Workshop: Building and optimizing dictation features

Topics

Description from the channel

Dictation has become a feature everyone is adding to their product. In this workshop, Andrew Galyan-Mann from the AssemblyAI forward-deployed engineering team builds a dictation feature into a mock Slack app, chapter by chapter, and shows how to get finished text on screen in under a second. What you'll learn: • Why the Sync API fits dictation better than async or streaming. Async has a floor of several seconds, and streaming means standing up a WebSocket for every short burst. Sync is a single POST request with a finished Universal-3.5 Pro transcript in the response. • How key terms prompting gets names and product vocabulary right before any cleanup happens. • How to warm the connection while the user is still recording, so the DNS, TCP, and TLS cost is paid before the upload and the request goes straight to inference. • How to run a cleanup pass on LLM Gateway with Qwen3.5 4B Fast, a small model optimized and hosted by AssemblyAI. Andrew walks through the prompt design: tell the model it is a function in a pipeline rather than an assistant, spell out which corrections to resolve, preserve intent, and give few-shot examples of what to fix and what to leave alone. • How to think about the latency budget across Speech-to-Text inference, transport, and cleanup, and which levers bring the total down. • A first look at the Dictation API, which wraps the whole loop in one call and returns both the verbatim transcript and the cleaned result. ▬▬▬▬▬▬▬▬▬▬▬▬ CONNECT ▬▬▬▬▬▬▬▬▬▬▬▬ 🖥️ Website: https://www.assemblyai.com 🐦 Twitter: https://twitter.com/AssemblyAI 🦾 Discord: https://discord.gg/Cd8MyVJAXd ▶️ Subscribe: https://www.youtube.com/c/AssemblyAI?sub_confirmation=1 🔥 We're hiring! Check our open roles: https://www.assemblyai.com/careers ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ #MachineLearning #DeepLearning