Gemini 3.8 Flash TTS clones a voice in 30 seconds: what it can do for your business's phone line and videos

On 23 September 2026 Google released two text-to-speech models, gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts, and with one of them 30 seconds of audio is enough to create a custom voice ¹. You don't need a server with graphics cards, and you don't need a recording studio.

The question that matters to you is simple: can I use it for the phone or for the company's videos without spending a fortune? For videos and recorded messages, yes, and the cost is almost irrelevant. For a voice that talks to customers live, the available information isn't enough yet for me to tell you to go ahead.

What has actually been released, plainly

The models come with a library of more than 2,000 ready-made voices ¹. If none of them suits you, you can create your own from a 30-second sample of your voice, or of a voice you have the right to use ¹.

The most interesting part isn't the cloning. It's that you can write a whole dialogue between several characters, each with their own voice and tone instructions, and get a single audio file. For anyone who makes tutorial videos or product presentations, this removes a good chunk of work.

Simon Willison, a developer who tries out these tools as soon as they come out, has put online a free playground to experiment with them from the browser using your Gemini API key ¹. It's the quickest way to hear the result before deciding anything.

What it costs, measured on a real test

In Willison's test, a dialogue of 1 minute 18 seconds was generated in about 20 seconds, at a cost of 2.74 cents ¹. Willison used the Flash model, not the cheaper Flash-Lite ¹.

In practice: for a welcome message, an out-of-hours voicemail or the voiceover of a video a couple of minutes long, the cost of the voice is the last thing to worry about. The time you spend writing a good script is worth far more.

Where it already works: videos, messages, demos

Here the case is clear. Whenever the text is written first and the audio is listened to afterwards, a synthetic voice does its job.

Think of the phone line message that gives opening hours, the instructions for booking, the voice of a video explaining how to assemble one of your products. Today you record them once and then never touch them again, because redoing them is a hassle. With a model like this you change one line of text and regenerate the file.

Multi-voice dialogues have a practical use too: a simulated phone call to train new staff, or a sales demo where customer and salesperson really talk.

Where you're not there yet: the customer on the line

A phone line that understands the customer and answers by voice is a different matter. The published test measures how long it takes to produce a complete file, not how long passes before the first word reaches the caller's ear. On a phone call, that delay is what makes a voice sound like a person or like an answering machine.

The source also says nothing about the quality of the Italian, or about regional accents. Don't assume that a voice that is excellent in English will sound natural to a customer in Bari or Bergamo. It has to be listened to.

My firm advice: don't put a cloned voice on live customer calls just because a new model has come out. Start with recorded messages, where a mistake costs little.

A voice doesn't become yours just because you recorded it

Google itself speaks of your own voice or a voice you have the right to use ¹. If you clone an employee's voice for the phone line, get written consent and make clear what happens if that person leaves.

There is also the customer's side. Callers need to understand that they are listening to a generated voice, especially if you then ask them for data or decisions. One sentence at the start of the message solves the problem and saves you embarrassment.

What to do on Monday

Pick a single recorded message that you really use today, for example the out-of-hours one on your phone line. Write the text in Italian, the way you would say it aloud.

Open the playground, try some of the ready-made voices first and only then, if needed, record 30 seconds of the voice of someone who has given you permission ¹. Let someone in the office listen to the result without telling them it's synthetic, and watch their face.

If it sounds good, use it for that message and stop there for a few weeks. If it sounds fake, do nothing: keep the recording you have and try again at the next update. If you'd like an opinion on where a synthetic voice makes sense in your specific case, we at AIR are happy to talk it through, with no rush to sell you anything.

The voice now costs a few cents. What matters is knowing where to use it.