Gemini 3.8 Flash TTS Is Free in AI Studio: Design AI Voices From a Prompt (Plus the Open-Weight Alternative)

Google shipped Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS on September 23, and the headline feature is one creators have wanted for years: type a description like “gravel-voiced Scottish sea captain, tired but kind” and get a brand-new voice back, then direct it line by line with stage directions, sighs and laughs. It works across 100+ languages and dialects, and it sits on top of a 2,000-voice library.

The part that matters for a budget: both models are free of charge on the Gemini API free tier and in the Google AI Studio speech playground. There are no open weights, so if you need something that runs on your own GPU, we cover the closest open alternative, Qwen3-TTS VoiceDesign, further down.

At a Glance

• What it is: Google’s new text-to-speech pair. Flash TTS is for character voices and creative direction; Flash-Lite TTS is the cheaper high-volume model for dubbing and voice agents.
• What’s free: Input and output are “free of charge” on the Gemini API free tier, and you can test both in AI Studio with a Google account.
• What it costs after that: Flash TTS is about $0.00225 per 10 seconds of audio (roughly $0.81 per hour) through Dec 31, 2026, then double. Flash-Lite is $0.0015 per 10 seconds.
• New tricks: prompt-based voice design, 30-second voice replication with consent checks, native two-speaker scenes, and inline tags like <laughs> and |mhm|.
• The catch: API-only, rate-limited, free-tier prompts can be used to improve Google’s products, and every clip carries a SynthID watermark.

What Gemini 3.8 Flash TTS Actually Adds

The previous Gemini TTS gave you 30 preset voices. According to Google’s announcement, the 3.8 generation replaces that with a full voice system:

• Generative voice design: describe role, accent and vocal character in plain English. Google’s demos include a Melbourne radio DJ, a tinny monotone robot and a Japanese dragon.
• 2,000+ library voices, including regional varieties like Mexican Spanish, Quebec French and Scots English.
• Voice replication from a 30-second sample of your own voice (or one you have rights to), gated by a verbal consent recording that must match the reference speaker.
• Save and reuse: designed voices can be stored so a narrator sounds the same in episode 12 as in episode 1.
• Voice remixing (adjusting timbre, pitch or accent on a library voice) is listed as “coming soon,” so don’t plan a workflow around it yet.

Directing a Performance, Line by Line

The more useful change for podcasters and animators is direction. Both models accept stage directions inside the script, and the speech-generation docs describe a speech_metadata annotation for per-turn style, accent, pace and tone. On top of that you get inline vocal bursts such as <laughs>, <sigh> and <gasp>, plus backchannel cues like |mhm| and |yeah| for reaction beats and comic timing.

Native two-speaker staging means one script can drive a whole interview or sketch with two separated voices. Google also claims voice quality and timbre hold across “hours” of continuous audio, which matters for audiobooks. We haven’t stress-tested a multi-hour render, so treat that as a vendor claim for now. Google’s own voice-design walkthrough video is the quickest way to hear what the tags do.

Is the #1 Benchmark Claim Real?

Google reports that Flash TTS ranks #1 on Hume AI’s Voice Design Benchmark with a score of 71.4, leads accent modeling at 60.8, and that Flash and Flash-Lite hold the #1 and #2 spots on Hume’s Overall Quality Index. It also cites top positions in blind preference tests on Voice Arena for Japanese, Brazilian Portuguese, Vietnamese, Arabic, Mexican Spanish and Hindi.

Two honest notes. Hume is a third party, which is better than a self-graded chart, but the numbers above are relayed by Google, not measured by us. And notice the Voice Arena list is specific languages, not an across-the-board English win. If your work is mostly English narration, run your own A/B against whatever you use today before switching.

Gemini 3.8 TTS vs. the Alternatives

Try It Free: Gemini 3.8 Flash TTS in Three Steps

Option 1, no code. Open the AI Studio speech playground, sign in with a Google account, accept the terms, and pick gemini-3.8-flash-tts. You can prompt a new voice, then drop it into the two-speaker screenplay editor. Nothing to install; any browser works.

Option 2, Python. Grab a free key from AI Studio’s API key page, export it as GEMINI_API_KEY, then run pip install -U google-genai. The minimal call, adapted from the official docs:

import base64
from google import genai

client = genai.Client()
interaction = client.interactions.create(
    model="gemini-3.8-flash-tts",
    input=[{"type": "user_input", "content": [{
        "type": "text",
        "text": "Welcome back to the show. <laughs> You will not believe this week.",
        "annotations": [{"type": "speech_metadata", "style": "warm late-night radio host"}],
    }]}],
    response_format={"type": "audio"},
    generation_config={"speech_config": [{"voice": "Kore"}]},
)
with open("out.wav", "wb") as f:
    f.write(base64.b64decode(interaction.output_audio.data))

Swap the model string to gemini-3.8-flash-lite-tts for the cheaper tier. Output is 24 kHz audio in a WAV wrapper.

