LALAL.AI, an AI-powered audio processing platform used by millions of audio engineers, video producers, journalists, podcasters, and localization teams worldwide, today announced the release of Lynx, its first neural network designed from the ground up exclusively for speech denoising.
Lynx is trained to separate speech from everything else: background music, crowd noise, mechanical interference, environmental sounds, and the full range of acoustic artifacts present in content produced outside controlled studio conditions. The result is a clean voice track ready for transcription, localization, text-to-speech input, or human voice replacement, without the manual cleanup steps that slow batch production.
"In speech denoising, you're not separating things that were recorded together by design. You're trying to recover a voice from an environment that was never meant to be a recording studio,” says Nik Pogorsky, LALAL.AI Product Owner and Co-founder. “We spent a year building a model that treats that as the actual problem, not a side case of something else.”
Lynx runs on a proprietary architecture developed over one year by the LALAL.AI research team. The model is six times smaller than Andromeda, LALAL.AI's flagship cloud stem separation model, which reduces computational load without compromising output quality. Lynx was trained on a manually curated dataset: since no reliable automated pipeline exists for cleaning the full dynamic range of real-world speech from diverse noise sources, the team spent months hand-selecting, filtering, trimming down, and preparing audio tracks covering conditions from quiet interviews to noisy field recordings.
"We consider the resulting training set one of the cleanest we’ve used. It totals several hundred hours of carefully filtered audio, supplemented by several thousand additional hours of more loosely cleaned content used during pretraining," blog post announcement says. "hat data spans far more than clean studio speech. It includes whispering, labored breathing, sniffling, children crying and laughing, crowd noise, exotic bird calls, and everyday ambient sound."
Lynx is available now through LALAL.AI's Voice Cleaner and Voice & Noise stem inside Vocal Remover and Stem Splitter, accessible via browser, mobile app, desktop app (cloud mode), and the LALAL.AI API for businesses, SaaS integrations, and enterprise deployments.
It supports the same file formats already used across LALAL.AI platform: MP3, OGG, WAV, FLAC, AIFF, AAC, M4A, AVI, MP4, MKV, MOV, and M4V.
The demand for dedicated voice cleaning is reflected in LALAL.AI's own usage data: the Voice & Noise stem, now powered by Lynx, is the platform's second most-used separation track, with more than 11.5 million audio splits processed on it in 2025.
Planned improvements to the Lynx architecture include enhanced separation of choral and group vocals, and improved isolation of speech recorded at distance from the microphone.
