Live Speech-to-Text with Foundry Local and C# for .NET 10

Real-time Speech Transcription with Foundry Local

A new guide outlines the creation of a C# console application that performs live speech-to-text transcription. This application leverages Foundry Local to run a specialized NVIDIA 0.6B Nemotron speech model, enabling real-time processing of microphone audio and delivering both interim and final transcription results as a user speaks.

Foundry Local's Role in Model Management

Foundry Local handles the complete lifecycle of the speech recognition model. This includes identifying the correct model variant, downloading it when necessary, caching it locally, loading it for inference, and unloading it once the application concludes. This simplifies the management of AI models for developers.

Integration with .NET and Specialized SDKs

While Microsoft.Extensions.AI provides abstractions for general AI scenarios like chat, live audio streaming requires direct interaction with provider-specific SDKs. The sample uses the native Microsoft.AI.Foundry.Local SDK's AudioClient for Foundry Local-specific capabilities such as live transcription sessions and raw PCM streaming. This approach allows developers to use abstractions when suitable and specialized SDKs for unique requirements.

Application Components and Requirements

The sample application targets .NET 10 and integrates Microsoft.AI.Foundry.Local.WinML for Foundry Local, NAudio for microphone audio capture, and the nemotron-speech-streaming-en-0.6b model from the Foundry Local catalog. It is important to note that this sample is Windows-only, as Microsoft.AI.Foundry.Local.WinML relies on Windows ML and NAudio.WaveInEvent uses Windows audio APIs.