| --- |
| license: other |
| license_name: mixed-open-licenses |
| license_link: LICENSE.md |
| language: |
| - en |
| tags: |
| - onnx |
| - speech-recognition |
| - speaker-diarization |
| - speech-language-pathology |
| pipeline_tag: automatic-speech-recognition |
| --- |
| |
| # SLP Sidekick speech models |
|
|
| These are the ten model files that SLP Sidekick downloads the first time it runs. SLP Sidekick turns a recording of a speech-therapy session into a listening sheet for speech-language pathologists. Everything runs inside the user's browser, and recordings never leave their computer. The only thing downloaded is this model pack, once. |
|
|
| The files are unchanged copies of the ONNX conversions published by [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) (k2-fsa, Apache-2.0). Only the file names are changed, so they can sit side by side in one folder. The app checks every file against the SHA-256 checksums below before using it. |
|
|
| | File | What it does | Original model | License | Size | SHA-256 | |
| |---|---|---|---|---|---| |
| | `parakeet-tdt-0.6b-v2-encoder.int8.onnx` | Speech recognition (main) | [NVIDIA Parakeet TDT 0.6B v2](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2) | CC-BY-4.0 | 652,184,296 | `a32b12d17bbbc309d0686fbbcc2987b5e9b8333a7da83fa6b089f0a2acd651ab` | |
| | `parakeet-tdt-0.6b-v2-decoder.int8.onnx` | Speech recognition (main) | same | CC-BY-4.0 | 7,257,753 | `b6bb64963457237b900e496ee9994b59294526439fbcc1fecf705b31a15c6b4e` | |
| | `parakeet-tdt-0.6b-v2-joiner.int8.onnx` | Speech recognition (main) | same | CC-BY-4.0 | 1,739,080 | `7946164367946e7f9f29a122407c3252b680dbae9a51343eb2488d057c3c43d2` | |
| | `parakeet-tdt-0.6b-v2-tokens.txt` | Speech recognition (main) | same | CC-BY-4.0 | 9,384 | `ec182b70dd42113aff6c5372c75cac58c952443eb22322f57bbd7f53977d497d` | |
| | `whisper-small.en-encoder.int8.onnx` | Second opinion on unclear lines | [OpenAI Whisper small.en](https://github.com/openai/whisper) | MIT | 112,442,483 | `8bdac288f369aa94ee2194059238c465ed82ea9d47ee8fa4a8c0a891873e462f` | |
| | `whisper-small.en-decoder.int8.onnx` | Second opinion on unclear lines | same | MIT | 262,223,042 | `710ccf890e10f3faa15f51ec346081a2723c9f3adb6e4da81c6573a5a6f877fb` | |
| | `whisper-small.en-tokens.txt` | Second opinion on unclear lines | same | MIT | 835,554 | `306cd27f03c1a714eca7108e03d66b7dc042abe8c258b44c199a7ed9838dd930` | |
| | `pyannote-segmentation-3.0.onnx` | Who is speaking when | [pyannote segmentation-3.0](https://huggingface.co/pyannote/segmentation-3.0) | MIT | 5,992,913 | `220ad67ca923bef2fa91f2390c786097bf305bceb5e261d4af67b38e938e1079` | |
| | `titanet-small.onnx` | Telling voices apart | [NVIDIA TitaNet small](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_small) | NeMo toolkit license (Apache-2.0) | 40,257,283 | `ad4a1802485d8b34c722d2a9d04249662f2ece5d28a7a039063ca22f515a789e` | |
| | `silero-vad-v4.onnx` | Finding the speech in the recording | [Silero VAD v4](https://github.com/snakers4/silero-vad) | MIT | 643,854 | `9e2449e1087496d8d4caba907f23e0bd3f78d91fa552479bb9c23ac09cbb1fd6` | |
|
|
| Total: 1,083,585,642 bytes (about 1.1 GB). |
|
|
| ## Using them without internet |
|
|
| Download the ten files into one folder (a USB stick works). In SLP Sidekick, choose **Use a models folder** and pick that folder. The checksums are checked the same way. |
|
|
| ## Credits |
|
|
| See [LICENSE.md](LICENSE.md). The models belong to their authors; SLP Sidekick only redistributes them under their licenses. These models are not medical devices, and a listening sheet is a starting point for the clinician's own judgment, not a diagnosis. |
|
|
