admion-whisper-small-int8

Admion Whisper Small INT8

This package contains an ONNX INT8 conversion of the multilingual OpenAI Whisper Small model for Admion's optional, fully offline transcription.

Provenance

• Original model: OpenAI Whisper Small multilingual
• Original checkpoint SHA-256: 9ecf779972d90ba49c06d968637d720dd632c55bbf19d441fb42bf17a411e794
• OpenAI Whisper revision: 31243bad24cc746f07d4c8bfdd2d974872cb1803
• sherpa-onnx exporter revision: 3dc7c569f31ca2cd4a20ed6f7db780327e6714c5
• Export command: python export-onnx.py --model small
• Quantization: ONNX Runtime dynamic QInt8 for MatMul weights

manifest.json records the complete pinned Python environment and the size and SHA-256 digest of every downloadable model file.

Licenses

OpenAI Whisper code and model weights are distributed under the MIT License. The sherpa-onnx exporter is distributed under the Apache License 2.0. Complete copies are included as LICENSE-OpenAI-Whisper.txt and LICENSE-sherpa-onnx.txt.

Intended use and limitations

Admion uses this model only on-device and treats every transcript and extracted value as an editable suggestion. MVP automatic transcription is limited to recordings of 20 seconds or less. Accuracy varies by language, speaker, device, noise, names, numbers, and code-switching. The original audio is retained until the user chooses to delete it.

Aggregate download statistics

Admion requests config.json once during each model installation. Hugging Face uses requests for this recognized query file to calculate its aggregate model download statistics. The count can include retries and is not a unique-user counter; Admion sends no capture content or app-generated identifier.