Whisper (sherpa-onnx) com atenção cruzada
Modelos OpenAI Whisper (tiny, base, small, turbo) exportados com
export-onnx-with-attention.py
do sherpa-onnx (Apache-2.0) e quantizados em int8.
A saída cross_attention_weights permite timestamps por palavra (DTW).
Compatíveis com ONNX Runtime >= 1.17 (IR 7, opset 13).
Pesos originais: OpenAI Whisper, licença MIT (Copyright (c) 2022 OpenAI).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support