ACS Call Automation playToAll silent in Teams Online Meeting (CallConnected + PlayCompleted, list_participants=0)

[cunddtech]

Microsoft Azure Support Ticket — Teams Bidirectional Media (Voice + Live Avatar)

Stand: 2026-06-30
Sprache Ticket: Englisch (Copy-Paste unten)
Portal: Azure Portal → Help + support → Create a support request
Schritt-für-Schritt (Deutsch): MICROSOFT_SUPPORT_TICKET_PORTAL_WALKTHROUGH_2026-06-30.md
Issue type: Technical → Azure Communication Services (primary) + cross-reference Microsoft Teams / Graph Cloud Communications

---

Kurz (Deutsch, intern)

Wir bauen STELLA — einen KI-Copilot für Teams-Meetings mit bidirektionaler Kommunikation:

Phase	Ziel
1 — Voice only	STELLA spricht hörbar in der Meeting-Leitung (TTS), hört Teilnehmer (STT), antwortet intelligent
2 — Voice + Video	Live-Avatar (Runway) als sichtbare Kachel im Meeting-Grid, synchron zu Audio

Blocker: ACS Call Automation (playToAll) meldet CallConnected + PlayCompleted, aber list_participants=0 → kein hörbares Audio (Ghost Leg). Das bleibt bestehen auch nach native onlineMeetings-Erstellung + Application Access Policy.

Bitte an Microsoft: Empfohlener Architekturpfad für produktionsreife Bot-Media in Teams (Audio rein/raus, später Video) — ACS vs Application-hosted Media vs andere.

---

Copy-Paste: Support Request (English)

Title

ACS Call Automation playToAll silent in Teams Online Meeting (PlayCompleted, participants=0) — guidance for bidirectional voice + live avatar video

Problem description

We are building STELLA, a multi-tenant AI meeting copilot integrated with Microsoft Teams. Our production stack uses:

• Azure Communication Services (ACS) resource: cdtech-communication (region: Europe)
• Microsoft 365 tenant: fe87c2f4-700c-4a68-9bd7-218fde87a857 (C&D Technologies GmbH)
• Organizer UPN: mj@ilovestella.de (also mj@cdtechnologies.de)
• Backend: Python/FastAPI on Azure Railway (https://api.cdtechnologies.de)
• Teams Graph App (application permissions): Client ID 4b1d32ba-6c39-4a0b-b320-31981ce0654b (GRAPH_CLIENT_ID_TEAMS)

We need guidance on the correct Microsoft-recommended architecture for:

1. Outbound audio (Phase 1): Bot TTS audible to all meeting participants (not operator-only Sidecar/browser audio).
2. Inbound audio (Phase 1): Bot receives participant speech for STT / conversational AI (bidirectional voice).
3. Outbound video (Phase 2): Live AI avatar (H.264) as a visible participant tile in the Teams meeting grid, lip-synced to TTS.
4. Scalable multi-tenant SaaS: Each customer uses their own M365 tenant; we orchestrate via delegated/app permissions and tenant-specific organizer mailboxes.

---

Current architecture (what we built)

┌─────────────────────────────────────────────────────────────────────────────┐
│ STELLA Backend (Railway, Linux, Python)                                      │
│  • Meeting create: Graph POST /users/{org}/onlineMeetings (native)           │
│  • Copilot session, TTS (Azure Speech de-DE-KatjaNeural → WAV)               │
│  • ACS Call Automation: connect_call(groupCallId) + playToAll(audioUrl)    │
│  • Callbacks: POST /api/teams/copilot-acs-media/events                       │
│  • Bot Framework: meeting chat text (works)                                    │
│  • Sidecar UI: browser STT chunks → backend (works, not meeting-wide)        │
└───────────────┬───────────────────────────────┬─────────────────────────────┘
                │ HTTPS                          │ Graph / Bot Framework
                ▼                                ▼
┌───────────────────────────┐    ┌────────────────────────────────────────────┐
│ ACS (cdtech-communication)│    │ Teams Online Meeting                        │
│ Call Automation API       │───▶│ groupCall / thread v2                       │
│ (invisible bot leg)       │    │ Human participants + meeting chat           │
└───────────────────────────┘    └────────────────────────────────────────────┘

