Open quote, audio containing a speaker, fixed dimensional speaker embedding, close quote.

You then compare that embedding against the embeddings in QDRAFT.

NVIDIA describes TITANET as a speaker representation model intended for speaker verification and diarization

and provides direct methods for extracting embeddings.

And here's the really encouraging part for your machine.

TITANET large is only about 23 to 25 million parameters.

The actual .NEMO checkpoint is only about 102 megabytes.

For comparison, even just the FP32 waits for 23 M parameters are only roughly 23 million times 4 bytes and 92 megabytes.

There will obviously be additional VRAM for activations, CUDA, PyTorch NEMO buffers, batching, etc.

NVIDIA doesn't appear to publish an official TITANET requires XGB VRAM figure, so I don't want to fabricate one.

But this is emphatically not an 8 gigabyte class model.

I would expect TITANET inference itself to occupy well under a couple of gigabytes

and probably substantially less with modest batch sizes.

More importantly, you can run it with very small...

...small batches because Berai isn't latency sensitive.

So on your RTX 3060 Ti 8...