Option 3, fully local and open. If you want no API, no watermark and no data sharing, Qwen3-TTS-12Hz-1.7B-VoiceDesign does prompt-based voice design under Apache 2.0. Try it first in the free Qwen3-TTS Hugging Face Space. To run it yourself, per the official GitHub repo:

conda create -n qwen3-tts python=3.12 -y
conda activate qwen3-tts
pip install -U qwen-tts
pip install -U flash-attn --no-build-isolation   # optional, lowers VRAM
qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign --ip 0.0.0.0 --port 8000

The download is 4.52 GB (a 3.83 GB model plus a 0.68 GB speech tokenizer). It is a 1.7B-parameter model in BF16 and wants a CUDA GPU; Qwen doesn’t publish a minimum, but an 8 GB card is a sensible floor. It covers 10 languages versus Gemini’s 100+. We covered other open voice options in our Tencent AuK breakdown and our look at Breeze TTS 2.

Honest Trade-Offs

• Free-tier data use. Google’s pricing page marks free-tier usage as “used to improve our products.” Don’t paste unreleased scripts or client material into the free tier.
• Rate limits. The free tier is throttled. Fine for testing and short pieces; a full audiobook will hit the ceiling.
• Prices double on January 1, 2027. Flash TTS goes from $9 to $18 per million audio tokens. Budget with the 2027 number.
• Watermarking. Every clip carries SynthID; replicated voices also carry C2PA credentials. That’s good for trust, but know it’s there.
• Region limits. Voice replication in AI Studio is unavailable in several regions, including the UK, EEA and India.
• Multi-speaker limits. One request handles up to two prebuilt voices. For scenes with your own designed or cloned voices, the docs say to render each turn separately and stitch the PCM yourself.
• Voice storage expires. Saved voices cap at 200 per project with a one-year lifetime; unsaved voice keys expire after seven days.

Workflow Tips

• Write voice prompts like casting notes. Age, accent, energy and one reference texture (“slightly husky,” “smiles while talking”) beat a list of adjectives.
• Save every voice you like immediately with storage on, or it disappears in seven days and you lose continuity.
• Use Flash-Lite for drafts and Flash for finals. Same direction tags, one-third less per second of audio.
• Keep direction tags sparse. One <laughs> per beat reads natural; three in a line reads like a laugh track.
• Prototype characters in the free tier, then decide. If a voice needs to live on your own machine, rebuild it with Qwen3-TTS VoiceDesign. Our Gradium Voice Design guide is another hosted option to compare.

If you are cloning a client’s voice rather than designing one, our video How to Clone Your Client’s Voice in Under 5 Minutes with ElevenLabs walks through the consent and setup side. More tool breakdowns land weekly on the Creative AI Show YouTube channel.

FAQ

Is Gemini 3.8 Flash TTS free?

Yes, within limits. Input and output are free of charge on the Gemini API free tier and in Google AI Studio, subject to rate limits. Paid usage costs about $0.00225 per 10 seconds of audio through December 31, 2026.

Can I use Gemini 3.8 Flash TTS audio commercially?

Google positions it for podcasts, games, audiobooks and dubbing, and partners like HeyGen and Wondercraft are integrating it. Check the Gemini API terms for your use case, and remember free-tier prompts may be used to improve Google’s products. Clips carry a SynthID watermark.

What are the system requirements for Gemini 3.8 Flash TTS?

None locally. It runs in Google’s cloud, so a browser (AI Studio) or Python with the google-genai package is enough. The open alternative, Qwen3-TTS VoiceDesign, needs a 4.52 GB download and a CUDA GPU.

What’s the difference between Flash TTS and Flash-Lite TTS?

Flash TTS adds prompt-based voice design and deeper performance control for character work. Flash-Lite TTS is tuned for high-volume, low-cost jobs like dubbing and voice agents, at $6 instead of $9 per million audio tokens.

Can I clone my own voice with Gemini TTS?

Yes, from a 30-second sample plus a verbal consent recording from the voice owner. It isn’t available in AI Studio in the UK, EEA or India.

Accurate as of September 26, 2026. Pricing, rate limits and leaderboard positions move fast.