15 · ACS, Speech, Copilot

Azure AI Speech

Speech-to-text and text-to-speech are services with their own language, voice, and approval rules.

Room HTMLAll downloads

Learn it in 5 minutes

Azure AI Speech can transcribe the caller and synthesize the prompt. The language of the prompt has to match the caller or every intent becomes unknown. A custom neural voice has approval rules. Do not assume you can clone a person.

Log the recognition result beside the call id. When the bot is wrong, you need the text it thought it heard, not a feeling.

Azure AI Speech can talk and listen. It is not emergency calling, and you log what it thought it heard.

Room HTML

Visual

Dynamics unified routingWorkstreamClassificationOutagecap 2Billingcap 1Generalcap 0
Language, intent, and capacity for Azure AI Speech. Overflow is a rule.
Speech is not the phone systemTTS · PromptSTT · RecognizePSTN · Not this service
It plugs into the recognize step. Counts in this drawing are teaching order, not customers.

SVGPNGHTML

Worked example

I pick a voice I am allowed to use, match the language, and log what the recognizer heard.

  1. Voice One you may use
  2. Language Matches the caller
  3. Result Logged
  4. Fallback Human

If it fails: The prompt language does not match the caller and every intent is unknown.

Checklist HTML

Quiz

A colleague reports: The agent is the only target of the public number, including emergencies. Where do you start?

Which description of Azure AI Foundry is accurate?

A colleague reports: A model change ships to the IVR with no eval. Where do you start?

Flashcards

Ecosystem

Glossary

  • Call Automation The ACS server-side calling API. Your app answers, plays, recognizes DTMF or speech, and transfers. You receive events. You do not drive a desktop click. Config: Subscribe to the events and time out the recognize step. Fault: The app never answers, so the caller hears ringing until the platform gives up.
  • Calling SDK The ACS client SDK for a human in a browser or a mobile app. Different from Call Automation, which is the server. Config: Pick the SDK if a person is on the page. Pick automation if a bot answers. Fault: A web page uses the calling SDK and nobody handles the server events.
  • Recognize step The Call Automation action that collects DTMF or speech. It needs a timeout and a path for ‘none of the above.’ Config: Map 0 to a human queue and test it. Fault: The bot repeats the prompt until the caller hangs up.
  • Azure AI Speech Microsoft’s speech-to-text and text-to-speech service. Copilot Studio and custom IVRs can use it. It is not emergency calling. Config: Choose a voice you are allowed to use and log the recognition result. Fault: The prompt language does not match the caller and every intent is unknown.

Sister technologies

Around the call: compliance recording, workforce management, emergency location, analytics, CRM screen pop, and an agent with a human escape. Each one has its own failure.

Config and troubleshooting

The prompt language does not match the caller and every intent is unknown.

Cross-platform bridges

CUCM

Cisco CUBE is one of the SBC families named in this lab’s certification lesson. A certified release can be the Direct Routing SBC between Unified CM and Teams.

The tenant still needs a verified domain on the SBC FQDN, and the user still needs Teams Phone. The studio translator does not open a live AXL session. Confirm the current certified release before you build.

Webex

Teams Phone and Webex Calling are separate clouds. This lab does not describe a feature that registers a Webex Calling user into a Teams tenant.

A number move is usually a carrier port, a certified SBC, or a meetings interop product. Confirm before relying.

UCCE

Unified CCE stays the ACD until you migrate it. Microsoft’s contact-center articles name Connect, Extend, and Unify. They do not name a UCCE connector in the notes this lab cites.

Voice into Teams is still Direct Routing or Operator Connect. The routing script stays on the contact-center product you bought. Confirm the vendor’s current model. Source: https://learn.microsoft.com/en-us/microsoftteams/teams-contact-center.

Future hooks

Copilot Studio, Azure Communication Services, and Graph are later tracks in this lab. Treat any uncited limit as confirm before relying.

BridgesEcosystem

Sources and go deeper

Opened 2026-10. A row marked confirm did not open. Wikipedia is background. The Teams blog row is the community index.

  • Azure AI Speech overview — Microsoft Learn — last checked 2026-10.
  • Microsoft Teams — Wikipedia — last checked 2026-10. Wikipedia fallback for background. Confirm the vendor page before you change a tenant
  • Microsoft Teams blog index — Community — last checked 2026-10. Community index for the Microsoft Teams blog. This is the index, not a claim that a post discusses this room

sources.json