Who Said What, in Under 300 Milliseconds

John Gottman's Love Lab can predict divorce with 93% accuracy from a single 15-minute disagreement, and the prediction runs on conversational dynamics - who criticized, who got defensive, who interrupted whom - rather than on the content of what was said. Every ambient wearable that could apply that science passively ships with the same blind spot: it captures what was said but cannot tell who said it, at least not in real time and not at the latency a proactive agent needs to act on a live conversation. In July, that constraint fell.

TL;DR

• pyannoteAI's Live-1 model shipped in July 2026 as the first production-grade streaming speaker diarization system - real-time speaker attribution at sub-300ms latency matching batch-mode accuracy.
• Speaker diarization - knowing who said what and when - is the missing prerequisite for relationship-aware ambient AI. Without it the wearable is a voice recorder; with it, the wearable builds a longitudinal graph of how you communicate with specific people.
• Gottman's research predicts relationship outcomes with 93% accuracy from 15 minutes of conversational dynamics - turn-taking, interruption patterns, contempt markers. Those signals require per-speaker attribution that only streaming diarization provides.
• The defensible product is the relationship graph: months of who-said-what data between specific people that no competitor can bootstrap on day one. The transcription itself is a commodity.

Why Does the Ambient AI Wearable Need Real-Time Speaker Attribution?

The wearable AI space has converged on a basic pipeline: always-on mic captures audio, ASR produces a transcript, and an LLM summarizes or answers questions about it. Every pendant on the market runs it. The transcript is a commodity.

But a transcript without speaker labels is a monologue from the wearable's perspective. It cannot distinguish your words from your partner's, cannot track who interrupted whom, and cannot detect that you used criticism while your partner used defensiveness - the two Gottman horsemen that, when they appear together, predict relationship deterioration with clinical precision.

Speaker diarization is the process of segmenting audio into "who spoke when." Until July 2026, production-grade diarization required batch processing: upload the full recording, wait, get results. That works for meeting notes and fails completely for a proactive agent that needs to intervene during a conversation. pyannoteAI Live-1 changes the constraint - sub-300ms latency, production accuracy matching their Precision-2 batch model, WebSocket streaming alongside any STT pipeline. Twelve years of diarization research compressed into a system that finally runs at conversation speed.

The Gottman Bridge

John Gottman's Love Lab research at the University of Washington observed over 3,000 couples and identified four conversational markers - criticism, contempt, defensiveness, and stonewalling - that predict divorce with 93% accuracy from a single 15-minute disagreement conversation. The prediction does not require understanding the content of what was said. It requires understanding the pattern of how it was said and by whom, which makes it a speaker-attributed signal by definition: you cannot detect a criticism-defensiveness escalation cycle without knowing which speaker is doing which, and you cannot measure the 5:1 positive-to-negative ratio that healthy relationships maintain without tracking sentiment per speaker across time.

The research is decades old. What streaming diarization unlocks in 2026 is the ability to apply it passively, continuously, from a body-worn device.

When I shipped smart glasses at Meta, we had always-on microphones and streaming ASR. What we did not have was real-time attribution good enough to build relational features on. The five-mic array on Ray-Ban Meta captures clean audio in noisy environments, but knowing who is speaking in a two-person conversation at dinner is a different problem than far-field noise rejection - it requires a model trained on millions of hours of conversational turn-taking. That model now exists as a production API.

The Five-Layer Stack

The architecture for a relationship-aware ambient wearable becomes concrete:

• Layer 1: Streaming ASR - word-level transcription from Whisper, Deepgram, or Nemotron. Commodity.
• Layer 2: Streaming diarization - pyannoteAI Live-1 running in parallel, attributing each word to a speaker. Newly possible at production quality.
• Layer 3: Relationship graph - longitudinal accumulation of per-speaker communication patterns. Who dominates, who withdraws, how dynamics shift over weeks and months.
• Layer 4: Pattern detection - Gottman's four horsemen, attachment styles, communication preferences. Applied to the graph, not individual utterances.
• Layer 5: Intervention - the coaching agent that surfaces an insight at the right moment because it has both the physiological context (SensorFM's when-to-say-it layer or equivalent) and the relational context (from the speaker-attributed graph).

The competitive moat sits at Layer 3. Transcription is commodity, diarization is now a purchasable API, and pattern detection is published behavioral science. But the graph - months of who-said-what between you and your partner, your manager, your closest friends - takes months to build and cannot be cold-started.

Why Now, Why Hard

Three things converged. Streaming diarization reached production quality in July 2026. The wearable form factor hit mainstream with Meta Glasses at $299 running a wearable-first model. And the relationship-AI market exploded - Nirva raised $8M for a mood-tracking pendant, Bee added relationship suggestions, Gottman's own Affective Software Inc. launched a digital assessment platform.

Everyone has the same hardware: mic, battery, Bluetooth. Everyone has access to the same ASR APIs. The differentiation is what you do with speaker-attributed conversation data over time. The pendant that knows you argued about money again on Tuesday, that your partner showed contempt three times in five minutes, that this pattern has been escalating for six weeks - that pendant is a relationship early-warning system, and people will pay for it in a way they never paid for note-takers.

The last technical blocker went away this month. What remains - the consent architecture, the trust UX, the months of continuous wear needed to accumulate the graph - is execution. Which is harder than it sounds, and it already sounds hard.