Multi-Speaker 2.0 is live–the world’s first low and high-resolution multi-speaker separator. One recording of several people in. A clean labeled track per person out, including the moments when two people talk at once. 8 kHz to 48 kHz, one model. Confidence scores every 20ms.

Accessibility isn’t a someday project anymore, and that includes the archive. Captions, audio description, and dubbing all need dialogue and M&E on separate tracks. Most catalog titles only have the finished mix. We separate them, so the work can start. 👇

Live dialogue isolation has a 9 to 15 ms window before lip-sync slips. Dialogue RT runs at 11 ms on a single Dell Pro Max with GB10. Dell wrote up how it got from an NAB demo to production: dell.com/en-us/blog/more-tha…

Editing a film or ad and need "a song about starting over"? You can now search @audionetwork's catalog by lyrics, not just genre and mood 🎶 Built with @AudioShakeAI's lyric transcription + alignment API. Case study: audioshake.ai/case-studies/u…

A finished mix used to mean dialogue, music and effects were fused for good. At Mix Sound for Film, Suzanne Kirkland and Stephen Oliver shared what changes when they're not: archives that can be reused, sports libraries that work across streaming, and more localized content.

Most voice AI demos run on clean audio. Real callers interrupt, talk over each other, + put you on speaker in the car. Thanks @RimeLabs, @LilyJClifford, @BrookeHopkins (@CovalDev), Josh Burnell (@Asurion) and @ShivIndap (@Upstart) for digging into that gap with us at #SFTechWeek

"Enormous amounts of that conversation already exist. The challenge is in turning it into usable training data." @themoko on @AudioShakeAI’s Refinery. audioshake.ai/products/refin…

Real conversation is messy and overlapping. Most training data isn't. The Refinery splits real-world recordings into one track per speaker and scores each for quality. Hundreds of millions of minutes processed with frontier labs. audioshake.ai/products/refin…

Most audio in the world was never recorded on separate tracks. @themoko breaks down source separation on @Dell's #ReshapingWorkflows with @LoganatDell, starting with how we pull vocals, instruments, or a single speaker out of a finished mix.

Most archive audio is stuck in a finished mix, which makes dubbing, clearing music and re-licensing hard. Suzanne Kirkland talked through how @AudioShakeAI's separation, @ortanamg, and @AWS Marketplace fit together to fix that, with Ortana and AWS at IBC.

Voice agents fail in predictable ways: overlapping speakers, background noise, crosstalk the model was never trained to hear. On Wednesday, @themoko and @lilyjclifford (@RimeLabs) get into what teams learn when they build that stack themselves. RSVP: partiful.com/e/qRX76YKI93J2I…

NVIDIA's Nemotron 3 diarizer hit #1 on Voice Arena. We ran it against ours on LibriCSS (8-speaker recordings). Diarizer vs. diarizer, no overlap: 35% fewer word errors. With Multi-Speaker isolating each voice, all recordings: 52% fewer word errors than Nemotron 3.

Upload a file → get a cue sheet back. @AudioShakeAI's music ID is now in @MusicReportsInc' Cuetrak®. We separate the music out of the mix first, then identify each cue — so buried and unlabeled tracks still get caught. audioshake.ai/press-releases…

A therapy podcast with 70+ episodes was about to lose its whole archive in a station move — licensed bumper music, baked into the mix, no session files. AudioShake's Music Removal stripped the cues out of every episode, leaving the speech untouched. audioshake.ai/case-studies/f…

Voice agents: great in the demo, harder on a real call. Join @AudioShakeAI + @RimeKLabs at #SFTechWeek for Voice AI in the Real World, a panel on training data, failure modes, and what teams built vs bought. Oct 7 · 575 Market St partiful.com/e/qRX76YKI93J2I…

"You have to hit between 10 and 15 milliseconds." @themoko and @LoganatDell on why Dialogue RT–speech isolation for broadcasters runs on a Dell Pro Max with GB10 — and what a workstation unlocks that the cloud doesn't.

Dialogue RT isolates dialogue from a live feed in 11 ms end to end — measured model input to isolated output. That's not noise suppression. It's two stems: dialogue, and everything else. On NVIDIA DGX Spark and Blackwell architecture GPUs. audioshake.ai/post/isolating…

Real conversation isn't tidy. People interrupt, agree over each other, laugh through the answer. This is the part every other tool gives up on. Multi-Speaker 2.0 keeps both voices as separate tracks, tells you how confident it is, and runs on anything from 8 to 48 kHz.

Four days at IBC, hundreds of conversations. The question we heard most: “Can it run on the feed, not the file?” Answer: yes. Live, less than 11 milliseconds. Another great year in the books.

Most audio in the world was never recorded on separate tracks. Jessica Powell breaks down source separation on Dell's #ReshapingWorkflows — how we pull vocals, instruments, or a single speaker out of a finished mix. Full episode: podcasts.apple.com/us/podcas…