Memory inherits every mistake the capture makes. Sonal Labs researches overlap-native attribution and builds the model and memory layer that depend on it.
“Patient denies chest pain.”
A cardiac symptom goes unrecorded.
Conventional pipelines make a forced choice when two voices land in the same frame. They pick the dominant speaker, drop the other, and hand a plausible-looking transcript to whatever comes next — a discharge summary, a referral, a care plan.
The error doesn't stay in the transcript. It becomes a symptom on the wrong chart. A deposition that collapses two voices into one. A care plan the patient never agreed to.
A memory system that occasionally gets words wrong is imperfect. A memory system that assigns clinical facts to the wrong person is adversarial.
A proprietary diarisation model built for the Cocktail Party Problem. Disentangles simultaneous voices, fingerprints speakers, stays stable through interruption and repair.
Optional video context — from smart glasses, CCTV, and cameras — fused with the audio track. Resolves ambiguity the microphone alone can't.
Sonal doesn't just record. It infers intent and triggers workflows. Evidence-grounded action items, reminders, follow-ups, and API-level actions.
A survey of overlap-aware techniques (EEND, TS-VAD, OSD), a reproducible baseline on open foundations, and an overlap-stratified evaluation methodology.
The dominant paradigm in speech AI treats “meetings” as the canonical multi-speaker scenario. Real conversation doesn't live there. Life unfolds in morning walks, phone calls while driving, family dinners, errands, worship — precisely where current systems fail most catastrophically. The paper formalises a different starting point.
The native overlap LLM at the centre of every Sonal system. Its job is attribution — knowing, frame by frame, who is holding the floor and what they said.
Built on EEND, TS-VAD, and open ASR foundations. Measured on overlap-stratified benchmarks, not aggregate numbers that hide the failure mode.
Continuously predicts where the current turn is going and when it will end, conditioned on prosody, syntax and context.
Holds, yields, asks to continue, or interrupts — the same decisions a good human interlocutor makes, as an explicit training objective.
Handles self- and other-repair. False starts, reformulations, mishearings — not errors, structure to be preserved.
Optional lip-movement and facial-cue conditioning for disambiguation when the acoustic signal alone isn't enough.
Doctor discusses X-ray results (audio) + image of the X-ray (video).
Generates a comprehensive patient visit summary — attributed, signed, auditable.
Multiple attorneys arguing (overlap event).
Verbatim diarisation of who said what, with timestamps suitable for record.
Camera sees FedEx truck + person dropping a box.
Logs ‘Parcel delivered’ event, alerts the homeowner via app.
Friend mentions an anniversary date in passing.
Auto-creates a calendar event — ‘Buy a gift for Sarah's anniversary’ — with the clip as evidence.
Whisper, Canary, Parakeet, Nova. Basic transcription is no longer a moat — it's infrastructure. Costs down ~90 % since 2020.
EEND, TS-VAD and overlap-aware techniques exist in research but aren't productised. This is the last hard problem in speech AI.
The hardware is ready and the market is growing 25 % a year. The missing piece is a software layer you can trust.
“Overlap-native diarisation is the technical foundation, not the product. The product is trustworthy memory.”