Build an AI Voice Agent with Azure Communication Services Call Automation (2026)

TL;DR — As of January 2026, ACS Call Automation bidirectional streaming is GA. You purchase a phone number through ACS, hook EventGrid to your webhook, accept the call with AnswerCall, start MediaStreamingOptions with bidirectional=true, and pipe frames into Voice Live API. The whole loop fits in a small ASP.NET service.

What you'll build

A C# ASP.NET service that answers ACS-routed calls, opens a bidirectional media stream, and bridges audio to Voice Live API. Mid-call the agent can call a tool to fetch order status, and a "transfer to human" intent triggers AddParticipant with a queue's phone number.

Prerequisites

1. Azure subscription with ACS resource and a purchased PSTN number.
2. EventGrid topic subscribed to ACS events.
3. Voice Live API enabled on a Foundry resource (same region recommended).
4. .NET 8, Azure.Communication.CallAutomation NuGet, Azure.Identity.
5. Public HTTPS callback URL (use ngrok or Container Apps for dev).

Architecture

flowchart TD
  PSTN[Caller PSTN] --> ACS[ACS Call Automation]
  ACS -->|EventGrid IncomingCall| API[ASP.NET Webhook]
  API -->|AnswerCall + MediaStreaming| ACS
  ACS <-->|wss audio frames| API
  API <-->|wss| VL[Voice Live API]
  VL --> GPT[gpt-realtime-mini]
  API -->|AddParticipant| QUEUE[Human Queue Number]

Step 1 — Wire the EventGrid IncomingCall handler

```csharp [HttpPost("/incoming")] public async Task Incoming([FromBody] EventGridEvent[] events) { foreach (var e in events) { if (e.EventType == "Microsoft.EventGrid.SubscriptionValidationEvent") { var data = e.Data.ToObjectFromJson(); return Ok(new { validationResponse = data.ValidationCode }); } if (e.EventType == "Microsoft.Communication.IncomingCall") { var d = e.Data.ToObjectFromJson(); await _calls.AnswerCallAsync(new AnswerCallOptions(d.IncomingCallContext, new Uri($"{Host}/callback")) { MediaStreamingOptions = new MediaStreamingOptions( new Uri($"wss://{HostNoScheme}/media"), MediaStreamingContent.Audio, MediaStreamingAudioChannel.Mixed, startMediaStreaming: true) { EnableBidirectional = true, AudioFormat = AudioFormat.Pcm24KMono } }); } } return Ok(); } ```

Step 2 — Accept the WebSocket and bridge to Voice Live

```csharp [HttpGet("/media")] public async Task Media() { if (!HttpContext.WebSockets.IsWebSocketRequest) { HttpContext.Response.StatusCode = 400; return; } var acs = await HttpContext.WebSockets.AcceptWebSocketAsync(); using var vl = new ClientWebSocket(); vl.Options.SetRequestHeader("Authorization", $"Bearer {await GetAadToken()}"); await vl.ConnectAsync(new Uri("wss://vox-foundry.cognitiveservices.azure.com/voice-agent/realtime?api-version=2025-05-01-preview&model=gpt-realtime"), default);

await SendSessionUpdate(vl);
var t1 = Pump(acs, vl, ParseAcsFrame);   // ACS -> Voice Live
var t2 = Pump(vl, acs, FormatAcsFrame);  // Voice Live -> ACS
await Task.WhenAny(t1, t2);

} ```

ACS bidirectional frames are JSON-wrapped base64 PCM at 24kHz mono — the same sample rate Voice Live wants natively. No resampling.

Step 3 — Frame format helpers

```csharp record AcsAudioFrame(string kind, AcsAudioData audioData); record AcsAudioData(string data, string timestamp, string participantRawID, bool silent);

byte[] ParseAcsFrame(byte[] raw) { var f = JsonSerializer.Deserialize(raw); return f.kind == "AudioData" && !f.audioData.silent ? Convert.FromBase64String(f.audioData.data) : Array.Empty(); } byte[] FormatAcsFrame(byte[] pcm) { var f = new { kind = "AudioData", audioData = new { data = Convert.ToBase64String(pcm), timestamp = DateTime.UtcNow.ToString("o") } }; return JsonSerializer.SerializeToUtf8Bytes(f); } ```

Step 4 — Tool: lookup_order via function calling

In session.update, register the tool. When Voice Live emits response.function_call_arguments.done, dispatch to your CRM SDK and reply with conversation.item.create (function_call_output) + response.create. Same pattern as OpenAI Realtime.

Step 5 — Transfer to human

When the user says "agent please", parse the model's signal (a tool call request_transfer is the cleanest), then:

```csharp await _calls.GetCallConnection(callConnectionId).AddParticipantAsync(new CallInvite(new PhoneNumberIdentifier("+1800SUPPORT"), new PhoneNumberIdentifier("+1YourACSNumber"))); ```

ACS handles SIP REFER under the covers; the AI can stay in the call as a transcriber or drop with HangUpAsync.

Step 6 — Recording + Contact Lens-equivalent

Enable StartRecordingAsync for compliance. ACS recordings drop into your storage account; pipe through Azure AI Speech batch transcription + Foundry sentiment analysis for post-call analytics.

Step 7 — Ship to Container Apps with managed identity

Same as the previous post: az containerapp create with --user-assigned for AAD, --ingress external, scaling to 50 replicas. WebSocket sticky session is automatic in Container Apps.

Pitfalls

• EventGrid validation handshake: must respond to SubscriptionValidationEvent with the validation code or the topic stays unverified.
• MediaStreamingOptions requires EnableBidirectional=true for the new GA bidirectional path; otherwise you get one-way out-of-call streaming (legacy).
• Audio format: ACS bidirectional only supports 24kHz mono PCM. If your downstream is 8kHz mu-law, transcode on the fly.
• Concurrent call quota starts at 100 — request increases via support.
• PSTN ingress costs ~$0.014/min in addition to Voice Live; cheaper than Twilio in most regions but watch billing.

How CallSphere does this in production

Explore a live demo and compare current plans to find the right fit for your business.

FAQ

Q: Can I use ACS without Voice Live? Yes — bridge to any STT+LLM+TTS stack. Voice Live just removes the integration tax.

Q: How do I get an EU number? ACS supports number purchase in 30+ countries via the portal; pick country during PurchasePhoneNumbers.

Q: Latency vs Twilio Media Streams? On East US 2 with Voice Live: ~750ms. With Twilio + OpenAI Realtime: ~650ms. ACS catches up in EU regions where Twilio adds a transatlantic hop.

Q: How do I do warm transfers? Use AddParticipantAsync then MuteParticipantAsync for the AI; the live agent picks up the same call leg.

Q: Can I record with redaction? Yes — pipe recordings through Azure AI Speech batch transcription with profanity=Masked + a Foundry redaction prompt for PII/PHI before storing.

Sources

• Call Automation AI Sample — ACS
• AI in Azure Communication Services overview
• Create next-gen voice agents with Voice Live and ACS — Microsoft Community Hub
• Azure-Samples/call-center-voice-agent-accelerator
• Azure-Samples/communication-services-AI-customer-service-sample