knn-vc ONNX export (fp32+int8) with parity reports

PROVENANCE.md

@@ -0,0 +1,41 @@
+# PROVENANCE
+
+| Field | Value |
+|-------|-------|
+| upstream_repo | https://github.com/bshall/knn-vc |
+| upstream_ref | v0.1 |
+| export_script_sha | 93765ec8873cd903289cfd9949aecb06ca68aff0 |
+| export_date | 2026-06-11T13:33:22Z |
+| torch_version | 2.10.0+cu128 |
+| onnx_version | 1.20.1 |
+| platform | Linux-6.18.20-1-lts-x86_64-with-glibc2.43 |
+| wavlm_model | microsoft/wavlm-large |
+| hifigan_checkpoint | prematch_g_02500000.pt |
+| wavlm_layer_extracted | 6 |
+
+## Upstream licence
+
+```
+MIT License
+
+Copyright (c) 2023 bshall (Stellenbosch University)
+Copyright (c) 2021 Microsoft Corporation (WavLM-Large weights)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+```

README.md

@@ -0,0 +1,41 @@
+---
+license: mit
+library_name: onnxruntime
+tags:
+- voice-conversion
+- onnx
+- vconnx
+- knn-vc
+pipeline_tag: audio-to-audio
+---
+
+# vconnx · kNN-VC (pure ONNX)
+
+ONNX export of [kNN-VC](https://github.com/bshall/knn-vc) (Baas et al.,
+Interspeech 2023, MIT) for the **vconnx** voice-conversion library:
+WavLM-Large layer-6 feature extractor + prematched HiFi-GAN vocoder; the
+kNN regression step is pure numpy at inference time. Any-to-any voice
+conversion at 16 kHz, CPU-only.
+
+| file | role | size |
+|---|---|---|
+| `wavlm_layer6.onnx` | feature extractor (fp32) | 387 MB |
+| `wavlm_layer6_q8.onnx` | feature extractor (int8) | 98 MB |
+| `hifigan_knnvc.onnx` | vocoder (fp32) | 63 MB |
+| `hifigan_knnvc_q8.onnx` | vocoder (int8) | 25 MB |
+
+Export parity vs the torch reference (max abs):
+WavLM see report ·
+HiFi-GAN see report
+(full reports in the repo).
+
+## Usage
+
+```python
+from vconnx import VoiceCloner
+out = VoiceCloner(engine="knnvc").clone_voice("source.wav", "reference.wav")
+```
+
+Provenance and upstream license: see `PROVENANCE.md`. Weights derive from
+[bshall/knn-vc](https://github.com/bshall/knn-vc) (MIT) and
+[microsoft WavLM](https://github.com/microsoft/unilm/tree/master/wavlm) (MIT).

config.json

@@ -0,0 +1,20 @@
+{
+  "engine": "knn-vc",
+  "components": {
+    "wavlm_encoder": "wavlm_layer6.onnx",
+    "wavlm_encoder_q8": "wavlm_layer6_q8.onnx",
+    "hifigan_vocoder": "hifigan_knnvc.onnx",
+    "hifigan_vocoder_q8": "hifigan_knnvc_q8.onnx"
+  },
+  "sample_rates": {
+    "input": 16000,
+    "output": 16000
+  },
+  "metadata": {
+    "opset": 14,
+    "wavlm_layer": 6,
+    "knn_default_k": 4,
+    "wavlm_source": "microsoft/wavlm-large",
+    "hifigan_source": "https://github.com/bshall/knn-vc/releases/download/v0.1/prematch_g_02500000.pt"
+  }
+}

hifigan_knnvc.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47ac3ce468e96476b432ad40977a9e90ff9a25bdefe7f2352fdba883997c6671
+size 66133611

hifigan_knnvc_q8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93fe768fd462739ff3bc956245664a3bb038eb450317284abab4af0795ce7e43
+size 26276028

hifigan_parity_report.json

@@ -0,0 +1,20 @@
+{
+  "overall_passed": true,
+  "tolerances": {
+    "max_abs": 0.001,
+    "mean_abs": 0.0001
+  },
+  "components": [
+    {
+      "name": "waveform",
+      "max_abs_delta": 2.4437904357910156e-06,
+      "mean_abs_delta": 2.0770487196841714e-07,
+      "shape": [
+        1,
+        1,
+        25600
+      ],
+      "passed": true
+    }
+  ]
+}

wavlm_layer6.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c0cc7c6648c10b11a0f57f2a5c2fa00b64a37e733860e8f734e8775afc98743
+size 405586979

wavlm_layer6_q8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b16c392b7402ca3065970a580bbda9ef5e29221c0ddc1267720e1f1e20a55a5f
+size 102208980

wavlm_parity_report.json

@@ -0,0 +1,20 @@
+{
+  "overall_passed": true,
+  "tolerances": {
+    "max_abs": 0.001,
+    "mean_abs": 0.0001
+  },
+  "components": [
+    {
+      "name": "hidden_states",
+      "max_abs_delta": 0.00051116943359375,
+      "mean_abs_delta": 1.3095690519548953e-05,
+      "shape": [
+        1,
+        49,
+        1024
+      ],
+      "passed": true
+    }
+  ]
+}