vibe-app/diar-streaming-sortformer-4spk-v2-gguf overview
Streaming Sortformer 4 speaker v2 — GGUF A GGUF build of nvidia/diar streaming sortformer 4spk v2 https://huggingface.co/nvidia/diar streaming sortformer 4spk …
Runs locally from ~140.3 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| diar_streaming_sortformer_4spk-v2.q8_0.gguf | GGUF | GGUF | 140.3 MB | Download |
Model Details
Model README
---
license: cc-by-4.0
base_model: nvidia/diar_streaming_sortformer_4spk-v2
tags:
- speaker-diarization
- sortformer
- gguf
- vibe
---
Streaming Sortformer 4-speaker v2 — GGUF
A GGUF build of nvidia/diar_streaming_sortformer_4spk-v2,
mirrored here so Vibe has a download source it controls.
Converted from the original .nemo checkpoint with NVIDIA's own converter
(NeMo-Speech.cpp, Apache-2.0), not from an ONNX export.
The result matches NVIDIA's published GGUF tensor-for-tensor: 971 tensors, same names, same shapes.
Files
| file | size | notes |
|---|---|---|
| diar_streaming_sortformer_4spk-v2.q8_0.gguf | 147 MB | Q8_0. What Vibe downloads. |
Only F32/F16/Q8_0 are safe for this model. The k-quant tiers are not: the speaker-cache
compression makes discrete near-tie decisions, and k-quant weight error can flip one and
permute speaker labels mid-stream.
Validation
Checked against the ONNX Runtime output it replaces, on two clips (12 s / 41 s):
identical speaker labels once binarised at 0.5, and every segment boundary identical
to the millisecond. Max absolute probability difference 0.018.
Caps at 4 speakers. 80 ms frames, 16 kHz mono input.
License
cc-by-4.0, inherited from the base model. Credit NVIDIA.
Run vibe-app/diar-streaming-sortformer-4spk-v2-gguf with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models