A step-by-step guide to building a live speech-to-text console app in C# using Microsoft's Foundry Local and the NVIDIA Nemotron 0.6B streaming ASR model. Foundry Local handles the full model lifecycle — catalog lookup, download, caching, loading, and unloading — without requiring API keys or manual file management. The app captures microphone audio via NAudio at 16 kHz mono PCM, streams it to the model through an async channel to handle backpressure, and reads interim and final transcription results as an async stream. Key advantages of running locally include on-device audio privacy, offline capability after initial download, and no cloud dependency. The post also explains when to use Microsoft.Extensions.AI abstractions versus provider-specific SDKs like the native AudioClient.
