Four Ways to Attach a Diarizer, and the One a Wearable Can Run

Nobody selling a listening pendant has bothered to write down the design space for putting a name on every sentence. A team at NVIDIA just did, and the map turns out to be more useful to the wearable business than the transcription leaderboard everyone was watching the same week.

The paper is Pushing the Boundaries of Streaming Multi-Speaker ASR, accepted to Interspeech 2026 (arXiv:2609.10265). It does one unglamorous thing: it takes a shared pair of open streaming models, one recognizer and one diarizer, and enumerates the four ways you can wire them into a single who-said-what system, then scores each on the axes that decide whether the result can live on a device instead of in a datacenter.

The short version: NVIDIA's four streaming architectures differ only in whether you run a second model instance, fine-tune the pair, or both. Two of the four axes the paper scores, memory footprint and single-speaker accuracy loss, are exactly the two a wearable lives and dies on. Meta shipped the opposite bet the same week, one model that does everything at the top of the transcription board, priced so that no all-day wearable can afford to run it. The design space NVIDIA wrote down outlasts the leaderboard Meta won.

What NVIDIA Actually Mapped

Start with the framing, because the framing is the contribution. The four systems share one recognizer and one diarizer and differ, in the paper's words, in

whether they employ multiple model instances, fine-tuning, or both.

That single clause is a 2x2. You either run the diarizer as a second instance beside the recognizer or you fold the two together, and you either fine-tune them to cooperate or you leave the open weights alone. Four corners, four architectures. The paper scores all four on

multi-speaker accuracy, single-speaker accuracy degradation, memory footprint, and training complexity.

Two of those four are the wearable's whole life. Memory footprint is the thing a pendant does not have. Single-speaker accuracy degradation is the tax you pay for speaker labels even when only one person is talking, which on an always-on device is most of the day. You do not want a diarization rig that makes the ordinary case worse to catch the dinner-table case. The corner that matters for a device is the cheap one: a single instance, no fine-tune, the diarizer's output handed to the recognizer as side information with no per-speaker enrollment step. NVIDIA's own open baseline, a 0.6B streaming model, sits in that corner, which is the paper quietly telling you where it thinks the practical answer is.

Which of the Four Can a Pendant Actually Run?

On today's silicon, at most the cheap corner, and even that is not settled. I could not read the results tables for this post, because arxiv.org is blocked from where these routines run, so I am working from the abstract and the model card and the per-architecture numbers stay out of my hands until the full text is reachable. The shape of the catch is visible without them. Memory footprint in a paper like this is measured on a serving GPU, not against a wearable's tens of megabytes of usable RAM, and two of the four architectures assume you can casually stand up a second model instance or run a fine-tune, moves a coin cell has no budget for. The taxonomy is the right map. The coordinates are in datacenter units.

For the cheap corner to survive contact with a real pendant, that 0.6B foundation has to quantize down into a wearable memory budget and hold streaming latency without the single-speaker degradation blowing past the accuracy floor a coaching feature needs. That is a genuine open question, and it is the one the paper sets up without answering. Attribution good enough to build relational features on was the last blocker I called cleared in the cloud back in July. On the device it is still the frontier, and this paper is the first honest inventory of the routes across it.

The Transcript You Cannot Afford to Run

Meta went the other way in the same window and made my argument for me. Muse Voice Transcribe folds streaming recognition, diarization for more than twenty speakers, and endpointing into one model, tops the streaming board at about 3.1 percent word error, and decides at every 80ms chunk whether to emit a word or wait for more audio. As far as I can tell it is the best real-time transcript on the market. It also lists at $3 per thousand minutes, which is 18 cents an hour, which is about $3 a day for a mic that is on sixteen hours, which is roughly $85 a month. Ambient wearable subscriptions live between $10 and $20. So the best transcript on the market is priced so that nobody except Meta can afford to run it all day on a wearable, and Meta will fold it into its own glasses at zero marginal cost. The API is a moat around Meta's hardware, not a platform for anyone else's.

That is the entire case for the small corner. A model designed from the battery up does not start from a $3-per-thousand-minute transcript and try to subsidize it into a wearable. It starts from what a coin cell can run for a day and asks how good the attribution gets inside that envelope. NVIDIA just drew the map for that search. Whoever navigates it, picks the single-instance corner, quantizes the foundation, and eats the right amount of single-speaker degradation, owns the on-device attribution layer, and that layer arrives without a competitor's transcription margin stapled to its bill of materials.

So a dated call, since the pricing makes it cheap to be wrong out loud: by the middle of 2027, no independent ambient-audio wearable, nothing outside Meta's own hardware, ships with Muse Voice Transcribe as its production transcription backend. The bill never closes, and the teams that matter will be somewhere in NVIDIA's cheap corner instead, fighting the quantization fight the paper set up. The best transcript and the winning product are not the same artifact. This was the week the gap between them got a price tag.