# Portable Speaker Head

This directory was exported from an Ascend/NPU checkpoint. It contains inference-only, CPU tensors and does not require `torch_npu` to load. Prefer `speaker_head.safetensors` in CUDA/runtime integrations. `speaker_head_cpu.pt` is an equivalent CPU-only PyTorch checkpoint with payload key `speaker_head`. See `export_manifest.json` for hashes and exact validation results.