| license: mit | |
| tags: | |
| - audio | |
| - voice | |
| - onnx | |
| - wavlm | |
| # pbTrueVoiceLeveler ONNX models | |
| ONNX models for **pbTrueVoiceLeveler** (idempotent voice leveling). | |
| | file | content | | |
| |---|---| | |
| | `wavlm_base_plus.onnx` | WavLM Base+ 12-layer feature extractor. input `audio[B,N]` (16k, RMS-normalized) -> `layers[B,T,12,768]` | | |
| | `tension.onnx` | TensionNet: clip features -> target LUFS/dynrange + bounded 16-band residual | | |
| | `compressor.onnx` | CompressorNet: clip features -> (strength, thr_off) | | |
| | `trim.onnx` | TrimNet: frame features -> keep logit [B,T] | | |
| | `contour.onnx` | LoudnessContourNet: frame features -> relative loudness [B,T] | | |
| | `stats.json` | standardization stats + band centers (pre/post-processing) | | |
| All fp32, opset 17, dynamic batch/time axes. Numerically equivalent to the PyTorch originals | |
| (max abs diff < 1e-4 for nets, < 1e-3 for WavLM). | |
| WavLM Base+ original weights: Microsoft WavLM (MIT). This is an ONNX export for inference use. | |