LALAL.AI unveils Andromeda AI model for cleaner audio

LALAL.AI has launched Andromeda, a new audio source separation model built on a transformer-based neural network. Designed specifically for stem separation and audio clean-up, the model aims to provide superior results for creators, particularly within the video podcasting space. Alongside the release, the company has provided practical guidance to help users enhance the overall sound quality of their productions.

The Andromeda engine has now been integrated as the default processing technology across several of the company's core tools. These include the Stem Splitter, Lead and Back Vocal Splitter, Echo and Reverb Remover, and Voice Cleaner. By centralising this technology, LALAL.AI intends to streamline the editing process, allowing for more precise isolation of audio elements and more effective removal of unwanted background noise or acoustic imperfections.

Andromeda follows six prior generations of the firm's work in digital signal processing and machine learning, according to the company. It said it trained the new model on expanded datasets and designed it for higher precision than earlier versions.

"After years of refining our audio source separation technology through multiple neural network generations, Andromeda reflects our deepest engineering efforts to date," said Oksana, Lead Contributor, LALAL.AI.

LALAL.AI described the model as using a transformer architecture. The company said it analyses audio across time, frequency, and tone.

"The model's transformer-based architecture allows it to analyze audio in terms of time, frequency, and tone - not just simple waveforms - resulting in separation clarity and consistency that creators across disciplines can rely on," said Oksana.

Performance claims

LALAL.AI said Andromeda improves separation speed and output quality. The company reported up to 40% faster processing and up to a 10% improvement in signal-to-distortion ratio over its predecessor.

The firm linked the changes to common issues in AI-based separation. It cited vocal bleed, unwanted reverb, and loss of high-frequency detail.

LALAL.AI also said Andromeda removes a trade-off that existed in earlier models. The company said users previously had to choose between extraction detail and bleed control. It said the new model combines those aims in a single engine. The company said this reduces manual clean-up work in digital audio workstations.

The move positions Andromeda as a shared foundation across the company's web-based tools. Stem Splitter targets vocal and instrumental separation. Lead & Back Vocal Splitter focuses on vocal layers. Echo & Reverb Remover concentrates on ambience artefacts. Voice Cleaner focuses on cleaning spoken audio.

Podcast focus

Alongside the model release, LALAL.AI has published a guide focused on audio improvement in video podcasts. The guide covers microphone choice, room preparation, gain staging, noise removal, and loudness consistency for distribution platforms including YouTube, Spotify and Apple Podcasts.

One section addresses the role of consistent recording conditions and standardised processing steps in maintaining clarity. The guidance links uneven levels, echo, and background noise with a perceived drop in production quality.

"Audio quality often decides how professional a video podcast feels," said Oksana.

"Even thoughtful conversations lose their impact if the sound is uneven, echo-filled, or obscured by background noise," said Oksana.

LALAL.AI said the guide does not assume specialist engineering knowledge or high-end equipment. It said consistent habits and structured processing can improve results in common podcasting setups.

Creator workflows

The company positioned Andromeda and its associated tools across several creator segments. It said musicians and producers can use stem extraction for remixing and practice. It said video editors can improve dialogue clarity and reduce background noise as part of post-production. It said podcasters can use the workflow guidance and apply clean-up steps before editing and publishing.

The company also pointed to broader content creation use cases. It said creators working across long-form video and short clips can use separation and clean-up tools as part of their editing process.

The field of stem separation has become increasingly competitive as creators seek more efficient methods for isolating vocals, instruments, and speech from mixed audio files. Consequently, transformer-based approaches have gained significant traction within audio processing circles. This shift is largely due to their ability to represent longer contexts within signals, offering a distinct advantage over certain earlier neural architectures.

Distribution plans

LALAL.AI said it plans to extend Andromeda beyond its web interface. The company said it is preparing mobile applications that run on the Andromeda engine. It also said it is preparing a VST plugin for use inside digital audio workstation environments.

The company did not provide a launch timetable for the mobile apps or the plugin. It framed both efforts as part of a broader roadmap that brings the same separation and clean-up engine into additional production environments.

"We see Andromeda as just the beginning," said Oksana. "Our focus is on empowering creatives with both the highest quality audio separation and practical guidance to make better audio - whether they're remixing tracks, preparing podcasts, or editing video content," said Oksana.