LALAL.AI v4 Update: Testing Audio Separation Quality and Accuracy

A refreshed AI model for better source separation

The recently updated LALAL.AI vocal-removal and source-separation service

uses the next-generation v4 Orion AI model to improve separation quality over the earlier version.
In this October 2023 review, we compare several recordings with the previous model to explore the changes.

LALAL.AI Video

What Is LALAL.AI?

If this update is your introduction to LALAL.AI, here is a quick explanation.
LALAL.AI uses AI to separate audio. Starting with a finished stereo mix, it primarily addresses these two tasks:

• Separate individual sources, including vocals.
• Remove noise captured during microphone recording.

The basic workflow is simple.

Drag and drop an audio file, such as MP3 or WAV, and select the source type you want to separate. The service processes it automatically, with downloading available according to your plan.

In addition to the browser-based web app, desktop applications for Mac and Windows and mobile applications for iOS and Android were available at the time of the review, allowing use across different environments. Check the requirements for your device and version.

Comparing Audio Quality with the Previous Version

Now we will use actual recordings to compare the separation accuracy of v4 Orion with the earlier v3 Phoenix model.

The available separation types at the time included vocals, drums, bass, noise, acoustic guitar, electric guitar and piano.

There were also beta options for synthesizer, strings and brass.

However, v4 Orion initially supported only vocals and noise separation, so those are the functions tested here. These are the October 2023 model limitations, rather than a description of today's service.

Before starting, keep this setup detail in mind: when loading a file, check whether v3 or v4 is selected as the AI model.

First, listen to the original recording before comparing the separated outputs.

In the reviewer's comparison, v3 left percussion in the extracted vocal, whereas v4 isolated the voice more cleanly in this example.
Let us listen to another passage.

The v3 output includes sound effects, while the v4 result separates them more effectively and sounds clearer to the reviewer.
Next, listen to the instrumental backing.

Some vocal remnants were audible in the v3 backing track; in this example, v4 removed them more cleanly.

You could preview the separation result before purchasing processing credits.

The preview was limited to one minute. Selecting Process the Entire File used your available credits to process the rest of the recording. This describes the original review's workflow and terms.

To save the results, click Download All or a download button for an individual file.

Next, let us try a recording with a different musical character.

Piano audible in the v3 vocal output is more effectively separated in the v4 example.
The reviewer also found v3 grainier, possibly because of bleed from other sources, while v4 sounded clearer.
Now listen to the instrumental backing.

In the v3 example, the lower register is harder to hear, perhaps because some strings were removed along with the vocals. The v4 result again suggests improved separation accuracy in this test.

Finally, we will test a recording with heavily reverberant vocals.

Hi-hat bleed heard at some points in the v3 vocal output is more effectively separated in the v4 example.
As before, let us also listen to the backing track.

The v3 backing retains some vocal content, while the reviewer found only a slight trace in the v4 result.

After testing these three contrasting recordings, the update appeared to improve separation across the examples tested. This is an observation about these samples, rather than a guarantee for every genre or recording.
For preparing remix material or karaoke backing tracks, the results suggest the possibility of higher-quality source separation in these examples.

Testing Noise Removal

Next, we will examine noise separation.
We recorded narration in a noisy environment, then processed it to separate the noise.
Compare the examples to hear how much noise was removed.

In the voice-only results, the reviewer could already hear a difference in separation accuracy.
Looking at the noise-only waveforms, v3 showed larger changes where speech occurred over otherwise steady noise. The smaller changes in v4 suggested better separation in this test.

This review has focused on the vocal and noise-removal functions supported by v4 at the time.
The other source types still used earlier AI models then. If comparable improvements were extended to them, the range of useful applications could grow further.

We looked forward to future updates.

Plans and Payment at the Time of the Review

Let us also look at the payment options described in the original October 2023 article. These are historical terms; check the official website for current pricing and availability.

The plans then included Lite, Plus and Pro, with High Volume options for heavier use.
At registration, the service offered ten minutes of credits, allowing users to try the separation quality without an initial purchase under those terms.

The article described the plans as differing in the amount of audio you could process, with no difference in sound quality or available functions
. It also described one-time purchases of processing credits rather than a subscription. Do not assume that these historical terms apply to current plans.

The appropriate plan depended on how much audio you intended to process.
For example, Lite then included 90 minutes of audio processing per purchase: enough for 18 five-minute processing operations.

Larger plans offered a lower cost per minute, so a higher-tier plan could be more economical for heavy use under that pricing structure.

The original article also mentioned a sale on Plus and Pro. That was a promotion at the time of publication, not a current offer.

That concludes our look at this LALAL.AI update.
The reviewer heard improvements in separation quality and looked forward to further development.
At the time, future developments mentioned included applying the v4 model to other source types and VST plug-in support. These statements describe expectations in the original review, not confirmation of current product features.

Frequently Asked Questions

Q1. What improved between LALAL.AI v3 and v4?

In the examples tested, v4 Orion reduced unwanted drum, percussion and piano bleed in extracted vocals. The reviewer also observed improved separation of environmental noise from a spoken recording.

Q2. Does LALAL.AI work with every musical genre?

The review tested three contrasting recordings, including one with strong vocal reverb, and found improved separation in those examples. Three tests cannot establish results for every genre or recording, so preview your own material before processing it fully.

Q3. Were the v4 improvements available for sources other than vocals?

In October 2023, v4 Orion supported vocal and noise separation. Other sources, such as drums, bass and guitar, still used earlier models. The article discussed possible expansion in future updates; check the current service for today's model availability.

Q4. What should I check before using LALAL.AI?

Check which AI model is selected when loading the file. The reviewed workflow provided a one-minute preview, letting you assess the quality before using credits for full processing. Current limits and terms may differ.

Q5. Can I use LALAL.AI outside a browser?

At the time of the review, Mac and Windows desktop apps and iOS and Android mobile apps were available in addition to the web app. Check current platform and version requirements for your device.