Dmitry Orlov's LALAL.AI launched Lynx, a new voice isolation model built for speech denoising in messy recordings, and the company is putting it directly into the products creators and developers already use.
LALAL.AI said in a thread on X that Lynx is 6x smaller than Andromeda, its prior flagship model, and is now the default model in Voice Cleaner. The company also says Lynx powers the Voice + Noise stem inside Vocal Remover and Stem Splitter, with availability across web, desktop, mobile, and API. No local GPU is required, according to the launch thread.
LALAL.AI on X The launch took place on July 13, according to a Podnews press release, making it a current product move rather than a resurfaced model announcement. LALAL.AI is also tying the release to a 30% discount on Annual Pro through July 28; the Pro tier includes 250 Fast Queue minutes per month, API access, and the VST plugin, according to the company's pricing page.
The sharper read is that LALAL.AI is moving voice cleanup from a feature around music splitting into a dedicated infrastructure layer for speech-heavy production. The company started as a vocal-removal tool, but its current product set now spans stem splitting, Voice Cleaner, Echo and Reverb Remover, Voice Changer, Voice Cloner, Lead/Back Splitter, desktop apps, mobile apps, a VST plugin, and an API for software integration. Lynx gives that suite a smaller model aimed at non-studio speech, where recordings come with crowds, engines, birds, footsteps, background music, mechanical sounds, and room artifacts.
A founder-led audio company gets more specific
LALAL.AI is operated by OmniSale GmbH, the Swiss company behind the product. OmniSale lists Orlov as CEO and founder, and describes LALAL.AI as its AI-powered audio stem separation service. The company's public materials place the team in the signal-processing lane: artificial intelligence, machine learning, mathematical optimization, and digital signal processing.
That background shows up in the product history. LALAL.AI says it developed Rocknet in 2020 to extract instrumentals and voice tracks, then released Cassiopeia in 2021, Phoenix and Voice Cleaner in 2022, mobile apps and Orion in 2023, and Andromeda in 2025. Andromeda used transformer architecture and emphasized cleaner vocals and faster processing compared with the previous system. Lynx is being presented as the next step in a narrower direction: speech denoising rather than general-purpose stem splitting.
Podnews identifies Nikolay "Nik" Pogorsky as LALAL.AI's product owner and co-founder. The announcement describes Lynx as the company's first neural network built specifically for speech denoising, aimed at real-world noise such as crowds, engines, footsteps, birds, background music, mechanical noise, environmental sounds, and acoustic artifacts. LALAL.AI has not published an independent benchmark comparing Lynx to competing voice isolation tools in the materials cited.
The market is shifting toward embedded audio preprocessing
Voice isolation used to read like a post-production tool for podcasters, musicians, and video editors. In 2026, it is increasingly a preprocessing layer for AI systems that need clean speech before transcription, dubbing, localization, captioning, moderation, search, or voice-agent interaction.
That is why LALAL.AI's API placement matters. A smaller model is useful if it reduces cloud cost, latency pressure, or deployment friction. The company says Lynx is 6x smaller than Andromeda, but has not disclosed parameter count, inference cost, latency, or third-party performance scores. The commercial bet is still legible: ship a model that can serve casual creators through the browser while also fitting into developer pipelines.
Competitors are pushing in the same direction from different starting points. AudioShake raised a $14 million Series A in October 2025 led by Shine Capital, with participation from Thomson Reuters Ventures, Origin Ventures, Background Capital, Indicator Ventures, and Precursor Ventures, and sells audio separation infrastructure to media and AI customers. In April 2026, AudioShake launched Dialogue RT, a real-time dialogue isolation model aimed at live broadcast workflows. Krisp launched VIVA SDK in July 2025 for voice AI agents and said it had passed 1 billion minutes of voice AI processing per month.
LALAL.AI's lane is different. AudioShake is selling heavily into enterprise media, broadcast, and AI data workflows. Krisp is built around meetings, call centers, and voice-agent infrastructure. LALAL.AI has a creator-product funnel that includes browser upload, desktop downloads, mobile apps, plugins, paid queues, top-ups, and API access. Lynx lets LALAL.AI carry that funnel into speech cleanup without asking users to treat voice denoising as a separate product category.
The numbers are useful, with limits
LALAL.AI says its Voice & Noise stem processed more than 11.5 million audio splits in 2025 and was its second most-used separation track. The company also says registered users increased from 3.8 million in 2024 to 6.79 million in 2025. Those are company-reported operating metrics, not audited financials.
The pricing mechanics show where LALAL.AI expects users to pay. Starter is free with 10 minutes, Lite is listed at $7.50 per month billed annually with 90 Fast Queue minutes per month, and Pro is listed at $15 per month billed annually with 250 Fast Queue minutes per month. The launch promotion cuts Annual Pro by 30%, according to the X thread and LALAL.AI's site. The Pro tier is also where LALAL.AI puts API access and the VST plugin, which makes it the plan most relevant to developers and working audio producers.
There is no verified institutional funding round, disclosed valuation, or cap table in the materials reviewed. That absence is part of the story. LALAL.AI is competing in a market where venture-backed infrastructure companies are selling audio separation into enterprise AI and media accounts, while LALAL.AI appears to be using a broad self-serve product surface to turn creator demand into paid usage.
Lynx gives Orlov's team a cleaner pitch for that strategy. A smaller, speech-specific model can serve the person cleaning a field interview, the video editor salvaging dialogue from a noisy clip, and the developer routing messy audio into transcription or localization. The unanswered question is performance: whether Lynx can beat specialized competitors in the recordings that matter most, and whether LALAL.AI will publish the benchmarks needed to prove it.