On 8 October 2026 AudioShake, an audio-separation company, announced The Refinery. It takes recordings where voices overlap and returns a separate, scored track for each speaker, for training AI. Every number is AudioShake’s own; no independent test exists.
What the company says it does
Confirmed that AudioShake announced it; the claims are its own. The Refinery, AudioShake says, “takes raw, real-world audio and returns structured, training-ready data” [1]. It works “directly from recorded audio, with no original stems or session files needed” (a stem is a separated track), and the stems are “isolated from the real recording itself — never synthesized, hallucinated, or filled in” [1].
Every multi-speaker output “includes confidence scores for speaker assignment and separation quality, so teams can filter and rank a dataset programmatically” [1]. It can also “identify and remove copyrighted music from training data” [1]. It is offered by API or inside a customer’s own systems [2].
What “4.1x fewer errors” means
AudioShake says that “in our LibriCSS evaluation, the separated tracks produced 4.1× fewer transcription errors than the tested open-source separation baseline” [1]. LibriCSS is a test set of overlapping speech. In AudioShake’s write-up the error rate was 9.17% against 37.75% for an open-source model that, in AudioShake’s words, was tested “outside its training domain” [3]. That is a comparison with an off-the-shelf model, run by the company.
Usage claims
AudioShake says it has “deployed early versions of The Refinery privately with frontier AI labs, and processed 100M+ minutes of audio” over the past year [1]. It names two customers, Luel (training data) and Rime (voice AI), plus unnamed “largest AI labs and data marketplaces” [1]. Luel’s chief executive, William Namgyal, is quoted: “AudioShake has helped us process thousands of hours of clean, speaker-separated data, enabling the world’s leading labs to build better models.” [1]
What this does not show
Pricing, which labs use the tool, and any independent evaluation are not stated, nor is how much model quality improves.
Related on TSN: micro1 Says It Will Spend $1bn on Company Data to Train AI Agents; Mecka AI and Parallel Systems: Paying for Robot Data and Autonomous Rail
Sources
- AudioShake, “Why We Built The Refinery: Teaching Machines How People Really Talk”, 8 October 2026 (company blog; source of the quotes and claims). https://www.audioshake.ai/post/why-we-built-the-refinery
- AudioShake, “Refinery” product page, read 10 October 2026 (company page; API and on-premises delivery). https://www.audioshake.ai/products/refinery
- AudioShake, “Multi-Speaker 2.0 technical evaluation” (company write-up; LibriCSS error rates and the baseline’s limits). https://www.audioshake.ai/post/multi-speaker-2-0-technical-evaluation
