Luel Multi-Speaker Separation

Luel Multi-Speaker Separation is Luel's multi-speaker separation product: from one conversation recording, it returns a clean track per person. Blog: https://www.luel.ai/blog/luel-multi-speaker-separation . Currently in closed beta.

Motivation: podcasts, meetings, interviews, and calls are almost always a single mixed track—you can turn the whole mix up or down, but you cannot lift one speaker out. This product sends back each person as their own track from first word to last.

Highlights:

• Trained on real conversations from Luel's contributor network (pauses, laughs, pace changes, near/far mics), not only short clean clips;
• On held-out real calls, Luel reports every person got their own track, nobody swapped onto someone else's, and ~99.3% of solo speech landed on the right track;
• Light on hardware: roughly real-time on modest cloud GPUs; handles 2–4 people, up to ~90 minutes, audio up to 48 kHz, returning tracks at the input rate;
• Also returns a who-talks-when timeline; non-person sounds (laughs, coughs) can land in a separate file;
• Strongest when people take turns; crosstalk, shared-room mics, and more languages are on the roadmap; most testing so far is English.

Problem

Mixed conversation audio is hard to edit per speaker, transcribe cleanly, dub, or turn into speaker-level training data. Separation unlocks per-person editing, transcription, dubbing, turn-taking study, and usable corpora from long real talks.

Design & limits

• Emphasizes whole-recording consistency: same person stays on the same track for an hour, not just short windows;
• Metrics are Luel's held-out call results—see their model page for the full board;
• Limit: on some calls (Luel cites up to about one long two-person call in four), an extra background/speech fragment track may appear—listen before relying;
• Closed beta; access by request.

Relation to Luel Speech

Same company, different product: Speech is low-resource TTS; this is conversation source separation.

Who it's for

• Podcast/meeting post, interview cleanup, call QA;
• ASR and voice-data teams that need per-speaker stems.

Getting started

1. Read https://www.luel.ai/blog/luel-multi-speaker-separation and the model-page results;
2. Request closed-beta access and describe your audio (speakers, length, language);
3. Submit 2–4 speaker recordings up to ~90 minutes; audition every returned track, watching for extras;
4. Feed stems into edit/ASR/data pipelines with a human check for anomalous tracks.