Philipp Hancke tests why voice agents sometimes answer their own greetings. In his setup, echo reached transcription in 21% of Edge sessions and 4% of Chrome sessions. AEC dumps show how echo cancellation, gain control, and VAD thresholds interact.
Blackbox Exploration
Google Chrome downloads a new neural AEC while you are not looking
Chrome quietly delivers a neural model that changes part of its echo-cancellation pipeline. Philipp Hancke reverse engineers the model and compares echo suppression and speech recognition against AEC3.
How OpenAI does WebRTC in the new gpt-realtime
Earlier this month, OpenAI released the GA version of its realtime API. This includes many capabilities that the Beta didn’t have, including video support. I started out doing an update to the The Unofficial Guide to OpenAI’s Realtime WebRTC API I made for the Beta release last November. I discovered there were enough WebRTC updates […]
Measuring the response latency of OpenAIs WebRTC-based Realtime API
As Chad mentioned in his post last week, we have been diving into what OpenAI is doing with WebRTC. Over the last months, we actually did a full teardown and compared OpenAI’s Realtime API to what powers chatgpt.com. What intrigued us most was how to measure response latency. One of the key metrics for any […]
How Cloudflare Glares at WebRTC with WHIP and WHEP
WebRTC blackbox reverse engineering experts Gustavo and Fippo take a look at Cloudflare’s new WebRTC implementation, how Cloudflare uses the new WebRTC-based streaming standards WHIP and WHEP, and the bold pronouncement that they can be a replacement to open source solutions.