| --- |
| license: mit |
| base_model: openai/whisper-tiny |
| tags: |
| - onnx |
| - whisper |
| - speech-recognition |
| - unity |
| library_name: onnxruntime |
| --- |
| |
| # Whisper Tiny ONNX |
|
|
| Selected ONNX exports of [OpenAI Whisper Tiny](https://huggingface.co/openai/whisper-tiny) for ONNX Runtime and Unity AI Inference 2.6.1 (Sentis). |
|
|
| The FP32 split decoder is sourced unchanged from [onnx-community/whisper-tiny](https://huggingface.co/onnx-community/whisper-tiny/tree/ff4177021cc41f7db950912b73ea4fdf7d01d8e7/onnx) revision `ff4177021cc41f7db950912b73ea4fdf7d01d8e7`. The BNB4 files are intended for ONNX Runtime. KitsuMate provides the ONNX mel preprocessing graph and this packaging; OpenAI and the ONNX Community contributors retain authorship of the model and conversion work. |
|
|
| ## Engine support |
|
|
| | Files | Precision | ONNX Runtime | Unity AI Inference 2.6.1 | |
| | --- | --- | --- | --- | |
| | `encoder_model.onnx`, `decoder_model.onnx`, `decoder_with_past_model.onnx` | FP32 | Yes | Yes, CPU validated | |
| | `encoder_model_bnb4.onnx`, `decoder_model_merged_bnb4.onnx` | BNB4 | Yes | No | |
| | `mel.onnx`, `tokenizer.json` | Shared preprocessing | Yes | Yes | |
|
|
| The unsuffixed files are FP32. All ONNX files are kept directly under `onnx/` so additional precisions can use filename suffixes without duplicating shared assets. |
|
|
| ## Files |
|
|
| | File | Bytes | SHA-256 | |
| | --- | ---: | --- | |
| | `onnx/mel.onnx` | 1,354,556 | `0fb448b82bde665099d8532502dd1d7f95751f3afcd3760e7b30a94ca0bddebf` | |
| | `onnx/encoder_model.onnx` | 32,904,992 | `6642befb640f950d4a8cbbd17834d59e7e75f575b81ccf213e06b050623ab1dd` | |
| | `onnx/decoder_model.onnx` | 118,397,483 | `ab79e3f2a9a3d98f159f853a3172120a38af7eb5f7863d706aa7d39c228f009e` | |
| | `onnx/decoder_with_past_model.onnx` | 113,638,998 | `0485135066eb1d36dcb04dbabd0cc1141c7cd8c442217abd0798d55fe2bc6bed` | |
| | `onnx/encoder_model_bnb4.onnx` | 8,578,451 | `c8596d911d020092caa4d80bf06929f4a0c5f573701a6882bc37dd761dc383bf` | |
| | `onnx/decoder_model_merged_bnb4.onnx` | 86,124,414 | `fd66992a760913909e01c151ea4b42de738ab7a5c6268ed4399fda545cb8dfbf` | |
| | `tokenizer.json` | 2,480,466 | `27fc476bfe7f17299480be2273fc0608e4d5a99aba2ab5dec5374b4482d1a566` | |
|
|
| ## License |
|
|
| Whisper model code and weights are released by OpenAI under the MIT License. See `LICENSE`. |
|
|