Optional (scaffold, Windows VM — not production-ready):
┌─────────────────────────────────────────────────────────────────────────────┐
│ teams-media-bot (.NET 8, Azure Windows VM)                                   │
│  • Target: Microsoft.Graph.Communications.Calls.Media (Application-hosted)   │
│  • StellaBridge HTTP: push PCM + H.264 frames                                │
│  • Today: ACS relay fallback only; GraphCommunicationsService = deferred     │
│  • URL: https://teams-media-bot.cdtechnologies.de (health often timeout)     │
└─────────────────────────────────────────────────────────────────────────────┘

Live avatar source (Phase 2):
┌─────────────────────────────────────────────────────────────────────────────┐
│ Runway Live API → WebRTC/video frames → transcode → Media Bot VideoSocket    │
│ (Sidecar bridge today: operator-only video; not in Teams grid)               │
└─────────────────────────────────────────────────────────────────────────────┘

Separate channels (important):

Channel	Technology	Status
Meeting chat (text, @mentions)	Bot Framework + Graph	✅ Works
Meeting audio broadcast (TTS to all)	ACS playToAll	❌ Ghost Leg — silent
Operator Sidecar TTS	Browser audio	✅ Operator hears only
Transcript post-meeting	Graph onlineMeeting transcript	✅ (after meeting ends)
Live STT from all participants	Not implemented	❌ Need inbound media path
Visible avatar tile	Graph Application-hosted media (planned)	❌ Not wired

---

Symptom (Ghost Leg)

When joining a native Teams online meeting (created via Graph onlineMeetings, not calendar fallback):

• ACS connect_call succeeds; webhook CallConnected received.
• ACS playToAll with publicly reachable HTTPS WAV URL succeeds; webhook PlayCompleted received.
• ACS list_participants returns count=0 consistently after connect.
• No audible media in Teams client for human participants.
• Sidecar/browser TTS works for operator only (not acceptable for product).

Application Access Policy (Teams PowerShell) already applied:

• Policy: Tag:STELLA-Graph-OnlineMeetings → AppId 4b1d32ba-6c39-4a0b-b320-31981ce0654b
• Granted to: mj@ilovestella.de, mj@cdtechnologies.de, mc@ilovestella.de
• Meeting create returns _source: onlineMeetings (MS… meeting ID) — Ghost Leg persists.

---

Reproduction evidence (2026-06-30, Policy Test)

Field	Value
Meeting subject	STELLA onlineMeetings Policy Test 2026-06-30
onlineMeeting ID	MSplYzJiODRlNy01MTJlLTQyMzUtOTYzYi01YjJlOTA0MjNhNmMqMCoqMTk6bWVldGluZ19OalpqTkdVek5HVXROVEkxT1MwMFlXRTVMVGt3Wm1FdE1HSTFNbUZpWlRZek1HRTFAdGhyZWFkLnYy
groupCallId (thread)	66c4e34e-5259-4aa9-90fa-0b52abe630a5
ACS call_connection_id	6d003a80-a129-48b2-a1ec-22e979b71feb
Join URL thread	19:meeting_NjZjNGUzNGUtNTI1OS00YWE5LTkwZmEtMGI1MmFiZTYzMGE1@thread.v2
Callback URL	https://api.cdtechnologies.de/api/teams/copilot-acs-media/events
Backend flags	TEAMS_COPILOT_AUDIO_PREFER_RAILWAY_ACS=true, Runway broadcast off
Human in call	Yes (Teams desktop, mic bar active, second machine)
Result	PlayCompleted, no audible output in meeting

Earlier retest (calendar fallback, same symptom):

• groupCallId: e6613890-2b4e-4b2d-80b9-1b716e146e36
• ACS call: 1e003a80-2f9a-4348-a143-3b8206013c61

---

Graph / Entra permissions (Teams app 4b1d32ba…)

Application permissions (admin consented), including:

• Calls.JoinGroupCall.All, Calls.InitiateGroupCall.All
• OnlineMeetings.ReadWrite.All, OnlineMeetings.Read.All
• Calendars.ReadWrite, Calendars.Read
• Chat.ReadWrite.All, TeamsAppInstallation.ReadWriteForChat.All (auto-install still 403 without Consent variant)
• User.Read.All, CallRecords.Read.All, etc.

Bot Framework app (separate registration) handles /api/teams/bot/messages — chat works; does not provide meeting media.

