Azure Language and Speech in Foundry Tools
Many exam questions describe a business need and ask which prebuilt AI feature solves it. For text and audio, the answer is almost always a feature of Azure Language or Azure Speech.
Azure Language
Azure Language in Foundry Tools (formerly Azure AI Language, and before that Text Analytics) analyses text. Most features are prebuilt, so you send text and get results with no training.
| Need | Feature |
|---|---|
| Positive, neutral or negative opinion | Sentiment analysis (and opinion mining) |
| Main talking points | Key phrase extraction |
| Language of the text | Language detection |
| People, places, dates, organisations | Named entity recognition (NER) |
| Link a name to a Wikipedia entry | Entity linking |
| Find and mask names, phone numbers, IDs | PII detection |
| Short version of a long document | Summarization |
| Answer from a knowledge base | Question answering |
Custom features exist when prebuilt ones are not enough: custom NER extracts your own entity types, and custom text classification assigns your own labels. These need labeled data and training.
Azure Speech
Azure Speech in Foundry Tools handles audio.
- Speech to text (recognition) converts spoken audio to text, in real time or in batch.
- Text to speech (synthesis) creates natural spoken audio, with neural voices.
- Speech translation translates spoken language in near real time.
- Speaker recognition identifies who is speaking.
For live conversations with a voice agent, Azure also offers Voice Live, which combines recognition, a model and synthesis in one low-latency connection.
Speech SDK building blocks
In Python the pattern is always the same:
- Create a SpeechConfig with your key and region (or endpoint). This is the connection to the service.
- Create an AudioConfig to choose the audio source (default microphone or a file) for recognition, or an AudioOutputConfig for the destination of synthesized audio (speaker or file).
- Create a SpeechRecognizer (audio to text) or a SpeechSynthesizer (text to audio) from those configs.
- Call a method such as
recognize_once_async()orspeak_text_async().
Remember: the config objects only describe settings. The recognizer and synthesizer do the work.
Exam traps
- Entity linking vs NER vs PII: NER finds entity types, entity linking ties them to a known source, PII detection finds sensitive data and can mask it.
- Key phrases vs summarization: key phrases are short terms, summarization produces sentences.
- Recognition vs synthesis: recognition is audio in, text out. Synthesis is the reverse.
- SpeechConfig vs AudioConfig: only SpeechConfig points to the Azure resource.
- Custom vs prebuilt: if the question says “your own categories” it needs a custom model. If it says “extract” common things, use prebuilt.
Practice questions
Choose an answer, then review the explanation to check your understanding.
1. A support team wants names, phone numbers and card numbers hidden in chat transcripts before the text is stored. Which Azure Language feature fits?
Show answer
Answer: B.PII detection finds personal data such as names and phone numbers and can return a redacted copy of the text.
2. A Python app must read text aloud and save the audio to a file. Which Azure Speech capability is needed?
Show answer
Answer: C.Speech synthesis (text to speech) turns text into audio. Recognition goes the other way.
3. Which object holds the key and region (or endpoint) when you start working with the Speech SDK?
Show answer
Answer: B.SpeechConfig carries the connection details. AudioConfig only chooses the microphone or file used as input or output.
4. You must find the main topics in thousands of product reviews without training a model. What do you use?
Show answer
Answer: B.Key phrase extraction is a prebuilt feature that returns the main talking points in text.