Azure.AI.VoiceLive (.NET)

Azure.AI.VoiceLive (.NET)

Real-time voice AI SDK for building bidirectional voice assistants with Azure AI.

Installation

dotnet add package Azure.AI.VoiceLive
dotnet add package Azure.Identity
dotnet add package NAudio                    # For audio capture/​playback

Current Versions: Stable v1.0.0, Preview v1.1.0-beta.1

Environment Variables

AZURE_VOICELIVE_ENDPOINT=https://<resource>.services.ai.azure.com/
AZURE_VOICELIVE_MODEL=gpt-4o-realtime-preview
AZURE_VOICELIVE_VOICE=en-US-AvaNeural
# Optional: API key if not using Entra ID
AZURE_VOICELIVE_API_KEY=<your-api-key>

Authentication

Microsoft Entra ID (Recommended)

using Azure.Identity;
using Azure.AI.VoiceLive;

Uri endpoint = new Uri("https://your-resource.cognitiveservices.azure.com");
DefaultAzureCredential credential = new DefaultAzureCredential();
VoiceLiveClient client = new VoiceLiveClient(endpoint, credential);

Required Role: Cognitive Services User (assign in Azure Portal → Access control)

API Key

Uri endpoint = new Uri("https://your-resource.cognitiveservices.azure.com");
AzureKeyCredential credential = new AzureKeyCredential("your-api-key");
VoiceLiveClient client = new VoiceLiveClient(endpoint, credential);

Client Hierarchy

VoiceLiveClient
└── VoiceLiveSession (WebSocket connection)
    ├── ConfigureSessionAsync()
    ├── GetUpdatesAsync() → SessionUpdate events
    ├── AddItemAsync() → UserMessageItem, FunctionCallOutputItem
    ├── SendAudioAsync()
    └── StartResponseAsync()

Core Workflow

1. Start Session and Configure

using Azure.Identity;
using Azure.AI.VoiceLive;

var endpoint = new Uri(Environment.GetEnvironmentVariable("AZURE_VOICELIVE_ENDPOINT"));
var client = new VoiceLiveClient(endpoint, new DefaultAzureCredential());

var model = "gpt-4o-mini-realtime-preview";

// Start session
using VoiceLiveSession session = await client.StartSessionAsync(model);

// Configure session
VoiceLiveSessionOptions sessionOptions = new()
{
    Model = model,
    Instructions = "You are a helpful AI assistant. Respond naturally.",
    Voice = new AzureStandardVoice("en-US-AvaNeural"),
    TurnDetection = new AzureSemanticVadTurnDetection()
    {
        Threshold = 0.5f,
        PrefixPadding = TimeSpan.FromMilliseconds(300),
        SilenceDuration = TimeSpan.FromMilliseconds(500)
    },
    InputAudioFormat = InputAudioFormat.Pcm16,
    OutputAudioFormat = OutputAudioFormat.Pcm16
};

// Set modalities (both text and audio for voice assistants)
sessionOptions.Modalities.Clear();
sessionOptions.Modalities.Add(InteractionModality.Text);
sessionOptions.Modalities.Add(InteractionModality.Audio);

await session.ConfigureSessionAsync(sessionOptions);

2. Process Events

await foreach (SessionUpdate serverEvent in session.GetUpdatesAsync())
{
    switch (serverEvent)
    {
        case SessionUpdateResponseAudioDelta audioDelta:
            byte[] audioData = audioDelta.Delta.ToArray();
            // Play audio via NAudio or other audio library
            break;
            
        case SessionUpdateResponseTextDelta textDelta:
            Console.Write(textDelta.Delta);
            break;
            
        case SessionUpdateResponseFunctionCallArgumentsDone functionCall:
            // Handle function call (see Function Calling section)
            break;
            
        case SessionUpdateError error:
            Console.WriteLine($"Error: {error.Error.Message}");
            break;
            
        case SessionUpdateResponseDone:
            Console.WriteLine("\n--- Response complete ---");
            break;
    }
}

3. Send User Message

await session.AddItemAsync(new UserMessageItem("Hello, can you help me?"));
await session.StartResponseAsync();

4. Function Calling

// Define function
var weatherFunction = new VoiceLiveFunctionDefinition("get_current_weather")
{
    Description = "Get the current weather for a given location",
    Parameters = BinaryData.FromString("""
        {
            "type": "object",
            "properties": {
                "location": {
                    "type": "string",
                    "description": "The city and state or country"
                }
            },
            "required": ["location"]
        }
        """)
};

// Add to session options
sessionOptions.Tools.Add(weatherFunction);

// Handle function call in event loop
if (serverEvent is SessionUpdateResponseFunctionCallArgumentsDone functionCall)
{
    if (functionCall.Name == "get_current_weather")
    {
        var parameters = JsonSerializer.Deserialize<Dictionary<string, string>>(functionCall.Arguments);
        string location = parameters?["location"] ?? "";
        
        // Call external service
        string weatherInfo = $"The weather in {location} is sunny, 75°F.";
        
        // Send response
        await session.AddItemAsync(new FunctionCallOutputItem(functionCall.CallId, weatherInfo));
        await session.StartResponseAsync();
    }
}

Voice Options

Supported Models

Key Types Reference

Best Practices

1. Always set both modalities — Include Text and Audio for voice assistants
2. Use AzureSemanticVadTurnDetection — Provides natural conversation flow
3. Configure appropriate silence duration — 500ms typical to avoid premature cutoffs
4. Use using statement — Ensures proper session disposal
5. Handle all event types — Check for errors, audio, text, and function calls
6. Use DefaultAzureCredential — Never hardcode API keys

Error Handling

if (serverEvent is SessionUpdateError error)
{
    if (error.Error.Message.Contains("Cancellation failed: no active response"))
    {
        // Benign error, can ignore
    }
    else
    {
        Console.WriteLine($"Error: {error.Error.Message}");
    }
}

Audio Configuration

• Input Format: InputAudioFormat.Pcm16 (16-bit PCM)
• Output Format: OutputAudioFormat.Pcm16
• Sample Rate: 24kHz recommended
• Channels: Mono

Related SDKs

Reference Links

When to Use

This skill is applicable to execute the workflow or actions described in the overview.

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.