Task 2.2: Implement AI solutions for text and speech by using Foundry

• APrompt a general-purpose chat model to name the language of each email

  Incorrect

  A chat model can identify languages, but its wording and formatting can vary from call to call, which makes automated parsing fragile.

• BUse language identification in Azure Speech on each email's text

  Incorrect

  Language identification in Azure Speech works on spoken audio, not on written email text.

• CCall language detection in Azure Language through the Azure Language SDK

  Correct

  Azure Language returns structured, predictable results: the detected language, its ISO 639-1 code and a confidence score between 0 and 1.

• DRun PII detection in Azure Language and read the language from its entity categories

  Incorrect

  PII detection finds personal information such as names and phone numbers. Its categories don't report the language of the text.

General-purpose models are flexible but generate text probabilistically, so results can vary between calls. When an app needs consistent, structured values such as a language code or confidence score, Azure Language's purpose-built analyzers are the better fit.

• AAdd the Azure Language MCP server from the Foundry Tool Catalog as a tool

  Correct

  The Azure Language MCP server exposes Language features as tools over the Model Context Protocol. A Foundry resource already includes access to Language, so no separate resource is needed.

• BThe Azure Language REST endpoint and key, pasted into the agent's instructions

  Incorrect

  Instructions are text the model reads, not a tool connection, and putting a key in them exposes a secret.

• CThe web search tool, so the agent can find language analysis services online

  Incorrect

  Web search grounds answers in public web content. It doesn't call Azure Language on the ticket text.

• DA fine-tuned version of the agent's model, trained on labeled support tickets

  Incorrect

  Fine-tuning changes how the model itself responds. It doesn't connect the agent to Azure Language's analyzers.

MCP is an open standard for connecting agents to tools and data. The Azure Language MCP server, available as a remote server in the Foundry Tool Catalog, lets an agent call Language features without custom integration code.

• AA WebSocket connection from the browser

  Incorrect

  WebSocket connections suit server-to-server integrations such as backend services. They add latency for a browser client.

• BA WebRTC connection from the browser

  Correct

  WebRTC is recommended for client-side apps such as web and mobile, and gives the lowest latency for real-time audio streaming.

• CA SIP trunk connection to a phone system

  Incorrect

  SIP is for telephony integrations such as call centers and IVR systems, not browser apps.

• DFast transcription calls to the REST API

  Incorrect

  Fast transcription returns the transcript of a recorded file. It can't hold a live, two-way spoken conversation.

The GPT Realtime API supports WebRTC, WebSocket and SIP. For most real-time audio, especially from a browser or mobile client, Microsoft recommends WebRTC for low-latency streaming.

• AUse a SpeechSynthesizer instead of a SpeechRecognizer for the audio files

  Incorrect

  SpeechSynthesizer produces audio from text. The app needs the opposite.

• BTrain a custom speech model with vocabulary taken from the lecture topics

  Incorrect

  Custom speech improves recognition accuracy for specialized words. It doesn't make a single-utterance call transcribe a whole lecture.

• CSet speech_recognition_language to the lecturer's exact locale

  Incorrect

  Setting the right locale improves accuracy, but recognize-once still stops after the first utterance.

• DSwitch from recognize-once to continuous recognition for the whole file

  Correct

  Recognize-once handles a single utterance of up to about 30 seconds or until silence. Continuous recognition keeps transcribing long audio.

The recognize-once operation transcribes one utterance, up to about 30 seconds or until silence is detected. For longer audio, use continuous recognition and stop it when the file has been processed.

• ASet speech_recognition_language on the SpeechConfig to the voice's locale

  Incorrect

  speech_recognition_language sets the language for recognizing spoken input. It doesn't select a voice for synthesis.

• BCall set_speech_synthesis_output_format() with a 24 kHz RIFF format

  Incorrect

  The output format sets the audio file type, sample rate and bit depth. It doesn't choose which voice speaks the text.

• CSet speech_synthesis_voice_name on the SpeechConfig to that voice

  Correct

  The voice name on the SpeechConfig tells the synthesizer which voice to use, and it takes priority over a synthesis language setting.

• DApply for limited access and train a custom voice for the confirmations

  Incorrect

  Custom voice creates a brand-specific voice and requires a limited access application. A standard voice from the gallery needs neither.

