| --- |
| license: mit |
| library_name: mlx |
| pipeline_tag: audio-to-audio |
| tags: |
| - speech-enhancement |
| - speech-denoising |
| - gtcrn |
| - mlx |
| - apple-silicon |
| - onnx-conversion |
| --- |
| |
| # GTCRN Simple - MLX |
|
|
| Apple MLX conversion of the GTCRN Simple streaming speech-enhancement model. |
| This repository contains FP32 safetensors deterministically converted from the |
| pinned Sherpa-ONNX release asset. It is a layout conversion only: there is no |
| retraining, quantization, fusion, or weight modification beyond the layout |
| required by MLX convolutions. |
|
|
| ## Files |
|
|
| - `model.safetensors` - 208,427-byte FP32 MLX weight file. Its SHA-256 is |
| `d503ed83a35b91f66af97a6a1b07a76425988db7d9744556d0565c336a7e6e3b`. |
| - `conversion.json` - deterministic conversion record binding the source ONNX, |
| audit, converter revision, weight SHA-256, byte count, and 179 tensors. |
| - `metadata.json` - GTCRN Simple protocol metadata shared with the pinned ONNX |
| asset. Its SHA-256 is |
| `070f320defc15d4e1781b0c5b9225d9f91031ed5f14f675efa497f244e640615`. |
| - `LICENSE` - the upstream GTCRN MIT license. |
|
|
| ## Provenance |
|
|
| - Upstream implementation and model: [Xiaobin-Rong/gtcrn](https://github.com/Xiaobin-Rong/gtcrn) |
| at `9202557caa577baf2ec2220bef82ba9f4b589dc1`. |
| - Pinned ONNX source: the Sherpa-ONNX `gtcrn_simple.onnx` release asset, SHA-256 |
| `e77603ac0c23dac3227dd2d7135b3a585cbee2679048aecfa886657d3ae1b534`. |
| - Conversion: [model-gtcrn](https://github.com/agentable/model-gtcrn) |
| `scripts/convert_onnx_to_mlx.py`, using its committed conversion audit. |
|
|
| The converted file preserves ONNX initializer names. ONNX Conv weights become |
| MLX OHWI; grouped ConvTranspose weights use the audited group-aware layout |
| transform. GRU, normalization, PReLU, matrix, and structural parameters retain |
| their source layouts as required by the MLX runner. |
|
|
| ## Validation |
|
|
| The MLX runner replays the pinned one-frame GTCRN graph, including all three |
| recurrent cache families. It is validated against the committed ONNX frame and |
| cache oracle with maximum absolute error below `5e-5` on Metal FP32 execution. |
| End-to-end waveform and corpus-quality qualification are not part of this |
| artifact release; they remain required gates in the consumer library before a |
| production release. |
|
|
| ## Usage |
|
|
| This artifact is consumed by the `engine: mlx` backend in |
| [model-gtcrn](https://github.com/agentable/model-gtcrn) on Apple Silicon with |
| the `mlx` build tag. The consumer manifest pins this repository at an immutable |
| commit and verifies the asset SHA-256 and byte count before MLX allocation. |
|
|
| ## License |
|
|
| The upstream GTCRN model is distributed under the MIT license. This derived |
| format conversion is redistributed under that same license; see `LICENSE`. |
|
|
