LALAL.AI Introduces Andromeda: The Next Generation of Audio Source Separation

After years of refining its audio source separation technology through multiple generations of neural networks, LALAL.AI introduces Andromeda — its most advanced and meticulously engineered model to date. Trained on expanded datasets, driven by deeper audio pattern analysis, and fine‑tuned for maximum separation accuracy, Andromeda sets a new benchmark for quality in vocal isolation and stem separation.

This model reflects the culmination of six generations of research in digital signal processing and machine learning. Each iteration built upon the insights of the previous one, distilling everything learned from millions of processed tracks and thousands of user feedback cycles worldwide. The result is a model that delivers both greater speed and precision, achieving up to 40% faster processing while significantly enhancing the clarity of extracted stems.

Clearer, Cleaner, Smarter: How Andromeda Handles Sound

Andromeda, like the previous LALAL.AI neural network Perseus, operates using transformer technology, which helps it analyze and separate complex audio more accurately. The new network processes tracks faster than the predecessor and maintains strong separation quality, handling vocals, instruments, and background sounds with improved precision.

Andromeda works by carefully separating audio layers to reduce common issues in AI separation, such as vocal bleed into other instruments, unwanted reverb, and loss of clarity in high frequencies. Instead of just looking at the sound as one simple waveform, Andromeda studies it in different ways — by time, frequency, and tone.

In practice, this means it doesn’t just detect “the voice” or “the drums”; instead, it registers timbre composition, resonance depth, and instrument presence across the stereo field. This kind of separation used to require manual cleanup in a digital audio workstation. With Andromeda, many of those post‑processing steps are now unnecessary. Even quiet elements, like soft backing vocals, come out sharper and easier to mix.

The Andromeda network is now integrated into several LALAL.AI web services: 

• Stem Splitter (for Vocal and Instrumental + Voice and Noise stems)
• Lead & Back Vocal Splitter
• Echo & Reverb Remover
• Voice Cleaner

Each of these tools benefits from Andromeda’s enhanced processing capabilities, allowing users to extract cleaner stems, reduce noise, and isolate vocals more effectively than before.

Foundations Built on Experience and Innovation

For over a year, the Andromeda network underwent intense training. The model was pre-trained for more than nine months alone, backed by a multi-stage data cleaning process. This process used unique filtering techniques developed by LALAL.AI’s team, applying the latest advances in digital signal processing and neural network research. These advanced methods helped build a stable, efficient model capable of high-quality stem extraction.

Although Andromeda’s architecture shares similarities with Perseus (transformer-based design), it includes multiple improvements in training and structure that yield faster and more reliable results. Compared to Perseus, Andromeda used about four times the amount of training data, which contributed significantly to its ability to handle a wider variety of audio sources.

Deep and Clear Extraction Unified

For the first time in LALAL.AI’s history, users no longer face a trade-off between detail extraction and cross-bleed prevention. Andromeda’s architecture breaks new ground: it replaces the two separate Enhanced Processing modes, Clear Cut and Deep Extraction, which were available in previous networks like Orion and Perseus, with a single, unified network.

There’s no need to choose between minimizing leakage or maximizing detail; the model intelligently balances both, ensuring exceptional stem quality with no manual toggling or compromise. Every vocal, instrument, and background element is isolated with meticulous care, and cross-bleeding, once an inevitable consequence of pushing for higher detail, is now all but eliminated. It means that Andromeda outperforms all previous options, bringing not only improved separation quality but also a simpler user experience.

Performance Improvements That Matter

In testing, Andromeda showed up to a 10% improvement in SDR (Signal-to-Distortion Ratio) over Perseus on benchmark datasets. SDR is a measure used to evaluate how well an audio source separation method reproduces the original signal without adding distortion or artifacts. While the numerical increase may seem modest, the true strength of Andromeda lies in its enhanced ability to separate a broader range of audio tracks more accurately — a quality that’s difficult to capture fully in metrics.

The frequency range remains consistent at up to 22 kHz with stereo processing, providing high-fidelity separation that LALAL.AI users expect. Andromeda’s advanced training on significantly more diverse datasets means it handles challenging audio scenarios and complex instrument mixes with greater precision, reducing artifacts and increasing consistency.

The loudness of the input files has less impact on the quality of separation than before. Softer or louder tracks won’t skew the results as much anymore, delivering a more consistent output regardless of volume differences. The progress behind Andromeda is largely due to the extensive use of DSP (Digital Signal Processing) techniques. DSP forms the core of LALAL.AI’s technology, providing essential tools to perform accurate and stable stem separation. Thanks to DSP, Andromeda achieves the high level of quality that LALAL.AI users expect.

How to Extract Voice or Vocals with Andromeda

If you want to see the new network in action, you can do so in the LALAL.AI Stem Splitter (Vocal and Instrumental and Voice and Noise stems), Voice Cleaner, Lead & Back Vocal Splitter, and Echo & Reverb Remover. Andromeda is set as the default in these services, but if you want to make sure the new network is on, compare it to previous generations, and/or tweak other settings, here is how, on the example of Stem Splitter, our most popular tool.

1. Visit the LALAL.AI Stem Splitter page. 

2. Toggle the stem selection menu. Select either Vocal and Instrumental or Voice and Noise from the dropdown list. 

3. Click the Select Files button and upload your audio or video track (containing the sound of voice or vocals).

The service supports MP3, OGG, WAV, FLAC, AIFF, AAC, M4A, AVI, MP4, MKV, MOV, and M4V formats.

4. After the track is processed, you will see the stem previews. Click the play button within the audio preview section to evaluate the vocal track. Tweak the settings again if needed.

5. Process the track in full. If you’re happy with what you hear in the preview, click the Split in Full button.

6. Download the results. Save the stems separately or grab them all at once by clicking the Download All button.

Now you’ve got vocals separated from the rest of the sound with exceptional precision of the new Andromeda neural network.

What’s Next for Andromeda and LALAL.AI

Andromeda is just the beginning of a new era for LALAL.AI. In the coming months, the network’s reach will expand far beyond web-based tools. Our team is working on integrating Andromeda into the LALAL.AI mobile apps, bringing high-precision stem separation to your phone or tablet for quick voice and instrumental splits on the go.

We’re also preparing the launch of the first LALAL.AI VST plugin, built around the Andromeda model. This plugin enables producers and audio engineers to perform instant in-DAW vocal and instrumental isolation with the same accuracy available online, but at a significantly faster speed and without leaving their workflow.

Stay tuned for updates on our social channels: Facebook, Instagram, TikTok, Reddit, X, and YouTube. Experience the future of AI music processing today with Andromeda at LALAL.AI!