Texts and conversations are the biggest source at Nordlicht Reisen GmbH: about 300 customer reviews a day, plus the calls to the hotline. Nobody reads or hears everything. This lesson shows which techniques turn text into usable features (text analysis) and how spoken language becomes text and back (speech). The exam names four text analysis techniques explicitly: key phrases, entities, sentiment and summarization. For speech, it asks about speech recognition and speech output.
Text analysis with Azure Language
Azure Language (on Learn "Azure Language in Foundry Tools") is a cloud service for text analysis. You send text, and the service returns results. According to Learn, the analysis runs "as-is": the model is not adapted to your data. Many of the features can be tried without code in the Foundry portal.
| Technique | What comes out | Nordlicht example |
|---|---|---|
| Key phrases (key phrase extraction) | the main topics of a text as a list | "The food was delicious and the staff wonderful" becomes "food" and "wonderful staff". |
| Entities (named entity recognition, NER) | people, places, organizations and quantities, sorted into predefined types | Hotel name, place and date in bewertung-2.txt |
| Sentiment (sentiment analysis) | a label (negative, neutral or positive) per sentence and per document, plus confidence scores | Is a hotel slipping in last week's reviews? |
| Summarization (summarization) | a short text from a long one | Capture a call in a few sentences |
| Language detection (language detection) | the language of a text | Separate German and English messages |
| Personal data (PII detection) | information that identifies people, in text and conversations | Replace names and addresses before saving |
For some techniques there are subtleties that the exam knows:
- Sentiment and opinion analysis: Opinion analysis (opinion mining) assigns the sentiment to individual aspects, such as "room" or "breakfast". According to Learn, the confidence scores lie between 0 and 1.
- Entities: The prebuilt recognition uses a fixed list of types. With custom recognition (custom NER), you train your own categories on your texts.
- Summarization: The extractive form selects important sentences from the original. The abstractive form writes new, concise sentences that are not in the text as such. For conversations, Learn names, for example, summaries of issue and resolution, as a call center needs them.
- Language detection: According to Learn, it recognizes more than a hundred languages.
The status of the features according to Learn
Important and close to the exam: Learn divides the features of Azure Language into "Core capabilities" and "Legacy capabilities". The core capabilities (recommended for new development) include the detection of personal data, language detection and entity recognition. Key phrases, sentiment analysis and summarization are listed under "Legacy". Their pages announce that the features are being retired from Azure Language and steer new projects to Foundry models. The date is on the pages; it does not belong in a lesson that stays valid for a long time. The techniques remain the same for the exam and for practice. Only the way to get there can be a language model instead of the feature. Whether the exam follows this is open.
Speech with Azure Speech
Azure Speech (on Learn "Azure Speech in Foundry Tools") offers, among other things, speech recognition, speech output, translation and language identification, likewise through a Foundry resource.
Speech recognition (speech to text) turns audio into text. Learn names three ways:
- Real-time transcription for live audio,
- Fast transcription for recorded files,
- Batch transcription for large volumes, processed asynchronously.
A base model is not always enough, for example with background noise or technical terms. Then you can train a custom speech model (custom speech) with your data. It stays private. Newer models that Learn names, for example Azure LLM Speech, improve quality and understanding of context.
Speech output (text to speech) turns text into natural-sounding speech. Neural voices stand behind it. With SSML (Speech Synthesis Markup Language), you control pitch, pronunciation, speed and volume. As voices, Learn names standard voices and custom voices. A custom voice (custom voice) is recognizable and unique to your brand or your product and stays private. For it, Learn points to its own guidance on responsible use, including limited access (limited access). What exactly applies, you read there.
Translation (speech translation) translates spoken language in real time, from speech to speech or from speech to text. Language identification (language identification) recognizes which language is spoken in the audio, alone or together with recognition and translation.
Nordlicht: The file anruf-hotline.wav goes through speech recognition. For many recordings at once, batch transcription fits; for a single file, the fast one. The result is text, and text is what text analysis works on: summarization, sentiment, detection of personal data (lesson 1.1). Several workloads are connected for this (lesson 1.4). Language identification helps when calls arrive in German and English.
Exercise 1
Which technique or feature fits? Choose one each and justify it in one sentence:
- Nordlicht wants all hotel names and places mentioned in the reviews as a list.
- An evaluation should show whether a hotel's reviews are rather positive or negative.
- A room description is read aloud by voice for visually impaired travelers.
- Hundreds of hotline recordings should be turned into text overnight.
- Before saving a conversation transcript, names and phone numbers should be replaced.
- A call should be captured in two or three sentences with issue and resolution.
Solution 1
- Entity recognition (NER). It returns places, organizations and other types from a fixed list.
- Sentiment analysis. The label negative, neutral or positive applies per sentence and document. For aspects such as room or breakfast, opinion analysis helps.
- Speech output (text to speech), controlled with SSML if speed or pronunciation should be adjusted.
- Batch transcription of speech recognition, because large volumes are processed without hurry.
- Detection of personal data (PII). It works on text and conversations.
- Summarization of conversations, here with issue and resolution.
Nos. 2 and 6 belong to the features that Learn marks as being retired. The technique remains; the way can be a Foundry model.