aufklarer's picture
ONNX export, gated against PyTorch on real speech
7247ec8 verified
|
Raw
History Blame Contribute Delete
1.08 kB
metadata
license: mit
library_name: onnx
tags:
  - speaker-diarization
  - segmentation
  - onnx
base_model: pyannote/segmentation-3.0

Pyannote Segmentation 3.0 — ONNX

Speaker segmentation for the server-side diarization pipeline (ONNX Runtime, CUDA EP with CPU fallback). Exported by models/pyannote-vad/export/convert_onnx.py in soniqo/speech-models.

input audio [1, 1, 160000] f32 — one 10 s window @ 16 kHz
output posteriors [1, 589, 7] f32 — per-frame powerset probabilities

Frame geometry differs from the LiteRT bundle. This graph runs the LSTM internally over the whole 10 s window and emits 589 frames (17.0 ms/frame). The LiteRT bundle decomposes the same model into 1 s chunks with explicit LSTM state (56 frames/chunk, 560 per window). A decoder ported from the LiteRT path must re-derive its frame-to-time mapping — reusing the old one shifts every segment.

Verified against PyTorch on real speech: powerset class agreement 1.0000, max abs diff ~1e-4 (benchmarks/parity_diarization_onnx.py).