AudioShake's The Refinery turns mixed audio into training data

What is changing?

AudioShake introduced The Refinery on October 8 to convert single-track recordings from company archives and licensed datasets into structured training data. Without requiring original stems or session files, it extracts overlapping speakers as separate, labeled audio tracks. The company says the output is isolated directly from real recordings rather than synthesized or filled in.

Built on Multi-Speaker 2.0, the workflow provides confidence scores for speaker assignment and separation quality, and can also split dialogue, music and background sound. In AudioShake's LibriCSS evaluation, the separated tracks produced 4.1 times fewer transcription errors than the tested open-source baseline. The company says early versions have processed more than 100 million minutes of audio.

The Refinery targets teams building ASR, speaker identification, TTS and conversational AI. Outputs can be delivered through an API or deployed on premises when data cannot leave an organization's environment. An official Hugging Face demo shows overlapping speech separated into individual channels and illustrates how the confidence scores work.