Model and Runtime Downloads

Model and Runtime Downloads

Downloads provide the files required to set up the local transcription pipeline; they do not upload audio or other user data.

Current artifacts

Artifact	Why it exists	Configured download size	Validation
PianoModel.a2m	Scorer, attribute model, and frontend data	50,747,100 bytes	Exact size, outer SHA-256, bounded extraction, manifest, component sizes and hashes
CUDA runtime pack	CUDA ONNX Runtime provider	207,317,426 bytes	SHA-256, bounded extraction, metadata/layout/ABI/provider checks, helper probe
DirectML runtime pack	DirectML ONNX Runtime provider	23,948,493 bytes	Same validation layers, including real-session helper test
cuDNN for CUDA 13	CUDA neural-network runtime DLLs	349,631,454 bytes	Exact size, SHA-256, bounded extraction, required DLL discovery

The current model identity is a2m-piano, version 1.0.1, schema 1. Its required bundle components are:

• scorer.onnx
• attributes.onnx
• frontend.npz
• manifest.json

Activation is transactional

Downloads go to temporary or staging paths. A2M validates before replacing the active artifact and keeps a backup during the swap. A failed candidate is not intentionally made active. Existing validated model bundles can be restored from a backup after an interrupted replacement.

SHA-256 checks integrity against values embedded in this application version. They do not independently prove who published a file.

Network behavior

The model and managed runtime packs come from configured A2M download endpoints. cuDNN comes from NVIDIA's developer download host. Transcription itself reads the selected local file and runs locally.

See Privacy, Networking and Security and File Locations and Data Management.