You just recorded a three-minute product narration and cut it into your video, and then you notice the problem. You went through the whole built-in voice library, and not one voice sounds like the person you have in mind. Too young, too much like a news announcer, the accent is wrong.

Many people reach the same step and do the same thing: they look for voice cloning and use an existing voice as a template. What they run into next is often not a technical problem but a different one: whose voice is this, and does that person know?

In the YouTube video "Gemini 3.8 Flash TTS with Voice Cloning", Sam Witteveen covers both questions in one video. This article reorganizes the key points of the video into an English narrative and adds what you should think through before you use it.

Two New Models, Two Different Jobs

According to the video, Google released two text-to-speech models this time: Gemini 3.8 Flash TTS and Gemini 3.8 Flash Light TTS.

They are positioned differently. Flash TTS leans toward creative work, such as character voices, audiobooks, podcasts, and games. Flash Light leans toward volume, such as dubbing conversion, large amounts of speech output, and voice agents, which is to say cases where cost matters more than polishing every line.

Google calls this its most expressive speech model so far. The video author's view is that this is partly true, but the independent numbers are more mixed than the official blog presents. Keep that sentence in mind, because we will come back to it.

Instead of Picking a Voice, Describe One

The feature the video finds most worth noting is voice design. You do not pick from a fixed list. You describe the voice you want in plain language, including traits such as role, accent, and age, and the model generates it from that. The video says more than 100 languages are supported.

The author also names the spot where this kind of model most often leaves people stuck: none of the default voices fits, so you end up going to cloning. His own approach is to generate several synthetic voices from slightly different descriptions, pick the best one, and then clone that artificial voice. The benefit is that the thing being copied is not a real person at all.

The voice library also has more than 2,000 voices and a large set of regional accents, which you can adjust and save for reuse.

The video demonstrates an "astronomer" character: late sixties, with a British accent. The author's comment is a practical one: it sounds a bit like a documentary narrator, but he felt the voice was much younger than the age in the description. It is also a reminder that a description is not a guarantee, so listen to a sample.

Directing Line by Line, and Two-Person Dialogue

Once a voice is chosen, you can give a performance direction for each line: this one whispered, this one rushed, this one a little sarcastic. You can also insert non-verbal tags in the text, such as a sigh, a laugh, or an inhale.

The video tested one sentence in three styles, flat, excited, and nervous to the point of whispering, and the author felt they sounded very different. He also particularly liked that the style affects the length of the audio, rather than forcing the content into a fixed duration.

In addition, one API call can carry two speakers, each with their own style. Google claims long recordings keep a stable voice without drifting. The author says he has recently run into long text drifting off in some open-source TTS models, so this point appealed to him. His impression of the two-person demo, though, was that it sounded a lot like an auto-generated NotebookLM-style show.

Next comes the most sensitive part of the whole video. The video says the way it works is that you provide a sample of about 30 seconds and the system generates that voice.

But there is a checkpoint in the process. The person being cloned must record a consent statement, and that statement has to match the reference speaker before the system will produce the cloned voice. The video also mentions that every output carries a SynthID watermark and comes with a C2PA credential.

The video also points out that the voice cloning feature in AI Studio is unavailable in quite a few countries, and even in some US states, because of legal requirements. That shows even Google is handling this carefully. Whether it is available in Taiwan, the video does not say, and this article has not checked it, so please confirm for yourself before use.

The author's own test result was somewhat unexpected: when he cloned his own voice from 30 seconds, it sounded "quite different" to him, and he guesses a longer reference recording would work better. He also thinks the open-source Qwen TTS is still very strong at cloning, but that Gemini has the advantage when it comes to generating a voice from a description.

This section is not the core of the video, but it is the part you have to take responsibility for yourself.

The biggest risk of voice cloning is not that the sound is not similar enough, but that it is too similar. A voice can be used to impersonate someone, to commit fraud, or to speak for a person without their knowledge. Google builds "the person's own consent recording" into the technical process precisely because the consequences of this kind of misuse are serious.

If you plan to use voice cloning, do at least the following.

One, clone only your own voice, or get clear, retainable written authorization from the owner of the voice, stating the purpose, scope, and duration.

Two, do not clone a celebrity, a colleague, a client, or any other third party, even if it is "just a test".

Three, if you publish the resulting audio, label it as AI-generated and keep the authorization documents and generation records.

