Micklavin/whisper-tiny-oga-int8

Whisper Tiny multilingual INT8 for OGA

Exported from openai/whisper-tiny at immutable source revision 169d4a4341b33bc18d8881c4b69c2e104e1cc0af using the Microsoft ONNX Runtime Whisper converter. The encoder and decoder use the OGA no-beam-search graph path, INT8 symmetric weight quantization, and external ONNX tensor data. The decoder exposes output_cross_qk_0 through output_cross_qk_3. Audio input uses 80 Mel bins.

Download the complete repository snapshot. Pass its local folder to ONNX Runtime GenAI 0.17.1 using og.Model(model_directory). Keep the tokenizer, audio processor, both ONNX graphs, and external weights together. All model references and manifest paths are relative to the package directory.

ONNX checks and OGA model/processor loading passed. Transcription accuracy, Windows ML provider execution, latency, and memory are not qualified for this artifact. Generation defaults to num_beams: 1.

Cross-QK tensors are alignment inputs; this package does not supply aligned word/segment timestamps. The optional jump_times graph is intentionally omitted. The manifest keeps timestamps: false pending a separate alignment implementation.

The source model is multilingual (the English-only .en checkpoint is not used). Source licensing is Apache-2.0; see the upstream model card and applicable notices. manifest.json includes source provenance and SHA-256 hashes for every package file.