KurenP/vtuber-live-subs-models

ReDimNet2+ speaker embedding, ONNX export

redimnet2plus.onnx is lab260/redimnet2-plus (ReDimNet2-B6 backbone, trained for channel and codec robustness; Apache-2.0) exported to ONNX for the vtuber-live-subs app, which identifies who is speaking in VTuber streams with it.

• Input wav: float32 [1, samples], 16 kHz mono, any length (one 25 ms frame or more).
• Output emb: [1, 192], L2-normalised. The model's own front end (hamming window, pre-emphasis, 72 log-mel bins, mean over time removed) is inside the graph, so no feature extraction is needed.
• Export: PyTorch 2 → ONNX opset 17 with the author's asv code (return_2d_output: True, out_channels: 224); checked against PyTorch to 3e-7 on random audio. Script: evaluation/speaker-redimnet-2026-10-26/export_onnx.py in the app's repository.
• CPU onnxruntime: 0.08 s per 3 s clip on an Apple M4 (4 threads).

All credit for the model goes to its authors (lab260; ReDimNet2 architecture by PalabraAI, MIT). This repository only holds a format conversion. Weights are redistributed under the original Apache-2.0 licence.