Is this page helpful?

Getting Started with the Nemotron ASR Plugin Pack#

Go from the extracted pack to a transcription, then choose the matching integration path. Run sample commands from bin/x64.

1. Prerequisites#

Requirement

Details

Platform

Windows x64

Audio

16 kHz, 16-bit, mono PCM WAV

All backends

Microsoft Visual C++ x64 runtime

CUDA

Compatible NVIDIA driver

Vulkan

Vulkan loader and driver

D3D12

Windows Direct3D runtime; preserve bin/x64/D3D12

Source builds only

Visual Studio 2019/2022 C++ workload and Packman network access

See Shipping Checklist for exact runtime files.

2. Run Your First ASR Sample#

The pack includes a test WAV at data/nvigi.test/nvigi.asr/jfk.wav.

.\nvigi.asr.sample.exe -m ..\..\data\nvigi.models --model parakeet-v2 -b cpu -w ..\..\data\nvigi.test\nvigi.asr\jfk.wav

This processes the complete WAV and returns final text. Use parakeet-v3 for the multilingual offline model.

.\nvigi.asr.sample.exe -m ..\..\data\nvigi.models --model streaming-en -b cpu -w ..\..\data\nvigi.test\nvigi.asr\jfk.wav --right-context 13

Known streaming aliases automatically use Start/Data/Stop. Partial text can arrive before the stream is finalized.

.\nvigi.asr.sample.exe -m ..\..\data\nvigi.models --model streaming-3.5 -b cpu -w ..\..\data\nvigi.test\nvigi.asr\jfk.wav --language en-US

Nemotron 3.5 supports language prompting and word/phrase boosting.

.\nvigi.asr.sample.exe -m ..\..\data\nvigi.models --model streaming-en -b cpu -w ..\..\data\nvigi.test\nvigi.asr\jfk.wav --vad

VAD reports speech state alongside ASR. It does not gate audio or end the stream.

.\nvigi.3d.exe

Select a backend and model from the UI, then record from the microphone or run a WAV.

Change -b cpu to cuda, vk, or d3d12 to select another backend.

Do not pass --streaming, language, or word-boost controls to Parakeet. The shipped command-line sample does not reject every invalid offline combination and may silently ignore controls or return a truncated transcript.

3. Integrate into Your Application#

3.1 Choose a workflow#

Need

Model family

Execution

Complete file or buffer

Parakeet v2/v3

evaluate() or evaluateAsync()

Live microphone or chunked audio

Nemotron streaming

Asynchronous Start/Data/Stop

English streaming

Streaming EN

Right context 0, 1, 6, or 13

Prompt-conditioned multilingual streaming

Streaming 3.5

Right context 0, 3, 6, or 13

Speech/silence metadata

Streaming model plus Silero

VAD capability and runtime chains

3.2 Minimum setup#

#include <nvigi_asr_nemotron.h>

const nvigi::PluginID pluginId =
    nvigi::plugin::asr::nemotron_ggml::cpu::kId;

nvigi::CommonCreationParameters common{};
common.utf8PathToModels = R"(C:\pack\data\nvigi.models)";
common.modelGUID = "{163BD6DC-248D-4238-BC9B-B76D2AB1C3DD}";
common.vramBudgetMB = 4096;

nvigi::ASRNemotronGGMLCreationParameters create{};
create.chain(common);
create.backend = nvigi::ASRGGMLBackend::eCPU;

Then load the selected interface, query capabilities, create an instance, submit audio, destroy the instance, unload the same interface, and call nvigiShutdown. Check every nvigi::Result.

3.3 Next steps#

Task

Read

Complete API lifecycle

Programming Guide

CLI, 3D UI, and source builds

Samples

Speech-state reporting

Silero VAD

Convert a supported NeMo ASR model to run with this plugin

NeMo ASR Model Conversion

Files required for deployment

Developer Pack Overview