To select a service, write a sentence with the entry and result: "audio to text," "photo to localised objects" or "invoice to fields." That step often resolves a great deal of the decision.
Objectives: choose Language, Speech, Vision and Document Understanding; recognition of their departures; to interpret custom models and scores; and following a demonstration without confusing an old interface with current capabilities.
Key concepts
| Service | Typical Input | Departure or task |
|---|---|---|
| Language | Text | Language, entities, themes, feelings, key phrases or translation |
| Speech | Audio or video with a voice | Transcription and Temporary Marks |
| Vision | Images or video | Classification, detection and visual location |
| Document Understanding | Documents and images of documents | Text, fields, tables and document type |
They are currently available with apis and other OCI interfaces. An output comes with structure and, depending on operation, confidence or coordinates. You have to check fields, quality and permissions before inserting them into a process.
Language and Speech
Language detects language, removes entities (NER), identifies key phrases, classifies themes and analyzes feelings. In "Reparation was good but delivery took too long," the feeling by appearance separates repair and surrender. It does not have to be reduced to a single valuation.
The custom models allow them to adapt classification and NER to their own categories with examples tagged. Detection of PII o PHI helps to identify personal or health information; The flow should decide how to mask, withdraw or protect what was found. Health NLP removes entities and clinical relations: do not automatically transform these data into a validated diagnosis.
Speech use ASR. A job processes asynchronous files and writes results to Object Storage. Live Transcribe attends audio in real time. JSON allows to process text and metadata. SRT It's for subtitles. The standardization represents numbers or dates legible, while the Diarrisation difference of speaker shifts, doesn't prove his civil identity.
an offensive language filter can remove, mask or label tag He keeps his word with a mark. Check confidence and audio scores if there's a doubt. Languages and functions depend on the model: Oracle ASR and Whisper have different coverings and an old list isn't enough to rule out a language. Query la Current Speech documentation.
Vision and Document Understanding
At Vision, photo classification assign labels to the scene. Object detection adds delimiting boxes to place instances. OCR recognizes visible text. Facing locates faces; That doesn't mean identifying people. If your domain's object isn't part of the pre-trained categories, evaluate a custom model and tagged data.
At document understanding, text extraction returns words and lines with location. tables extraction holds rows and columns; carnation-value extraction links fields as a supplier, date and total Documentary classification identify types as an invoice or receipt.
Some old films and scills have been shown with document AI within vision. To extract documentary structure, this guide uses today's service Document Understanding, according to his documentation. Don't learn that historic association as a permanent rule.
Step-by-step example
A repair file contains a call, a photo of the equipment and an invoice.
- Speech transforms the call review words of low confidence that change the meaning of the problem.
- Language categorizes text and distinguishes feeling about care and repair.
- Vision locates photo components if the model recognizes them. document them when they aren't identified.
- Document Understanding classifies the invoice and removes fields and table lines.
- The application checks that quantity, prices and taxes explain the total before registration of expenditure.
In console, inspect structured result as well as labels about image. A misread seal or text can end up in a wrong column: that the request finish correctly does not prove that every field is correct.
Common errors
- Choose Language directly for an audio file without transcribing it.
- To merge detection boxes with general classification of image.
- To merge OCR with an extraction of links between rows, columns and keys.
- To say that a preset model recognizes all objects of your industry.
- Using high score as an automatic invoice approval.
Selection exercise
You need subtitles for a recording, place three equipment in a photo and recover quantity and price from each row of an invoice. What would you wear?
Solution
Speech with SRT output for subtitles; Vision with detection of objects for location Document Understanding with tables extraction for rows and columns. If you also need supplier and total as fields, add carne-value extraction. OCR only returns text and its location. It's not a guarantee of proper tabular structure.
Sources and review
Query Language, Vision and the previous Speech and Document Understanding sources. Repeat the demonstrations at MyLearning describing what every field of response represents.