Google: Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS and its cheaper sibling Gemini 3.8 Flash-Lite TTS were announced on 2026-09-23. Google frames the pair as a shift from preset voices to voice design: you describe a character in natural language instead of picking a stock speaker, then direct the performance line by line. Both models are generally available in the Gemini API and Google AI Studio, with Flash TTS also reaching Gemini Notebook and Flash-Lite TTS reaching Google Vids.

Model Specifications

Google's model card notes both models are built on Gemini 3 Pro and list a January 2025 knowledge cutoff, which it flags as a known limitation alongside possible hallucinations and occasional timeouts.

Key Capabilities

• Generative voice design: build voices from scratch by prompting for role, accent, age, pitch, and texture. Google's blog cites a fire-breathing dragon and a Melbourne DJ as examples, and the site claims coverage across more than 100 languages and dialects.
• Voice replication from a sample: recreate a vocal profile from a 30-second consenting sample. The workflow requires a verbal consent recording from the voice owner that matches the reference speaker.
• Line-by-line direction: stage directions, pacing, emotion, and accent shifts are scripted rather than baked in. Flash TTS also supports native two-speaker staging from a single script.
• Vocal bursts and backchanneling: inline tags such as <laughs>, <sigh>, and <gasp> plus interjections like "mhm" for conversational texture.
• Long-form stability: Google claims steady timbre and pacing across hours of audio with minimal speaker drift, which is what makes audiobooks and podcasts viable.

Flash-Lite TTS trades the deepest creative controls for throughput. Google positions it for high-volume dubbing, large content libraries, and high-traffic voice agents, not for designing a character voice from nothing.

Benchmarks

Google's own evaluation sheet, published 2026-09-23, reports these numbers. They are vendor-run evaluations using production checkpoints, so treat them as vendor claims rather than independent measurements.

The pattern in Google's own data is that Flash TTS wins on voice design and on non-English locales such as Japanese, Vietnamese, Arabic, and Hindi, while Flash-Lite TTS often posts higher English and Brazilian Portuguese Elo. That is a useful signal: pick Flash-Lite when you need cheap English speech, and Flash TTS when the voice itself or a non-English market is the point.

Safety and Watermarking

Every clip produced by the Gemini Audio models carries a SynthID watermark, and Google says the replication path requires explicit consent verification. Replicated voices also carry C2PA credentials. If you are shipping synthetic voice into a product that faces regulators or platform review, the consent gate is the feature to test first, because it sits directly in your onboarding flow rather than behind an API flag.

Pricing

Google had not published a first-party price for either model on the speech generation page when this entry was written on 2026-09-24, so no dollar figure is quoted here. Confirm current rates on the Gemini API pricing page before you budget, and note that the pages with the model definitions were not reachable from this network at the time of writing.

Getting Started

1. Open the speech generation playground in Google AI Studio from the product page and confirm which model id your project gets.
2. Prompt a voice with role, accent, and texture in one sentence, then direct delivery with a second line of stage directions.
3. If you need replication instead of design, prepare consent audio for the speaker before you start, not after.
4. Test your target language explicitly. Japanese, Vietnamese, Arabic, and Hindi are the markets Google's own numbers support best.

Limitations

• 8K input tokens: long scripts must be chunked, which is the usual source of drift in audiobook pipelines.
• Vendor benchmarks only. Hume and Voice Arena numbers here come from Google's evaluation sheet, and the separate pages describing them were not directly reachable from this network.
• No published price. Do not plan a cost model until you check first-party rates.
• English parity is not a given. Flash-Lite posts the higher English Elo in Google's table, so the "bigger" model is not automatically the better pick for a US English product.

FAQ

How is this different from Gemini 3.1 Flash TTS?

Gemini 3.1 Flash TTS is the previous audio sibling. Google's own Voice Arena table puts 3.8 Flash TTS ahead in every language it lists, including Japanese and Hindi, and the 3.8 generation adds generative voice design.

Can I clone any voice?

Google's published workflow requires a verbal consent recording from the voice owner matching the reference speaker. Plan for that gate if you are building on the replication path.

Is there a cheaper option?

Yes. Gemini 3.8 Flash-Lite TTS is the high-volume tier and is the one Google points at dubbing and podcast pipelines.

Alternatives

• Gemini 3.1 Flash TTS: the previous generation audio sibling, still listed.
• sanoTTS: open-weight route if you need to self-host speech.
• Breeze TTS 2: another multilingual TTS option covered on this site.

Conclusion

Gemini 3.8 Flash TTS is the clearest example so far of voice synthesis turning into a design tool rather than a preset picker, and Google's own numbers put it first on voice design across more than one independent-style arena. The caveats are real: the benchmarks are vendor-run, there is no published price, and the 8K input ceiling shapes how you build. If your product depends on a distinctive voice or a non-English market, this is the model to evaluate this week.