Four, if a voice design can give you a synthetic voice that belongs to no one, that is usually safer than cloning a real person.

The above is a general risk reminder, not legal advice. Rules on likeness, voice, and personal data differ from place to place, so consult legal counsel before commercial use.

Benchmarks: How to Read "Number One"

Back to the sentence from earlier. The video says Google's numbers mainly cite the benchmark from Hume AI, claiming that Flash TTS ranks first on Hume's voice design benchmark, especially strong on accents.

The author then points out a conflict of interest: one of the two authors of that blog post was, until January this year, the CEO and chief scientist of Hume AI, and at that time Google DeepMind licensed Hume's technology and hired a group of people from the company. The author stresses that he is not saying the numbers are wrong, only that grading yourself on a benchmark you designed is not necessarily the best way to judge.

A more independent reference is the Elo preference ranking from Artificial Analysis: Gemini 3.8 Flash TTS ranks second and Flash Light ranks sixth, with Flash sitting just ahead of Qwen's new Audio 3 TTS Plus. The video says Gemini stands out on pronunciation-related items, which matters for voice agents that have to read out numbers and ID codes. Multilingual is a Google strength, but Cartesia and ElevenLabs are also improving fast.

On price, the video's position is that Google lands in the middle: more favorable in price than some Cartesia and ElevenLabs plans, and than Minimax too, but far from the cheapest. More interesting, Flash Light is not much cheaper than Flash, which differs from the situation with the text-model versions. After listening, the author felt that Light tends to sound a bit lazier and may stretch longer, and since Flash costs about 50 percent more, unless the content is long-form he would probably pick the one with better sound quality.

Closing: Ask a Different Question

Back to that three-minute narration from the beginning.

You thought the question was "which voice sounds closest". The video itself only demonstrates the description-based design and cloning workflow; this article's suggestion is this: first use a description to design a voice that belongs to no one, and when you need to clone, clone only yourself or someone who has clearly authorized you.

The next time you are tired of scrolling through the voice library, ask yourself two questions first: whose voice is this, and did they agree? Once those two questions have answers, then talk about how close it sounds.

FAQ

Q1: What is the difference between Gemini 3.8 Flash TTS and Flash Light TTS?

According to the video, Flash leans toward creative uses such as characters, audiobooks, podcasts, and games, while Flash Light leans toward volume uses such as dubbing conversion and voice agents. The video also notes that the price gap between the two is small, with Flash about 50 percent more expensive.

Q2: What kind of sample does voice cloning need?

The video demonstrates about 30 seconds of reference audio, and requires the person being cloned to record a consent statement. The author thinks a longer sample may work better, but that is his guess, not an official guarantee.

Q3: Can I clone someone else's voice?

You should not do so casually. You need clear authorization from the person. The process in the video also requires the person to record a consent statement that is matched against the reference voice. Copying someone's voice without consent may involve impersonation and other legal problems.

Q4: Can I use voice cloning in my region?

The video says this feature in AI Studio is unavailable in quite a few countries and some US states. Whether it is available in Taiwan, this article has not checked, so please go by Google's official information.

Source Statement

This article is a rewrite and summary, not a sentence-by-sentence translation. Its content is reorganized from the YouTube video below. The video's transcript was available only as an Arabic version rather than the original English subtitles, so proper nouns such as names, companies, and model names may have been transliterated or mistranslated. Statements about people and benchmarks are the opinions of the video speaker and are attributed as such. No sponsorship disclosure appeared in the transcript. Price descriptions are only relative comparisons and no specific list price was obtained, so please check official pages. The voice cloning risk reminders are general and are not legal advice.

Video title: Gemini 3.8 Flash TTS with Voice Cloning. Channel: Sam Witteveen. Link: https://www.youtube.com/watch?v=GDUBXR-ql78. Publication date: 2026-09-24.

Further Reading

If this topic interests you, we have published related articles on this site: Gemini 3.8 Live and extended thinking, and how to choose for a voice agent, building a real-time voice AI agent with the Gemini Live API, 2026 Gemini API pricing for Flash, Pro, and Flash-Lite, and a guide to Gemini Token pricing.

Before You Pick the Next Voice

If you want to see the model list, balance, and usage from different models in one place, you can take a look at AI Token King: https://aitokenking.com.tw/home. Whether the speech models mentioned in this article can be called through that platform should be checked against its official model list.