A studio you type at, and glasses with no camera
Edition of 2026-09-24
# AI For All ### Thursday, September 24, 2026
As of today, anyone who makes audio for a living can describe a voice in one sentence and get it, or reproduce a real human voice from half a minute of recording, through a product that is already live on Google's developer platform.
The preset menu is gone
Google DeepMind published two text to speech models on Wednesday, and they are live today inside Google AI Studio, the Gemini API, Gemini Enterprise, Gemini Notebook and Google Vids. The change is not a better voice. It is the end of the list. Instead of choosing from a fixed set, you describe what you want, role, accent, character, and the tool builds it. Google says voice design reaches more than *100 languages and dialects, and that the ready library now holds 2,000+ production-ready voices*.
The detail that matters if you work with performers is copying. Google says a voice can be replicated from a *30-second audio sample*, and it attaches consent verification, an invisible watermark and a signed provenance record to the output. That is a direct answer to the obvious objection, and it is what separates a studio tool from a liability. The cheaper of the two models is built for high volume dubbing and voice agents, where the cost of each clip decides the purchase.
Read the announcement at Google DeepMind.
The cheapest way in is the pair that cannot film you
Meta's first camera free AI glasses went on pre order on Wednesday at *$349 USD, shipping in October. Ray-Ban Meta Audio weighs 43 grams, the lightest in the range, runs for 12 hours* on one charge, and keeps the parts people actually use: music, calls, live translation, and an assistant. Meta brought Muse, the personal agent it launched this month, to the glasses, and said the new frames reach new markets starting today, among them Singapore, South Korea, Mexico and the United Arab Emirates.
The camera equipped models stay on sale, so the interesting question is not technical. It is whether a large share of buyers wanted the assistant all along and simply would not wear a lens in public. Meta is betting that they did.
Read the announcement at Meta.
A model found a tool that humans had already walked past
Anthropic says its agents found a new biological system with nothing more than a prompt from its scientists. Roughly *950 agents searched a large public database of DNA sequences for 21 hours* before one of them noticed a repeating pattern sitting next to an odd looking gene, a pattern earlier studies had passed over. The laboratory then confirmed it at the bench.
Anthropic says the system carries properties that have only ever appeared together in systems which are programmable and can cut, copy and paste DNA, the family that CRISPR belongs to. Feng Zhang, one of the pioneers of CRISPR genome editing, reviewed the preprint and called the identification genuinely intriguing and worth further investigation. The caution is honest and it is the point: nobody yet knows what the system does.
Read the announcement at Anthropic.
The platform that hosts the video is now the editor
At its Made on YouTube event, YouTube announced a conversational editing assistant for Shorts and the YouTube Create app, built on Gemini, that takes plain language instructions to reorder frames, trim footage and sync music to the beat. It also put a number on how much of this decision making it already runs: title and thumbnail tests have passed *40 million experiments* since the feature launched in 2024.
Its own survey puts the share of American video posters who have already used AI to create or edit at *72%*. That is the audience, not a sample of enthusiasts. Most of the new tools are still in testing, so the honest reading is direction rather than something you can use this morning.
Read the announcement at YouTube.
Weak signal
Nvidia published a small open model that identifies who spoke when in a recording, following several speakers at once. It sits on a public model hub and anyone can download it. No headline number carries it. What it shows is structural: the routine parts of speech work are becoming free, private, and small enough to run on hardware a business already owns.
The question for the next few weeks is not which synthetic voice sounds best. It is what a recording still proves, now that a convincing one takes half a minute.