---

What we need from Microsoft

Please advise on the recommended production path for our goals:

A) Outbound meeting audio (Phase 1 — voice only)

1. Is ACS Call Automation playToAll into Teams native online meetings a supported scenario? If yes, why would list_participants=0 + silent audio occur despite CallConnected/PlayCompleted?
2. Should we use Graph Cloud Communications / Application-hosted media bot instead (Windows VM + Calls.AccessMedia.All + media platform certificate)?
3. Is there a service-hosted media bot or Teams Real-time API path that avoids Windows VM ops but still delivers audible mixed audio?
4. Any tenant-level enablement required beyond Application Access Policy (e.g. Cloud Communications error 7504, policy flags)?

B) Inbound audio — bidirectional voice (Phase 1)

5. How should a bot receive participant audio for STT in a Teams meeting?
   • Graph AudioSocket / subscribed media?
   • ACS receive path?
   • Teams transcription API (too high latency for live dialog)?
6. Minimum architecture for full-duplex voice (human speaks → bot hears → bot responds with TTS in meeting)?

C) Outbound video — live avatar (Phase 2)

7. Recommended approach to inject live H.264 video (AI avatar, ~25 fps) as a visible meeting participant:
   • Application-hosted media only?
   • Can ACS carry video into Teams grid or audio-only?
8. Can audio and video use the same Graph call leg, or must they be separate?
9. Licensing / SKU requirements (Teams Premium, Copilot, etc.)?

D) Multi-tenant SaaS

10. For customer tenants (not our own): recommended onboarding — Application Access Policy per tenant, consent model, hosted vs customer-deployed media bot?
11. Any certification or Teams store requirements for a copilot with media injection?

---

Alternatives we are evaluating (please comment)

Option	Our understanding	Question for Microsoft
ACS Call Automation	Simple HTTP, Linux-friendly; current Ghost Leg	Supported for Teams meetings? Fix for participants=0?
Application-hosted Media Bot	Windows VM, Graph Communications SDK, AudioSocket/VideoSocket	Is this the only supported path for grid video + reliable audio?
Service-hosted media bot	Microsoft hosts media stack	Available for our scenario / tenant size?
Teams RTMP / live events	Not interactive meeting	Not suitable — need 2-way
Screen share / computer sound	Operator workaround only	Not productizable
Azure OpenAI Realtime + ACS	Voice AI patterns	Overlap with Teams meeting join?

---

Business context

• Product: STELLA — AI copilot for regulated industries (energy consulting, sales meetings).
• Use case: AI assistant joins customer Teams meetings, introduces itself (ISB compliance script), answers questions audibly, later with live avatar.
• Requirement: Bidirectional communication; voice-first MVP, video avatar second phase.
• Scale: Multi-tenant; each customer brings their own M365 tenant.

---

Severity / impact

• Severity: High — blocks core product capability (audible AI in Teams meetings).
• Workaround: Operator shares "computer sound" from browser Sidecar — not acceptable for customers.

---

Attachments / logs we can provide on request

• ACS webhook event sequence (CallConnected, PlayCompleted timestamps)
• Redacted Railway logs for call_connection_id above
• Graph meeting create response (onlineMeetings)
• Architecture ADR: docs/architecture/TEAMS_LIVE_AVATAR_MEDIA_BOT_DECISION_2026-06.md (private repo — summary included above)

---

Contact

• Primary: Markus Jaszczyk — mj@ilovestella.de / mj@cdtechnologies.de
• Organization: C&D Technologies GmbH
• Tenant ID: fe87c2f4-700c-4a68-9bd7-218fde87a857

---

Nach dem Absenden

1. Correlation IDs aus Azure Portal notieren und in docs/ops/TEAMS_ACS_GHOST_LEG_INCIDENT_2026-06-30.md ergänzen.
2. Antwort mit Architektur-Empfehlung → ADR aktualisieren + P1 Media Bot Scope anpassen.
3. Falls 7504 / tenant enablement nötig: separates Enablement-Ticket referenzieren.

Verwandte interne Docs

• TEAMS_ACS_GHOST_LEG_INCIDENT_2026-06-30.md
• TEAMS_GRAPH_MEDIA_BOT_P1_HANDOFF_2026-06-30.md
• TEAMS_VOICE_AUDIO_TEST_HANDOVER_2026-06-29.md
• services/teams-media-bot/README.md