Sunday roundup: six posts from a week in voice AI

A developer's weekly roundup covers six posts on voice AI, including the EU AI Act's synthetic audio watermarking rules that took effect August 2, requiring TTS providers to implement marking pipelines. Kakao released Kanana-o, a Korean TTS model scoring 94.50 on Korean InstructTTSEval with natural language emotion control. Nous Research launched Hermes Agent v0.20.0, an open-source voice agent framework with streaming TTS and pluggable backends. The roundup also highlights rapid AI progress, noting SWE-bench scores jumped from 49% to 95% in a year, and Cloudflare Kitesurf's cost-efficient browser engine for agents.

Six posts this week. My keyboard filed a noise complaint somewhere around post four. Here is what I wrote and what stuck with me. The synthetic audio watermarking rules took effect August 2. I wrote about what TTS builders actually need to do about it. Short version: if your API serves generated audio, you need a marking pipeline before anyone downstream can ship. Speechify has been on this for months. Not every provider has. Kakao dropped Kanana-o, which scores 94.50 on the Korean InstructTTSEval and handles emotion through natural language. You write "say this excitedly" instead of <prosody rate="fast" and it just works. I want every TTS API to go this direction. SSML is powerful but nobody wakes up excited to write angle brackets. Nous Research released Hermes Agent v0.20.0 with streaming TTS, barge-in, on-device wake words, and pluggable STT/TTS. First open source voice agent framework that feels genuinely competitive with the proprietary options. The pluggable backend means no single-vendor lock-in at the voice layer. That matters. Throwback post comparing August 2025 to now. Best model then scored 49% on SWE-bench with 200K context. Today Claude Fable 5 scores 95% with 1M context. The thing that surprised me most is how fast context windows grew. A year ago you could not fit a whole codebase in one prompt. Now you can. That changes everything more than the benchmark scores do. Simon Willison turned a 2022 tweet into a playable 3D browser game with Claude Fable 5. One prompt, two images, and the model shipped a full game with mechanics, dialogue, and an inventory system. I wrote about it because this is the kind of output that keeps me optimistic about where this is going. Code generation for CRUD apps is predictable. Code generation for a raccoon heist game because someone thought it would be fun is not. That gap is where the interesting stuff lives. Covered Cloudflare Kitesurf, Agent Plugins, OpenAI's Astra pause, ByteDance's training run, and Anthropic's hardware plans. Kitesurf keeps feeling under-covered for how big the shift is. A browser engine that costs 3-7x less than Chromium for agent workloads changes the economics of running agents at scale. I wrote two more posts today that are in review. Claude Code cross-session messaging and a deeper look at Cloudflare Kitesurf. They will land tomorrow or Monday depending on when I get back to my desk and stop writing long enough to review them. At https://lukeocodes.dev/eu-ai-act-voice-watermarking https://lukeocodes.dev/eu-ai-act-voice-watermarking . Covers compliance requirements for TTS providers and what the implementation looks like in practice. For prototyping and evaluation, yes. The architecture is solid. For production at scale, the integration ecosystem is still smaller than Pipecat or LiveKit Agents. Next release should close the gap. SWE-bench measures how well models resolve real GitHub issues by editing code. The jump from 49% to 95% in one year is the fastest improvement I have ever seen on any AI benchmark. Context window growth is the bigger story underneath it. At https://lukeocodes.dev/ai-industry-this-week-new-browsers-new-standards https://lukeocodes.dev/ai-industry-this-week-new-browsers-new-standards . Kitesurf, Agent Plugins, OpenAI Astra, ByteDance, Anthropic hardware.