but Titanet is not the thing I'd expect to cause difficulty.

And unlike Whisper, speaker embedding extraction is comparatively lightweight.

You could even run Titanet on CPU if necessary,

and reserve the GPU entirely for Whisper.

But Titanet is not a drop-in replacement for all of Highlighter.

There's an important distinction.

Titanet large by itself answers.

Open quote, what does this speaker sound like?

Close quote, it produces the embedding.

It doesn't by itself answer.

Open quote, speaker one talks from 3.2 to 17.6 seconds.

Speaker two talks from 17.6 to 21.3 seconds.

Close quote, that's diorization.

NVIDIA has its own full NEMO diorization system.

One option uses Titanet MSDD, multi-scale diorization decoder.

NVIDIA's telephonic MSDD package, for example,

combines Titanet L with an MSDD model of only around 5.8 metaparameters.

So there are actually three interesting architectures for V-Ri.

Option A, assembly AI plus Titanet.

This is the cleanest technically.

Assembly AI, transcription, word timing, speaker A, B, C, D, then locally.

Titanet embeddings for speaker A's accumulated audio.

Q-drant, speaker A equals the moon.

You don't need Pioneer to talk.