The Pulse
AudioShake’s Refinery Separates Overlapping Voices for AI Training
AudioShake introduced The Refinery on October 8, 2026, to turn existing recordings into speaker-separated training data. The company says its system has processed more than 100 million minutes of audio and reports fewer transcription errors
AI.info Team · Written with AI ·
“AudioShake has helped us process thousands of hours of clean, speaker-separated data, enabling the world’s leading labs to build better models.”
William Namgyal, CEO & Co-founder, Luel
AudioShake targets the gap between a recording and a dataset
AudioShake published details of The Refinery on October 8, 2026, pitching it as a way to convert existing audio recordings into structured material for training speech and conversational AI. The system separates voices from finished recordings, including when people interrupt or talk over one another. It does not require the original recording session’s separate tracks.
The company says the need comes from a mismatch between how people speak and how many speech datasets are prepared: training examples often contain clean, isolated voices, while ordinary conversation includes overlap, background noise and brief interjections. Archives and licensed datasets may hold those exchanges, but finished recordings usually combine the voices and other sounds into a single mix. Extracting individual speakers from that mix is the product’s central task.
Separate tracks preserve interruptions and backchannels
The Refinery builds on AudioShake’s Multi-Speaker 2.0 technology. Traditional diarization can label when two people speak at once, but does not by itself produce a separate audio track for each voice. AudioShake says its system returns labeled speaker tracks while retaining the overlap, interruptions and short acknowledgments that can matter to downstream training.
The company also describes confidence scores for speaker assignment and separation quality. Those scores are meant to help teams rank outputs, keep stronger segments and send uncertain ones for human review rather than listening through an entire archive. The product can also separate dialogue, music and background sound from finished recordings, according to AudioShake.
The 4.1-times result has a defined benchmark
AudioShake reports that separated tracks produced 4.1 times fewer transcription errors than a tested open-source separation system in its LibriCSS evaluation. Its technical evaluation identifies the comparator as MERL TF-Locoformer using a Libri2Mix checkpoint, and reports a 9.17% concatenated minimum-permutation word error rate for Multi-Speaker 2.0 versus 37.75% for that comparator. The company says both systems went through the same transcription pipeline using Whisper large-v3.
The comparison has a stated limit: AudioShake says the public TF-Locoformer checkpoint was used outside its training domain. The result therefore describes an off-the-shelf comparison, not a matched-training test that would establish one architecture’s general advantage. AudioShake’s technical write-up also says accuracy degrades when sustained speaker overlap passes roughly 20%.
AudioShake says early deployments processed 100 million minutes
AudioShake says early, private versions of The Refinery processed more than 100 million minutes of audio over the past year. It names Luel, a multimodal training-data company, and Rime, a voice-AI company, among its customers, along with unnamed AI labs and data marketplaces. The company does not break out how much audio came from each customer or how the processed material changed the performance of a trained model.
The launch makes a specific case for audio separation: existing recordings could provide training examples of speech as people actually exchange it, rather than only clean, turn-by-turn speech. The announced benchmark measures transcription errors after separation, not the gains a model gets from training on the resulting corpus. For now, AudioShake’s disclosed evidence is a large processing-volume claim, named customers and a company-run separation test with a stated out-of-domain comparison.