What if you could transform your text into a compelling audiobook—without ever recording a word? In this talk, we’ll dive into using Azure AI Speech to turn long-form content into immersive audiobooks. You’ll discover the differences between real-time and batch synthesis, the quirks of SSML formatting, how to instruct Azure OpenAI to generate valid rich SSML markup to explicitly direct the speech synthesis, what it takes to train a custom voice, and the strengths and pitfalls of AI-generated narration. Expect real-world examples, hands-on insights, and a discussion on the evolving role of AI in storytelling.

Finland is a relatively large country by area with over 400,000 kilometres of power grid, most of which is still located above ground. Storms, heavy snow, and fallen trees can cause power outages that are expensive to fix at short notice, especially in the more remote parts of the country. Future problems can be recognised by photographing the power grid from the air, but going through the resulting images is very mind-numbing and time-consuming work. This session is based on a real customer project: instructing a vision model in Microsoft Foundry to pre-screen helicopter photographs of the power grid and flag potential faults, from trees leaning over the lines to damaged poles, so that humans only need to review the roughly ten per cent of images where something might actually be wrong. Confirmed findings flow onward as coordinates and work orders to deal with the discovered problems. We will walk through the complete solution: images landing in a data storage, automation firing them off to a vision-capable AI model, detections stored in a database, and a custom web app where a human reviews the AI's findings before they are forwarded to the field work management system. Just as importantly, you'll hear honest talk about what worked and what didn't: why the model spots some fault types reliably yet misses others entirely, and why a perfectly normal situation can look like a problem. In this session, you will learn how to: - Build a serverless image-analysis pipeline on Azure using Blob Storage, Functions, Table Storage, and a vision-capable GPT model in Microsoft Foundry. - Work within the practical limits of language-model vision: image resolution constraints, why naively splitting large images backfires, and how clear something needs to be before the model will spot it. - Design a human-in-the-loop review flow for a process where AI cannot be fully trusted. - Handle real-world API concerns when processing tens of thousands of images. - Judge when an off-the-shelf language model plus human review beats training a custom computer-vision model, and understand the cost difference. The project was delivered on a deliberately limited budget, because management was not yet convinced the investment would pay off. It did. My hope is that this session sparks ideas for how you can innovate and streamline major business processes with AI, beyond chatbots and text generation, in the places where the big money and savings actually are. This session is well-suited for developers, business decision makers, and consultants who want to see what practical, affordable AI looks like when it is applied to a large, unglamorous, and genuinely expensive business problem.