Gemini 3.8 voice cloning: where it works and what consent covers

A voice from a 30-second clip, if it is yours to use

Google's new text-to-speech models can copy a voice from a short recording and then read any script in it. A narrated product video, an audio version of a blog post or a training module in the founder's voice no longer needs a new recording session for every script. You record once, then type.

The models are Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, released on 23 September. Text-to-speech means you paste in a script and get back a spoken audio file. Google calls the copying feature voice replication. Most people call it voice cloning. Google's announcement puts it this way:

Recreate consistent vocal profiles from just a 30-second audio sample of your voice or a voice you have the rights to use

Google attaches two conditions. The first is consent. By Google's account, cloning needs a second recording from the same person:

users must provide a verbal consent recording from the voice owner that matches the reference speaker before a voice can be created.

Google's developer documentation spells out what that means. You need two real recordings from the same adult. One is a 10 to 30 second sample of natural speech. The other is that person reading a fixed sentence out loud, which in English the docs give as:

I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model.

The second condition, by Google's account, is a hidden mark on everything the models produce. Google states:

every audio clip generated by our Gemini Audio models is watermarked with SynthID.

Google describes SynthID as an imperceptible watermark woven into the audio: listeners hear nothing, and detection software can still flag the speech as AI-generated. Google says cloned voices also carry C2PA credentials, a tag attached to the file that records where it came from.

Why this week: a low price that doubles in January, and apps with no code

The release is the cause. Two details make it matter beyond developers.

The first is cost, and it comes with a date. Google's pricing page, as read on 26 September, lists the audio from Flash TTS at $9 per million tokens through December 31, 2026, and $18 from January 1, 2027. Tokens are the small units Google counts text and audio in. The same page converts the audio rate into $0.00225 per 10 seconds now and $0.0045 from January. That is about 1.35 cents per minute of finished audio this year, and about 2.7 cents from New Year's Day. A ten-minute narrated module comes to about 13.5 cents of audio this year and 27 cents after. Flash-Lite TTS is listed at $6, then $12, which is about 0.9 cents a minute now and 1.8 cents from January. The script you send in is billed on top, at $0.50 per million tokens for both models, then $1.00.

The same table lists a free tier, free of charge, and one row that matters to anyone uploading a voice. Under "Used to improve our products", Google's page says Yes for the free tier and No for the paid tier.

The one independent measurement we found is Simon Willison's. He is a developer and writer who ran his own test on launch day, 23 September:

It took ~20 seconds to generate 1m 18s of audio using Gemini 3.8 Flash TTS (not the cheaper Flash-Lite), at a cost of 2.74 cents.

That is one clip, a scripted dialogue between two pelicans, and the cost is the one he reports. It is not a price list. Plan with Google's listed rate, including the January doubling.

The second is reach. Google's launch post lists the models going to "everyone" in two apps that need no code: Flash TTS in Gemini Notebook, and Flash-Lite TTS in Google Vids. Developers get both through the Gemini API, the connection other software uses to send Google a script and get audio back, and through Google AI Studio, Google's browser workspace for trying its models. Access for enterprises through Gemini Enterprise is listed as coming soon.

Read that list carefully. It names models, not features. None of the pages we reviewed says whether voice cloning is part of what reaches Notebook and Vids. The models are there. Cloning may not be.

As for why Google shipped cloning now, Willison has a guess, and it is his reading, not Google's:

I guess voice cloning is widely enough available now from other providers that Google are no longer hesitant to ship it.

Google's safeguards, and the objections on record

Google's case is the one above: a consent recording that must match the voice sample, a watermark on every clip, and a regional limit. On that limit, a footnote on Google's announcement says voice replication through AI Studio is not available in Illinois, Texas, the European Economic Area (the EU plus Norway, Iceland and Liechtenstein), the UK, Switzerland or India. Google does not say why. The footnote covers AI Studio. The pages we reviewed do not say whether the same limit applies through the API or in the consumer apps.

The criticism is thin, and it is worth being plain about how thin. As of 26 September, our search of The Verge, TechCrunch and Ars Technica found no critical piece about this release. We also found no statement on it from SAG-AFTRA, Equity or NAVA, three organisations that represent voice actors. The objections we found come from four pseudonymous users in one thread on Hacker News, a discussion forum popular with software developers.

The first is about the watermark. A user called gruez suggested a workaround. Generate speech in a generic voice with Google's model, then turn it into the target voice with a cloning tool running on your own computer:

That'll probably get rid of any C2PA watermarks too.

