.NET Library for OpenAI WebRTC End Point

This repository contains a .NET library for interacting with OpenAI's real-time WebRTC API. It provides helper classes to negotiate peer connections, send and receive OPUS audio frames and exchange control messages over a data channel.

Features

• Establish a RTCPeerConnection (WebRTC) with OpenAI using a REST based signalling helper.
• Send audio samples or pipe them from existing SIPSorcery media end points.
• Receive transcript and other events via the data channel.
• DataChannelMessenger class to assist with sending session updates, function call results and response prompts.
• Designed to work with dependency injection (ASP.NET) or standalone alone applications (Console & WinForms).

Installation

Install the library from NuGet:

dotnet add package SIPSorcery.OpenAI.WebRTC

Usage

Console/WinForms Direct WebRTC Connection to OpenAI Realtime End Point

See GetStarted example for full source.

using SIPSorcery.OpenAIWebRTC;
using SIPSorcery.OpenAIWebRTC.Models;

// Create the new WebRTC end point provided by this library (nothing starts yet).
var webrtcEndPoint = new WebRTCEndPoint(openAiKey, logger);

// Initialise default Windows audio devices and wire up event handlers to the WebRTC end point.
var windowsAudioEp = InitialiseWindowsAudioEndPoint();
webrtcEndPoint.ConnectAudioEndPoint(windowsAudioEp);

// Tell the WebRTC end point to start the connection attempt to the OpenAI Realtime WebRTC end point.
var negotiateConnectResult = await webrtcEndPoint.StartConnect();

// Wait for the connection to establish and then optionally update the session, start a conversation 
// and process data channel messages.

webrtcEndPoint.OnPeerConnectionConnected += () =>
{
    Log.Logger.Information("WebRTC peer connection established.");

    var voice = RealtimeVoicesEnum.verse;

    // Optionally send a session update message to adjust the session parameters.
    var sessionUpdateResult = webrtcEndPoint.DataChannelMessenger.SendSessionUpdate(
        voice,
        "Keep it short.",
        transcriptionModel: TranscriptionModelEnum.Whisper1);

    if (sessionUpdateResult.IsLeft)
    {
        Log.Logger.Error($"Failed to send rsession update message: {sessionUpdateResult.LeftAsEnumerable().First()}");
    }

    // Trigger the conversation by sending a response create message.
    var result = webrtcEndPoint.DataChannelMessenger.SendResponseCreate(voice, "Say Hi!");
    if (result.IsLeft)
    {
        Log.Logger.Error($"Failed to send response create message: {result.LeftAsEnumerable().First()}");
    }
};

webrtcEndPoint.OnDataChannelMessage += (dc, message) =>
{
    var log = message switch
    {
        RealtimeServerEventSessionUpdated sessionUpdated => $"Session updated: {sessionUpdated.ToJson()}",
        RealtimeServerEventConversationItemInputAudioTranscriptionDelta inputDelta => $"ME ⌛: {inputDelta.Delta?.Trim()}",
        RealtimeServerEventConversationItemInputAudioTranscriptionCompleted inputTranscript => $"ME ✅: {inputTranscript.Transcript?.Trim()}",
        RealtimeServerEventResponseAudioTranscriptDelta responseDelta => $"AI ⌛: {responseDelta.Delta?.Trim()}",
        RealtimeServerEventResponseAudioTranscriptDone responseTranscript => $"AI ✅: {responseTranscript.Transcript?.Trim()}",
        _ => $"Received {message.Type} -> {message.GetType().Name}"
    };

    if (log != string.Empty)
    {
        Log.Information(log);
    }
};

Example Output:

[20:45:29 INF] AI ✅: Hello! How can I assist you today?
[20:45:40 INF] ME ✅: Tell me a nursery rhyme and use as many emojis as you can in the transcription.
[20:45:44 INF] AI ✅: 🍼🎶 Humpty Dumpty sat on a wall, 🥚⬆️🌉 Humpty Dumpty had a great fall. 🥚💥⤵️ All the king's horses 🐎👑 and all the king's men 👨‍✈️👑 couldn't put Humpty together again! 🥚❌⚒️🐣
[20:46:06 INF] AI ✅: You're welcome! 😊 Anytime!
[20:46:06 INF] ME ✅: Thank you.

ASP.NET WebRTC Bridge: Browser <- ASP.NET Bridge -> OpenAI Realtime End Point

See BrowserBridge example for full source.

using SIPSorcery.OpenAIWebRTC;
using SIPSorcery.OpenAIWebRTC.Models;

// Set up an ASP.NET web socket to listen for connections.
// The web socket is NOT used for the connection to OpenAI. It's a convenience signalling channel to allow the browser
// to establish a WebRTC connection with the ASP.NET app.

app.Map("/ws", async (HttpContext context,
    [FromServices] IWebRTCEndPoint openAiWebRTCEndPoint) =>
{
    Log.Debug("Web socket client connection established.");

    if (context.WebSockets.IsWebSocketRequest)
    {
        var webSocket = await context.WebSockets.AcceptWebSocketAsync();

        // Create the ASP.NET WebRTC peer to connect to the browser.

        var webSocketPeer = new WebRTCWebSocketPeerAspNet(
            webSocket,
            CreateBrowserPeerConnection,
            null,
            RTCSdpType.offer);

        // Start the WebRTC connection attempt to the browser.

        var browserPeerTask = webSocketPeer.Run();

        // Start the attempt to connect to the OpenAI WebRTC end point in parallel with the 
        // browser connection which is already underway.

        SetOpenAIPeerEventHandlers(openAiWebRTCEndPoint);
        var openAiPeerTask = openAiWebRTCEndPoint.StartConnect(config);

        // Wait for both WebRTC connections to establish.

        await Task.WhenAll(browserPeerTask, openAiPeerTask);

        // Wire up the event handlers to connect the browser's audio to the openAIWebRTCEndPoint instance. 
        // This is the equivalent of connecting the audio to local Windows audio devices but in this case
        // the Browser audio stream is being wired up to the OpenAI audio stream.
        // It's much simpler to connect the browser directly to OpenAI, and this library is not needed for that.
        // The advantage of having an ASP.NET app in the middle is for things like capturing the audio transcription
        // or using local functions that can be handled by the ASP.NET app.

        ConnectPeers(webSocketPeer.RTCPeerConnection, openAiWebRTCEndPoint);

        Log.Debug("Web socket closing with WebRTC peer connection in state {state}.", webSocketPeer.RTCPeerConnection?.connectionState);
    }
    else
    {
        // Not a WebSocket request
        context.Response.StatusCode = StatusCodes.Status400BadRequest;
    }
});

Examples

Several sample applications demonstrating different scenarios are available in the examples folder:

• GetStarted – minimal console program that connects your microphone to OpenAI.
• AliceAndBob – runs two OpenAI peers and routes their audio between each other with a waveform display.
• LocalFunctions – showcases the local function calling feature.
• GetPaid – extends local functions to simulate payment requests.
• DependencyInjection – illustrates using the library with .NET DI.
• ASP.NET Get Started – ASP.NET application bridging a browser WebRTC client to OpenAI.
• ASP.NET Local Function – ASP.NET application that builds on the Get Started example and adds a local function to tailor OpenAI responses.

Each example folder contains its own README with usage instructions.

License

Distributed under the BSD 3‑Clause license with an additional BDS BY‑NC‑SA restriction. See LICENSE.md for details.