Sunday roundup: a quiet publishing week in voice AI
Two posts this week, and it was a light one on the blog. Not a light week in voice AI though.
Wednesday: What multilingual voice agents cost
This was the post I wanted to write since Flux Multilingual shipped at the end of April. The stitched stack versus unified model tradeoff is the kind of architectural question every voice agent team hits around month three of a build, and most of the answers online come from vendors with a horse in the race. I tried to lay out the actual engineering costs: latency accumulation across model handoffs, the brittleness of language detection routing, why turn detection breaks across languages. Deepgram’s numbers on end-of-turn latency under 400ms went through a few rounds of internal review before I published. I wanted to be sure they held up.
Thursday: The inequity of contributing guides
This one came from a different place. I spent years in DevEx watching the same pattern play out. Well-meaning maintainers write a CONTRIBUTING.md that documents their own internal process, and the result is a filter that excludes everyone who does not already know how the project works. The CI automation argument is the part I keep coming back to. If it can be a linter rule, it should be. Let machines gatekeep the mechanical stuff. Keep the human side a conversation, not a checklist.
What else happened that week
ElevenLabs launched Music v2 on Wednesday the 27th, a model that can switch genres mid-track. Opera to heavy metal in one prompt. I did not have time to test it that week but the technical direction is interesting: conditional generation that respects section boundaries and lets you edit one part of a song without regenerating the whole thing. That architecture has parallels to how voice agent pipelines work.
CNN published a piece on May 29 about AI voice cloning scams, reporting that Americans lost $893 million to AI related fraud last year. The number stopped me. Not because it is surprising, voice cloning tools have been accessible for a while. But because the industry still does not have a standard disclosure mechanism. Some providers watermark. Some do not. Some detect cloning attempts. Some do not. The gap between what is technically possible and what is actually deployed in production keeps getting wider.
Deepgram shipped a Self-Hosted release on May 28 with Nova-3 Medical batch model improvements (97.20% KRR on medical term recognition) and profanity filtering for multilingual models. The Medical model improvement is the sort of update that does not make headlines but changes the procurement conversation for healthcare providers who were evaluating voice AI and waiting for better domain accuracy.
Frequently asked questions
Where is the multilingual voice agents post?
At https://lukeocodes.dev/what-multilingual-voice-agents-cost. It covers the engineering tradeoffs between stitched multilingual stacks and unified models like Deepgram Flux Multilingual, including latency, code-switching, and turn detection across languages.
What did the contribution guides post say?
At https://lukeocodes.dev/the-inequity-of-contributing-guides. It argues that most CONTRIBUTING.md files act as gatekeeping documents that exclude newcomers, and that maintainers should automate mechanical rules (linting, formatting) while keeping human interaction flexible and conversational.
What was significant about ElevenLabs Music v2?
Launched May 27, 2026, with the ability to switch genres mid-track, edit individual song sections without regenerating, and handle complex vocals and arrangements. Available through ElevenCreative and ElevenMusic platforms.
How much did Americans lose to AI voice cloning scams?
$893 million in 2025 according to the FBI, covering voice cloning, AI-generated phishing emails, and romance scams. CNN reported on the trend on May 29, 2026, highlighting the gap between available detection technology and what is actually deployed.