That is a prediction, hedged with "probably", and it names the C2PA tag, not SynthID. Nobody in our sources tested it. The point underneath it still holds, and Willison's comment above supports it: cloning tools exist outside Google. As Google describes it, the consent check governs cloning done through Google. It cannot stop someone cloning a voice with another tool, and a watermark only helps if it is still in the audio when someone checks for it.

Google's pages also leave one question open. They say the consent recording must match the sample. They do not say whether the check can tell a real consent recording from a synthetic one. A user called kmoser asked exactly that:

Does it detect synthesized consent recordings?

Another user, pixl97, replied that no such check would last:

There is no world in which something like that works over a few months.

That is an assertion. The thread offers no test behind it, and Google's pages offer none against it.

The last objection is practical: can you use this at work at all? A user called rcr-anti wrote:

there's no alignment across the three platforms they have, consumer, prosumer, cloud.

The same comment says the models do not even have the same capabilities from one platform to the next. Google's own launch list has that shape: one model in Notebook, the other in Vids, enterprise access still to come, and a regional limit written for AI Studio only.

As of 26 September, none of the sources we reviewed reports an independent test of whether SynthID survives editing, compression or re-voicing, or of whether a synthetic consent clip would get through.

Voiceovers without a consent problem

Split the job in two.

Most voiceover work doesn't need anyone's voice: product explainers, audio editions of posts, internal training. Use one of Google's library voices, or, on Flash TTS, a voice designed from a written description, which Google's announcement says that model can create. Neither voice belongs to a real person. Google's consent requirement is written for cloning, and so is the regional footnote.

Clone a voice only when the voice itself is the point: the founder narrating company updates, or a presenter your audience already knows. Before you do, check three things.

Where you are. If you work in one of the listed places and planned to clone in AI Studio, the footnote says you can't. If you work through a company Google account, confirm that the product you actually have includes the feature. Google's own list puts different models in different products.

Who records the consent. It has to be the voice owner, reading Google's sentence in their own voice. Google says that recording is checked against the sample.

What Google keeps. Per Google's docs, the cloned voice is stored by default in your Google project, the account space where Google keeps your API settings and usage.

The docs list a retention of one year for stored voices. The alternative is a scrambled voice key that your own software keeps instead of Google, and for that the docs list seven days. The documentation we read does not say how long Google keeps the consent recording. If you are asking a colleague to lend their voice, that is worth knowing before they record.

Some of this work repeats, like turning every new blog post or changelog entry into an audio clip. That is a small automation, not a weekly chore. In Latenode, the heart of it is one short code step. A JavaScript step there can install any NPM library, the ready-made code packages developers share, so it can call the Gemini API with your own Google API key, the password-like code that tells Google whose account to bill. It sends the post's text and asks for a library voice or one designed from a description, never a cloned one. The same step can then hand the finished file to your own storage or website through that service's NPM library, if it publishes one. Leaving cloning out is deliberate. Cloning needs the voice owner's recorded consent and is limited by region, and neither belongs in a job that runs unattended. The two bills stay separate. Google charges for the audio on your Google account, at the rates on Google's pricing page, which double on January 1, 2027. Latenode meters each run by the CPU-seconds it consumes, the computing time it uses, so a shorter run costs proportionally less. There is no per-run minimum charge, and Latenode's free tier includes 10,000 CPU seconds a month.

Use the voices nobody owns, clone only with the owner in the room

Use Gemini 3.8 TTS for voiceovers now if your scripts change often and a library or designed voice will do: explainers, audio editions, training. By Google's listed rate, audio from Flash costs about 1.35 cents a minute until the end of 2026 and about 2.7 cents from January 1, 2027, and nobody's rights are involved.

Clone a voice only if it is yours, or belongs to someone who will record Google's consent sentence themselves. Before you promise a client cloned narration, check that the route you will actually use works where you are. If you are in the EU, the UK or another listed place, assume cloning in AI Studio is closed to you. For Notebook and Vids, look in the product before you plan around it, because Google's announcement doesn't say whether cloning is there.

Do not count on the safeguards to protect the people whose voices are at stake. By Google's account, the consent check stops a stranger's voice being cloned inside Google's own products, and that is worth something. Past that point, what remains is the watermark. As of 26 September, none of the sources we reviewed reports a test of whether it survives editing, and other cloning tools already exist that Google's check does not reach. If your listeners should know a voice is synthetic, the only reliable way is to tell them yourself.