---
language: en
tags:
- onnx
- nemotron
- speech-recognition
- asr
- audio
- nvidia
- cpu
- cuda
- dml
- directml
license: other
---

# Nemotron ASR ONNX Models

ONNX-exported Nemotron-3.5 ASR streaming model (0.6B) for speech recognition.

## Variants

| Variant | Provider | Description |
|---------|----------|-------------|
| `cpu/` | CPUExecutionProvider | CPU-optimized (fp32 decoder/joint, int4/int8 encoder) |
| `gpu-cuda/` | CUDAExecutionProvider | NVIDIA GPU via CUDA (fp32) |
| `gpu-dml/` | DmlExecutionProvider | DirectML for Windows GPU (fp32) |

## Model Architecture

- **Encoder**: Conformer-based audio encoder
- **Decoder**: Transformer-based text decoder
- **Joint**: Transducer joint network
- **VAD**: Silero VAD for voice activity detection

## Usage with NemotronSpeech (C#)

```bash
git clone https://github.com/DimQ1/nemotron-speech-csharp
cd nemotron-speech-csharp
dotnet run -c Release -- models-onnx/gpu-cuda --mic --language auto
```

## Conversion

Converted from NVIDIA NeMo `.nemo` checkpoint using [Olive](https://github.com/microsoft/Olive).
Original model: `nemotron-3.5-asr-streaming-0.6b.nemo`

## License

See original NVIDIA Nemotron license.