Modelpublic

Youngwon/whisper-large-v3-turbo_timestamped-external-data

sourceHugging Facemitupdated 1d agoView on Hugging Face

0likes16downloads

Model Card

whisper-large-v3-turbo_timestamped (external data)

This repository re-packages the q4 and q4f16 ONNX weights from onnx-community/whisper-large-v3-turbo_timestamped (revision b3f77bf9a8c4d5ea3415827033d1ffea7955fd9a), an ONNX export of openai/whisper-large-v3-turbo (MIT).

The only change is the file layout. Top-level weights moved from inside the .onnx file into a separate .onnx_data file (ONNX external data). The weight values are the same. With this layout, onnxruntime-web (WebGPU) can upload the weights straight to the GPU, which uses less WASM memory in the browser. Tensors inside subgraphs stay in the .onnx file.

Used by transbee.

file	weights
`onnx/encoder_model_q4f16.onnx`	`onnx/encoder_model_q4f16.onnx_data`
`onnx/decoder_model_merged_q4f16.onnx`	`onnx/decoder_model_merged_q4f16.onnx_data`
`onnx/encoder_model_q4.onnx`	`onnx/encoder_model_q4.onnx_data`
`onnx/decoder_model_merged_q4.onnx`	`onnx/decoder_model_merged_q4.onnx_data`

whisper-large-v3-turbo_timestamped (external data)

License