# Complex AI product examples Example: smart speakers Command (trigger word/wake word): "Hey device, tell me a joke" Steps to process the command (aka **AI pipeline**): - Detect/Recognize the trigger word / wake word - Audio -> "Hey device"? (0|1) - A -> B mapping - Speech recognition - Audio -> "tell me a joke" - Different A -> B mapping - Text transcript - Intent recognition - Understand what you want to do - Command categories: joke, time, music, call, weather, etc - Map the input (text transcript) to an output (command category) - Execution - Execute the recognized command Joke: "Why are there so many shocking results in AI? Because AI is the new electricity" For a more complex command: "Hey device, set timer for 10 minutes" * Trigger word/wake word detection * Speech recognition (audio -> set timer for 10 minutes) * Intent recognition (set timer for 10 minutes -> timer) * Execution * Extract the duration from the transcript (10 minutes -> 10) * Set timer for 10 minutes * Let me know when 10 minutes is up * ... * Start the timer for the asked duration using a specialized program Other examples: - Play music - Volume up/down - Make call - Current time - Unit conversion - Simple question - ... Other example: self-driving car. Steps for deciding how to drive: - Recognize/detect objects/persons around using Image/radar/lidar/other sensors data (e.g., GPS, maps, etc) - Motion planning - How to move forward toward the goal, avoiding any obstacles, following lanes and respecting other rules - Steer/Accelerate/Break Key steps: - Car detection - Use 360 cameras and other sensors (radar, lidar, etc) - Pedestrian detection - Lane detection - Traffic light detection - Obstacles detection ... - Trajectory prediction - Where everything is likely to go next - ... - Motion planning - Output the path to follow and speed/angle etc