How were the speaker consistency scores generated in the documentation?

[0xDigest]

Hi all,

The speaker consistency scores listed in INFERENCE.md appear to have been updated via pull request 143. How were these embeddings/scores generated? Is it possible to reproduce or validate this somehow?

[0xDigest]

Was this done by pulling the generation.decoder_hidden_states, using the last hidden state of each tuple, pooling, and doing something like cosine similarity?

[ylacombe]

I used this script to create the speaker similarity dataset per sample and then computed the stats with classic tools. Hope it helps