AI-901 lesson 4: text analysis and speech
Text analysis takes existing text and returns something structured or shorter: the main phrases, the named things, the feeling, or a condensed version. Speech services convert between audio and text: recognition (speech to text) turns spoken words into a transcript, and synthesis (text to speech) turns written text into spoken audio. The exam asks which technique produces the output a scenario needs, so the skill is matching each technique to the shape of its result.
Four text analysis techniques
These capabilities are offered by Azure Language, part of Foundry Tools (older articles call the group Azure AI services). Each is a preconfigured feature: send text, read the result.
| Technique | What comes back | Example output |
|---|---|---|
| Keyword extraction (key phrase extraction) | The main topics as a list of short phrases | “late delivery”, “refund”, “damaged box” |
| Entity detection (named entity recognition) | Items in the text, each tagged with a category | “Rotterdam” as a location, “12 March” as a date, “Contoso” as an organisation |
| Sentiment analysis | A positive, neutral or negative label with confidence scores, for the text and for sentences | “negative, 0.92” |
| Summarization | A shorter version of long text or a conversation | A three-sentence digest of a call transcript |
Summarization can be extractive, which selects the most important sentences from the source, or abstractive, which writes new sentences that capture the meaning. Microsoft’s documentation currently lists key phrase extraction, sentiment analysis and summarization among its legacy capabilities, supported for existing use, while named entity recognition is a core capability. For fundamentals, the exam concern is what each does, not its lifecycle status.
Telling them apart
One sentence of input can feed all four, and the question decides which. Take the sentence “The parcel from Contoso was three days late and the driver was rude.” Keyword extraction lists phrases such as late parcel and rude driver. Entity detection finds Contoso as an organisation and “three days” as a duration. Sentiment analysis labels it negative. Summarization would only matter with longer text.
Look at what the scenario wants to receive: a list of topics, names with categories, a mood, or less text.
Speech recognition and synthesis
Azure Speech in Foundry Tools provides both directions.
- Speech recognition (speech to text) converts audio into text. It can transcribe a live stream in real time, process a recorded file quickly, or work through large volumes of recordings in batches. Typical uses are captions, meeting transcripts and call recordings. Custom models can help when audio has noise or specialist vocabulary.
- Speech synthesis (text to speech) produces natural-sounding spoken audio from text using neural voices. Typical uses are voice assistants, audiobooks, navigation prompts and reading content aloud. Markup can adjust pitch, speaking rate and pronunciation.
Azure Speech also covers speech translation and pronunciation assessment, but the two core directions are the ones to know first.
Worked example: a 150-person energy supplier handles calls
A 150-person energy supplier records customer calls and wants three things. First, searchable written records of each call: that is speech recognition. Second, a flag on calls where the customer sounds upset, plus a list of the meter numbers and street names mentioned: sentiment analysis and entity detection applied to the transcript. Third, a short summary line for each call in the case notes: summarization. Finally, the supplier wants its phone menu to read out the customer’s balance in a natural voice, which needs speech synthesis. Note the order for the first two requirements: text analysis works on text, so audio must be transcribed before sentiment or entities can be found. Later lessons show how a multimodal model can sometimes handle audio directly, but the building blocks stay the same.
Easy to confuse
| Often mixed up | The difference |
|---|---|
| Keyword extraction vs entity detection | Keywords are the main ideas as phrases; entities are specific items with a category such as person or location |
| Sentiment vs summarization | Sentiment returns a feeling label; summarization returns shorter text |
| Speech recognition vs synthesis | Audio in, text out vs text in, audio out |
| Text analysis vs generation | Analysis returns facts and labels about given text; generation writes new content |
| Entity detection vs information extraction | Entity detection tags items in running text; extraction (a later lesson) returns named fields from documents and other media |
Key points
- Keyword extraction lists main phrases; entity detection tags names, places, dates and similar items; sentiment gives a label; summarization shortens.
- Summarization can be extractive (selected sentences) or abstractive (newly written sentences).
- Speech recognition converts speech to text; speech synthesis converts text to speech.
- Text analysis needs text, so spoken input is transcribed first.
- Azure Language and Azure Speech are part of Foundry Tools, once grouped as Azure AI services.
Check yourself
Question 1. A bank wants to scan written complaints and quickly list the product names, branch cities and dates mentioned in each one. Which technique fits best?
- A. Entity detection
- B. Sentiment analysis
- C. Summarization
- D. Speech synthesis
Show answer
Answer: A
Finding categorised items such as products, locations and dates is entity detection. Sentiment analysis gives a feeling label, summarization shortens text, and speech synthesis produces audio.
Want more questions like this? Full AI-901 practice tests →
Question 2. A museum wants its guide app to read exhibit descriptions aloud to visitors through headphones. Which capability does this need?
- A. Speech recognition
- B. Keyword extraction
- C. Speech synthesis
- D. Sentiment analysis
Show answer
Answer: C
Turning written text into spoken audio is speech synthesis. Speech recognition does the opposite, keyword extraction returns phrases, and sentiment analysis labels feeling.
Want more questions like this? Full AI-901 practice tests →
Question 3. A call centre wants to know whether recorded customer calls are mostly positive or negative. Which sequence of capabilities is needed?
- A. Speech synthesis, then summarization
- B. Speech recognition, then sentiment analysis
- C. Speech recognition, then keyword extraction only
- D. Sentiment analysis directly on the audio file, then synthesis
Show answer
Answer: B
The audio must first become text through speech recognition, and sentiment analysis then labels that text. Starting with synthesis produces audio rather than consuming it, and keyword extraction or summarization would not give a positive or negative label.
Want more questions like this? Full AI-901 practice tests →
This lesson completes the text and speech part of the first exam domain, Identify AI concepts and capabilities, which carries 40 to 45% of AI-901, and the recap page AI concepts explained lists these techniques side by side.