marksverdhei

/

Qwen3-Voice-Embedding-12Hz-1.7B

like 26

Hello.
Thank you for providing standalone code for the speaker embeddings.
I've noticed an issue where differing lengths of padded batched inputs are not being considered by the model.
This results in different embeddings for a given audio segment depending on if the embedding was computed in a batch or not.
This is (presumably) unintentional behaviour ?