Text to speech needs the text to speak and the voice to use. With the Speech SDK, you choose a standard voice by setting its name on the SpeechConfig; the output format only changes how the audio is encoded, and custom voices are a separate, limited access option.

• ANoise suppression and echo cancellation on the incoming audio

  Correct

  Voice Live includes noise suppression and echo cancellation, so the assistant hears the caller clearly and doesn't pick up its own voice.

• BWriting the assistant's instructions and choosing which tools it may call

  Incorrect

  The developer still defines the assistant's instructions and tools; Voice Live handles the speech experience around them.

• CTraining a custom brand voice without a limited access application

  Incorrect

  Custom voice requires a limited access application; Voice Live doesn't remove that requirement.

• DTranslating statement PDFs into other languages while keeping their layout

  Incorrect

  Translating documents with their formatting is done by document translation in Azure Translator.

• EEnd-of-turn and interruption detection, so callers can pause or cut in naturally

  Correct

  Advanced end-of-turn detection and robust interruption detection are built into Voice Live's conversational features.

Voice Live is a fully managed API that combines speech recognition, generative AI and text to speech in one low-latency interface, with conversational features such as noise suppression, echo cancellation, interruption handling and end-of-turn detection.

• AAn Azure AI Search resource

  Incorrect

  Azure AI Search provides indexing and retrieval, not speech recognition.

• BAn Azure Storage account

  Incorrect

  Storage holds files such as recordings; it doesn't run speech recognition.

• CThe Microsoft Foundry resource

  Correct

  Azure Speech is available through a Foundry resource, whose endpoint and key authenticate the Speech SDK connection.

• DAn Azure Key Vault for secrets

  Incorrect

  Key Vault can store secrets safely, but it isn't the service the SDK connects to.

To use Azure Speech from code, create a Foundry resource and use its endpoint and key, or Microsoft Entra ID, to authenticate the Speech SDK. Azure Speech is one of the Foundry Tools that the resource provides.

• AThe Speech CLI

  Correct

  The Speech CLI lets you use Azure Speech without writing code, and most Speech SDK features are available in it.

• BThe Azure Language SDK

  Incorrect

  The Language SDK is a code library for text analysis, not audio transcription.

• CThe Speech SDK in a short Python script

  Incorrect

  The Speech SDK transcribes audio, but it's a client library: using it means writing and running application code.

• DThe fast transcription REST API from a custom client

  Incorrect

  Fast transcription returns transcripts quickly, but calling it from a client you build is writing an application.

Azure Speech can be used through the Speech SDK, REST APIs or the Speech CLI. The SDK and REST APIs are for code you write; the CLI is a command-line tool that exposes most SDK features without writing code.

• Aazure-ai-textanalytics

  Incorrect

  That package is the Azure Language client library for text analysis.

• Bazure-ai-voicelive

  Correct

  Python Voice Live applications use the azure-ai-voicelive package, along with packages such as pyaudio and azure-identity.

• Cazure-ai-contentunderstanding

  Incorrect

  That package calls Content Understanding analyzers.

• Dazure-cognitiveservices-speech for image input

  Incorrect

  The Speech SDK handles speech features, and it doesn't take image input.

The Voice Live sample code handles the session, audio devices, streams and interruptions. In Python it relies on the azure-ai-voicelive package, with pyaudio, python-dotenv and azure-identity.

• AIn the model leaderboards, which run speech tests on each model

  Incorrect

  Leaderboards compare language models on benchmarks; they don't host service playgrounds.

• BOn the AI services tab of the Build page's Models section

  Correct

  The new portal lists AI services, such as Azure Speech and Azure Language capabilities, for testing on the AI services tab.

• COn the resource's keys and endpoint page in the Azure portal

  Incorrect

  That page shows the key and endpoint your code uses to connect. It isn't in the Foundry portal and offers no interactive testing.

• DIn the model playground, after deploying a chat model from the catalog

  Incorrect

  The model playground tests deployed models with prompts and parameters. Azure Speech and Azure Language are Foundry Tools that the portal offers on the AI services tab.

In the new Foundry portal, the Build page's Models section has an AI services tab with playgrounds for Foundry Tools such as Azure Speech speech to text and text to speech and Azure Language features.