How to deploy offline？

[Kevin-BKQ]

How to localize and deploy the pyannote/seech-separation ami-1.0 project??? I downloaded pyannote/parallelism ami-1.0 locally, but the speech brain/ spkrec-ecapa-voxceleb@5c0be3875fda05e81f3c004ed8c7c06be308de1e How to adapt this?

[arcticloud]

The thing to understand is that a pyannote pipeline is not one model — its config.yaml references several hosted models (for diarization: a segmentation model + an embedding model; for speech-separation, a separation model too). "Deploy offline" means caching every one of those and then telling the stack not to phone home.

Reliable recipe:

1. On a machine with internet, pre-download everything into the HF cache:

   huggingface-cli download pyannote/speech-separation-ami-1.0
   huggingface-cli download pyannote/segmentation-3.0
   # plus whatever embedding/separation model the config.yaml lists

   Open the pipeline's config.yaml and download each model id you see under pipeline.params — that's the authoritative list, so you don't miss one.

2. Copy the cache (~/.cache/huggingface/) to the offline box. If the pipeline pulls a SpeechBrain model (older ECAPA), also copy SpeechBrain's pretrained_models/ dir, since SpeechBrain caches separately from HF.

3. Force offline mode so nothing tries to reach the network:

   export HF_HUB_OFFLINE=1
   export TRANSFORMERS_OFFLINE=1

   Then Pipeline.from_pretrained(...) loads entirely from the local cache.

If you'd rather not rely on the cache at all, you can instead point config.yaml at local file paths for each model and load the pipeline from that local yaml. The hash you saw (spkrec-ecapa-voxceleb@5c0be38...) is just a pinned revision — once it's in the cache at that revision, offline mode resolves it fine.