WeSpeaker ResNet34-LM — ONNX
Speaker embeddings for the server-side diarization pipeline (ONNX Runtime, CUDA EP with CPU fallback). Exported by models/wespeaker/export/convert_onnx.py in soniqo/speech-models.
| input | fbank [1, T, 80] f32 — precomputed kaldi log-mel, T dynamic |
| output | embedding [1, 256] f32 |
Takes precomputed fbank, not raw audio. pyannote compute_fbank uses torch.fft.rfft, and aten::fft_rfft has no ONNX lowering — an in-graph frontend cannot be exported at opset 18. The caller computes kaldi fbank (80 mels, 25 ms / 10 ms, dither 0, waveform scaled by 1<<15); per-utterance mean centering is applied inside the graph. Same contract as the LiteRT bundle, so swapping runtimes needs no feature-code change.
Verified against PyTorch on real speech: cosine similarity 1.000000 (benchmarks/parity_diarization_onnx.py).
- Downloads last month
- 26
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for soniqo/WeSpeaker-ResNet34-LM-ONNX
Base model
pyannote/wespeaker-voxceleb-resnet34-LM