replace only its speaker recognition side with tight-net timers.

This is the smallest engineering challenge.

You continue, Pianote, speaker A, B, C boundaries,

but stop treating Pianote's current embeddings' latching behaviour

as the authority for identity.

Instead, collect, say, 20 to 60 seconds of clean audio belonging to speaker A,

feed that into tight-net, and search those embeddings against QDRAP.

That may solve your 1 in 30 recognition problem

without touching transcription at all.

Half this is the experiment I'd do first.

Before paying assembly AIs to our $64 a month,

I'd test tight-net locally against your existing known speakers.

It's almost absurdly cheap computationally compared with Whisper,

23 mem parameters and a 100 megabyte checkpoint.

Take, perhaps, 20 known to moon recordings,

20 of your father, 20 of you, 20 of Dame Aunty,

several different microphones, phones, environments.

Generate tight-net embeddings and measure.

Same-person cosine similarities versus different-person similarities.

That will tell us something vastly more useful

than choosing an arbitrary QDRAP threshold.

If, for example, you're using QDRAP,

for example, you discover, open quote,

the moon versus the moon.