review and hoping an A&R recognizes the words. 

Sometimes, artists use vocals that they think are royalty-free or accidentally recreate a famous hook from their memory. They don't have label lawyers on retainer to check this stuff before release. The “aha” moment was when we realized that we could just extract the vocals and turn them into text to automatically catch similarity. Melody can be subjective at times, and proving that two melodies are the same often ends up in a messy court battle. Lyrics, on the other hand, are absolute, indexable text data. 

“A phrase like ‘I love you, baby’ will trigger a text match on thousands of songs”

Filtering out standalone, generic lines was the hardest part of turning a similarity score into an actual “should I be worried” verdict. A phrase like "I love you, baby" will trigger a text match on thousands  of songs, so it was complicated to tune the logic to look for unique consecutive word chains and specific syllable structures, rather than just relying on a high "bag-of-words" similarity score.

“Musixmatch, Whisper, GPT-4o-mini, ElevenLabs, and LALAL.AI are the best tools for what we’re trying to do.”

Musixmatch was the obvious choice for lyric data; we have massive databases via Musixmatch, which makes it a much more programmable problem to solve. Meanwhile LALAL.AI is perfect for clean stem extraction; I would choose both of them in a production environment regardless!

A huge part of SoundSafe’s reliability traces back to stem separation quality

SoundSafe relies on LALAL.AI under the hood, so we were eager to learn what role our stem splitter played in the entire pipeline.

Heavily distorted vocals, intense ad-libs, or vocals buried in reverb or other effects are the biggest technical limitations. If the audio processing is too heavy, Whisper gets the transcription part wrong. If the text is wrong, then the Musixmatch query returns useless information, and the entire process is borked. 

Without the LALAL.AI-powered vocal isolation, it would be so much worse: Whisper gets confused by the complex production of these songs and starts hallucinating words for the noises it’s hearing. The word error rate spikes, so LALAL.AI is the filter that actually makes Whisper a viable step in the process. If there’s a bad transcription, then there are bad results. If the stem separation leaves too much instrumental bleed, the transcript is wrong. 

“LALAL.AI's extraction quality basically directly dictates SoundSafe's accuracy. Luckily, LALAL is very good :)”

Integrating the LALAL.AI API into the project was one of the easiest integrations we had. The docs were very clear, so it took us very little time to set up. The biggest positive surprise was the speed of the processing and how cleanly the vocal model handled backing harmonies without leaving weird audio artifacts. 

The road ahead of SoundSafe: A hackathon project or a proper industry standard

There’s definitely a path to production. The ideal buyers are sync agencies and indie labels as they receive tons of tracks from independent producers and need a fast, automated way to surface tracks that deserve closer review before they're pitched without paying  someone hundreds of dollars an hour to review it.

In terms of the new features roadmap, I’d definitely add batch processing so labels can check whole albums at once, but my main focus would be building a plugin for music software like Ableton or FL Studio. 

Right now, people check for copyright issues after a song is already fully finished and exported. I want them to be able to run a check while they are making the track. 

“Catching a problem while the singer is still in the recording booth saves everyone a massive amount of time and money compared to fixing it weeks later.”

Besides, there’s definitely an interest in extending clearance checking beyond lyrics into melody and instrumental fingerprinting. Once the lyric part is solid, the next step is taking the other half of LALAL.AI's output (the instrumental stem) and running it through an audio-fingerprinting model to check for uncleared drum breaks or synth riffs.

Follow LALAL.AI on Instagram, Facebook, Twitter, TikTok, Reddit, LinkedIn, and YouTube to keep up with all our updates and special offers.