3.3.2. Speech and Computer Vision Workloads
💡 First Principle: Speech and vision workloads are about perception — turning audio or images into information, or the reverse. The key is direction: recognition/analysis goes from media to data; synthesis/generation goes from data to media. Naming the direction names the workload.
On the speech side: speech recognition (speech-to-text) converts spoken audio into written text; speech synthesis (text-to-speech) generates spoken audio from text. On the vision side: image classification (what is this image overall?), object detection (what objects are present and where?), and optical character recognition / OCR (read text out of an image) all interpret existing images, while image generation creates new images from a text prompt.
⚠️ Exam Trap: Image classification vs. object detection: classification gives one label for the whole image ("this is a street scene"); object detection finds multiple items and their locations ("car at these coordinates, person at those"). Scenarios that need counting or locating things call for detection, not classification.
Reflection Question: A warehouse wants to count how many pallets appear in a photo and where each one is. Is that classification, object detection, or OCR? Why?
