Qwen3-ASR-0.6B-Encoder-XAttn

The audio encoder of Qwen/Qwen3-ASR-0.6B, extracted as a standalone transformers model (trust_remote_code=True), 18 layers, d_model 896. Weights are unchanged from Qwen3-ASR-0.6B.

This variant adds the code for DuplexJev's cross-attention fusion block (Q = layer 18, K = layer 14, V = layer 9). The encoder weights are identical to the plain variant; the trained fusion weights ship with each DuplexJev-A connector.

Used by: DuplexJev-A-* (cross-attention connectors). You normally do not load it yourself: Decider.from_pretrained("adventists-ai/DuplexJev-...") fetches it.

License: Apache-2.0, as the original Qwen3-ASR-0.6B. See https://github.com/adventists-ai/duplexjev.