---
title: Nemotron-3-Diarization
canonical_url: "https://www.modelscope.cn/models/mlx-community/Nemotron-3-Diarization"
md_url: "https://www.modelscope.cn/models/mlx-community/Nemotron-3-Diarization.md"
repository: mlx-community/Nemotron-3-Diarization
last_updated: 2026-09-23
license: openmdw-1.1
pipeline_tag: voice-activity-detection
tasks:
  - voice-activity-detection
model_type:
  - nemotron_diarization
base_model:
  - nvidia/Nemotron-3-Diarization
base_model_relation: finetune
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 68
stars: 0
tags:
  - mlx
  - mlx-audio
  - speaker-diarization
  - streaming-sortformer
---

# Nemotron-3-Diarization

> Nemotron-3-Diarization - mlx-community 在 ModelScope 开源的模型。Nemotron 3 Diarization (MLX)

mlx-community/Nemotron-3-Diarization 是 ModelScope 魔搭社区上的voice-activity-detection模型，采用 openmdw-1.1 许可，基于 nvidia/Nemotron-3-Diarization 构建。

- **Repository**: mlx-community/Nemotron-3-Diarization
- **License**: openmdw-1.1
- **Tasks**: voice-activity-detection
- **Base model**: nvidia/Nemotron-3-Diarization
- **Tags**: mlx, mlx-audio, speaker-diarization, streaming-sortformer
- **Downloads**: 68
- **Stars**: 0
- **Last updated**: 2026-09-23

Source: https://www.modelscope.cn/models/mlx-community/Nemotron-3-Diarization

---

# Nemotron 3 Diarization (MLX)

Converted from [nvidia/Nemotron-3-Diarization](https://huggingface.co/nvidia/Nemotron-3-Diarization) for use with [mlx-audio](https://github.com/Blaizzy/mlx-audio).

It predicts up to eight speakers at 10 ms resolution from 16 kHz mono audio.

```python
from mlx_audio.vad import load

model = load("mlx-community/Nemotron-3-Diarization", strict=True)
result = model.generate("meeting.wav")
print(result.text)

# Incremental results retain speaker identities through the AOSC and FIFO.
for result in model.generate_stream("meeting.wav"):
    for segment in result.segments:
        print(segment.start, segment.end, segment.speaker)
```

For live PCM input, call `model.feed(chunk, state)` with a state from
`model.init_streaming_state()` and 16 kHz mono chunks. Flush the final partial
chunk and lookahead with `model.feed([], state, final=True)`. Timestamps are
absolute within the recording. Labels are generic arrival-order speaker IDs;
the model does not identify people. Overlapping speakers may be active together.

Consult the [upstream model card](https://huggingface.co/nvidia/Nemotron-3-Diarization)
for training data, evaluation and license information.