diar-streaming-sortformer-4spk-v2.1-onnx
Overview
This repository contains an ONNX export of NVIDIA's streaming Sortformer diarizer 4-speaker v2.1 model.
Source
| Field | Value |
|---|
| Upstream model | nvidia/diar_streaming_sortformer_4spk-v2.1 |
| Upstream source revision | fafaab5faa1617a0ca52d38dd3dc4bd636800d3d |
| Export tool/script | ONNX export from NVIDIA Sortformer checkpoint (community packaging) |
| Quantization recipe | FP32 ONNX (onnx/model.onnx) |
Precision and Packaging
Export tooling, precision, and quantization are recorded in the Source table above. This packaging mirror does not publish independent parity benchmarks; validate on your target execution provider before production use.
Files
| File | Description |
|---|
onnx/model.onnx | ONNX model file |
Intended Use
Use the ONNX model for speaker diarization experiments or integration work where a streaming four-speaker diarization model is appropriate.
Training Data
The upstream model card lists Fisher English, NIST SRE 2004-2010, LibriSpeech, AMI Meeting Corpus, VoxConverse v0.3, ICSI, AISHELL-4, DIHARD Challenge 3 development data, NIST SRE 2000 Disc8 split1, NOTSOFAR1, AliMeeting-train, and DiPCo.
Evaluation
The upstream model card reports Diarization Error Rate (DER) results across DIHARD III, CALLHOME, AliMeeting, AMI, NOTSOFAR1, and CH109 evaluation sets. This repository does not document ONNX-specific parity results.
Limitations
- The upstream model is designed for streaming diarization and a maximum of four speakers.
- The upstream model card notes degraded performance on recordings with five or more speakers, non-English speech, noisy data, and out-of-domain data.
- This repository does not document the exact ONNX export command or parity validation.
License
Use of the source model is governed by the NVIDIA Open Model License Agreement.
