Google pushes voice-cloning consent onto the developer who uploads the 30-second clip

Thirty seconds is about the length of a voicemail, and it is all the audio these models need to make a replica of a real person's voice ^[9]. The rule attached to that path is a rule for the customer: secure the speaker's consent before generating the replica ^[10]. SiliconANGLE's report does not say how Google checks that the consent exists.

So the record lives in your systems, held by whoever needs a voice this week. If a narrator is cloned in March and leaves the company in September, somebody has to be able to produce the permission and say what it covered.

Google's pitch on the two models is that the APIs are highly similar, so running them side by side is relatively simple ^[2]. The difference in output is described in relative terms: Flash TTS sounds better and costs more, Flash-Lite is tuned for cost efficiency and inference speed ^[4][3]. Language coverage comes as a number. Flash speaks 130 languages at launch and Flash-Lite 101 ^[5], so 29 languages exist on the expensive model only ^[6].

For an English-and-Spanish product, the tier choice is a cost test you can run on real scripts. For a support organisation covering a long tail of locales, Flash-Lite is unavailable in those 29 markets at any price, and the routing rule has to carry the list.

The ranking comes from an evaluation Google ran itself. Google tested the two models on an audio quality benchmark developed by the startup Hume AI, where they took first and second place ^[15], and says they beat several competing models on multiple language-specific versions of Voice Arena, a benchmark scored on human feedback ^[16]. "These models enable creators, developers, and enterprises to create richer, more expressive audio experiences, while enabling improved user experiences in products like Gemini Notebook and Google Vids," Google staffers Leland Rechis and Alan Cowen wrote in a blog post ^[17].

Detection tooling can establish that a clip came from Google's models; whether the speaker agreed is a separate record. The provenance work covers the file: SynthID embeds a watermark humans cannot hear but detection tools can read ^[13], and a C2PA record travels with every generated file noting when it was made and whether it has been modified since ^[14].

What sorts this work before it ships is whether the voice belongs to an identifiable person, and whether the audio outlives the project that commissioned it. A voice built from a natural language prompt, with timbre and pacing set by parameters ^[8], needs nothing more durable than the prompt string in version control. A replica of a real employee, embedded in documentation that stays up for years, needs a consent record that outlasts that employee and a way to retire the voice if they withdraw. The customization option Google says is still to come, editing one of the 2,000-plus prepackaged voices ^[11][7], falls in the first group. The 30-second path falls in the second, whatever the size of the audio file.