Cleaning Up Vocal Rips With LALAL.AI Lynx

Grabbing vocal samples from the internet or vinyl? Here’s how to clean up voice rips with LALAL.AI and its new algorithm, Lynx, and use them in tracks.

There’s nothing like a vocal sample to spice up a track. Whether that be an acapella, a stem, or a bit of dialog from an old movie, a well-timed voice can add depth and interest to most any production. However, not every sample will be as clean as you might want. YouTube videos can be poor quality. Depending on the source, there could be background noise. Even a record might be too noisy to work well.

Thankfully, there are all kinds of methods these days to clean up vocal samples and remove unwanted noise. Depending on your chosen genre, you may actually want that noise. But if not, you’re going to need to clean it up before it hits your DAW.

LALAL.AI, possibly best known for it’s excellent stem seperator, has a new algorithm available for cleaning up vocals called Lynx. Today, we’re going to put it through its paces and see how it fares on a variety of vocal samples, and then use them in tracks.

What Is LALAL.AI Lynx?

LALAL.AI has a suite of AI-powered vocal removal, cloning and cleanup services. Now the company has announced Lynx, a new voice isolation neural network for removing noise and other artifacts from voices. According to their blog post, they spent a year training it to recognize sonic dirt and separate it from recordings of the human voice.

Using Lynx is fairly simple. You can work with it on the website, in a desktop app, an iOS app for iPhone or iPad, or as a plugin in your DAW (Pro tier only). We’re using it on our desktop. 

Working With Lynx

Start by opening the software. Next, drop the audio file that you want to clean onto the app. You’ll be presented with a number of different processing options. Voice & Noise is for cleanup and contains the new Lynx algorithm. Vocal & Instrumental, on the other hand, is for separating vocals from music and creating acapella stems. We’ll look at this one later.

Once you’ve selected Voice & Noise, you’ll come to a new screen with a few processing options. v7 Lynx should already be selected under Neural network, but if it’s not, make that change. Next, select the noise canceling level you prefer, either Mild, Normal, or Aggressive. We found Normal strikes a good balance between cleanup and retention of quality. Lastly, you can also decide whether to de-echo vocals or not. This is useful if the audio was recorded in a reverberant room, for example. 

Once you’re done, make sure “Create preview” is selected so you don’t get charged if you make a mistake and then hit Start Processing. The app will separate the vocals from the noise and render them as two individual files. If you like what it’s done, hit the Full button to save your renders.

Let’s try processing three different types of vocals.

Movie Dialog Cleanup

Bits of movie dialog can help contribute a sense of drama in a song. One excellent resource is films in the public domain, older movies that no longer have any copyright protection. YouTube is full of these, many of them films from the first half of the 20th century. As you might expect, though, these old movies can sound scratchy, as with this bit of dialog we sourced from a horror film. 

The horror:

After running it through LALAL.AI, with noise canceling level set to Normal, the sound is much more usable. 

Here’s the processed file:

After a little trimming and time-stretching, and with some additional EQ and compression, here’s how it sounds as part of a speed garage song with loops from Splice.

Full track:

Acapella Vinyl Clean Up

Producers have been sampling acapellas into their tracks for decades. That’s nothing new. What is new is the kind of technology we now have to clean up the audio. Back before stem separation made this so much easier, we used to have to buy dodgy acapella vinyl compilations at DJ specialty stores. Being bootlegs, you’d have to deal with low-quality transfers — often recorded off one record and then pressed to another — as well as scratchy vinyl surfaces.

To see how LALAL.AI might do in this situation, we created our own version of a dodgy acapella. First, we ripped the vocal-only version of Fingers Inc and the legendary Chuck Roberts’s “My House” off YouTube at 240p. Next, we dropped it into Ableton Live and used iZotope’s freeware Vinyl plugin to add scratches, pops and electrical noise.

Here’s what that sounds like:

To clean it up, we added that dirty file to the cue in LALAL.AI. Given the degree of noise, we first tried the Aggressive noise canceling level. While it did a good job removing noise, it was a little too heavy-handed on the original audio, so we dialed it back to Normal. We also chose to de-echo the vocals, seeing how the original track used a slapback echo on Roberts’ voice. It doesn’t do a perfect job but it’s much better than it was before.

Here’s how that turned out:

As LALAL.AI doesn’t just remove but separates noise from the vocal, it gives you both files so you can hear what’s been subtracted.

The removed noise:

Lastly, we incorporated the clean file into a UKG song with loops from Splice, adding saturation, EQ and compression plus reverb and delay to help it sit in the mix.

In the beginning:

Remixing With Vocal Stems

For this last use-case, we wondered how LALAL.AI would do cleaning up a vocal that it had pulled from a piece of music itself. To put it through its paces and replicate a remix workflow, we chose a song recorded live, complete with obnoxious overdubbed cheers. And what could fit the bill more than 1970s rock band Kiss? They have a new live album coming out called Kiss Destroys Anaheim 76. The song is “Do You Love Me?”

We started by putting the beginning of the track into the app and selecting Vocal & Instrumental. As mentioned before, this algorithm doesn’t offer a cleanup mode, just stem separation. But, given that the source file contained clapping, cheering, and other sounds, the result wasn’t as clean as we would have liked.

Here’s the original snippet:

And how the app rendered the vocal stem:

It’s pretty good, but we reckon it could get even better. We went with a normal level and applied de-echo to get it as clean as possible for our track.

The final result:

Lastly, here’s that isolated Kiss vocal in an acid house track with loops from Splice, bolstered with saturation, EQ, compression and other effects.

Limosine acid:

LALAL.AI Lynx is available at a variety of price points. There’s a free starter tier, Lite for $7.50 a month, and Pro for $15 a month. There are also one-time top-ups.

Follow Attack Magazine