Live Speech-to-Text in C# with Foundry Local: A 0.67 GB Model and No Cloud Call

Bruno Capuano published Beyond Chat: live Speech-to-Text with Foundry Local and C# on the .NET blog on August 4, 2026. It is a console app that reads your microphone and prints transcription as you speak, with no network call and no per-token bill. The interesting part is not the demo, it is that streaming ASR is now a plain async stream in the Foundry Local SDK.

The session is the whole API

Foundry Local exposes live transcription through an audio client hanging off a loaded model. There is no IChatClient here: this is the native SDK surface, not Microsoft.Extensions.AI.

var config = new Configuration { AppName = "transcriber" };
await FoundryLocalManager.CreateAsync(config, logger);
var mgr = FoundryLocalManager.Instance;

await mgr.DownloadAndRegisterEpsAsync();

var catalog = await mgr.GetCatalogAsync();
var model = await catalog.GetModelAsync("nemotron-speech-streaming-en-0.6b")
            ?? throw new InvalidOperationException("Model not in catalog");

await model.DownloadAsync(p => Console.Write($"\r{p:F2}%"));
await model.LoadAsync();

var audioClient = await model.GetAudioClientAsync();
var session = audioClient.CreateLiveTranscriptionSession();
session.Settings.SampleRate = 16000;
session.Settings.Channels = 1;
session.Settings.Language = "en";

await session.StartAsync();

You then push raw 16-bit mono PCM at 16 kHz with session.AppendAsync(chunk) and read results off session.GetStream():

await foreach (var result in session.GetStream())
{
    var text = result.Content?[0]?.Text;
    if (result.IsFinal)
        Console.WriteLine($"\n[FINAL] {text}");
    else if (!string.IsNullOrEmpty(text))
        Console.Write(text);
}

IsFinal is a transcript-state marker, not a stream terminator. Treat it as “this span will not be revised again” and keep reading.

The trap: NAudio’s callback is synchronous

Microphone capture uses NAudio.WaveInEvent, and its DataAvailable event handler is synchronous while AppendAsync is not. Firing AppendAsync from the handler gives you unbounded fire-and-forget calls that ignore SDK backpressure. The sample routes audio through a bounded channel instead:

var audioChannel = Channel.CreateBounded<byte[]>(
    new BoundedChannelOptions(50) { FullMode = BoundedChannelFullMode.DropOldest });

var appendTask = Task.Run(async () =>
{
    await foreach (var chunk in audioChannel.Reader.ReadAllAsync())
        await session.AppendAsync(chunk);
});

waveIn.DataAvailable += (s, e) =>
{
    var buffer = new byte[e.BytesRecorded];
    Buffer.BlockCopy(e.Buffer, 0, buffer, 0, e.BytesRecorded);
    audioChannel.Writer.TryWrite(buffer);
};

Dropping the oldest 100 ms chunk under pressure is the right call for live captioning. Blocking the audio callback is not.

Which package, which model

On Windows, install Microsoft.AI.Foundry.Local.WinML, which routes through the Windows ML runtime for broader hardware acceleration. Everywhere else, Microsoft.AI.Foundry.Local. Same API surface either way. Add NAudio for capture, and note that WaveInEvent is Windows-only, so the sample falls back to synthetic PCM elsewhere. The docs list .NET 9.0 SDK or later as the floor; the blog sample targets .NET 10.

The English model nemotron-speech-streaming-en-0.6b was quantized from 2.47 GB down to 0.67 GB and runs faster than real time on CPU. If you need more than English, swap the alias for nvidia-nemotron-3.5-asr-streaming-multilingual-0.6b and set session.Settings.Language = "auto".

Finish with session.StopAsync() and model.UnloadAsync(), and use model.RemoveFromCacheAsync() when you want the disk back. Full walkthrough in the Microsoft Learn how-to.