Google released Gemini 3.8 Flash TTS and Flash-Lite TTS on 23 September for directed voice generation, multilingual speech and high-volume applications.

Google added two text-to-speech models to the Gemini family. Gemini 3.8 Flash TTS is designed for detailed creative direction and voice design, while Gemini 3.8 Flash-Lite TTS targets larger-volume, lower-cost workloads. Both are rolling out through the Gemini API and Google AI Studio, with broader product availability varying by model.

Why it matters

Text-to-speech systems increasingly generate performances rather than simply read text aloud. Google says developers can specify pacing, emotion, accents, conversational sounds and line-by-line acting cues. That changes the product decision from choosing a fixed synthetic voice to directing a reusable vocal identity for an audiobook, game, dubbing workflow or voice agent.

The models support more than 100 languages and dialects. Gemini 3.8 Flash TTS can create a voice from a written description or reproduce a permitted voice from a 30-second sample. Google says voice replication requires a matching consent recording from the speaker. The company also applies SynthID watermarking and C2PA credentials to generated audio.

Those safeguards matter because a short sample can make voice replication useful and dangerous. A legitimate team can preserve a narrator’s sound across a production. An impersonator can use the same capability for fraud or harassment. Consent checks reduce casual abuse, but their effectiveness depends on identity matching, account controls and whether the watermark survives editing or re-encoding.

Google reports that Flash TTS ranked first on Hume AI’s Voice Design Benchmark and that both new models placed at the top of Hume’s quality index. It also cites blind preference results from Voice Arena across several languages. The rankings are useful launch evidence, but they do not settle performance for every accent, speaking style or long-form project.

Availability and limits

Developers can begin testing both models through Google AI Studio and the Gemini API. Google says Flash TTS is also rolling out in Gemini Notebook, while Flash-Lite TTS is reaching Google Vids. Enterprise API access is described as coming soon rather than immediately available everywhere.

Voice replication has regional restrictions. Google’s launch note says the feature is not available through AI Studio in the European Economic Area, the United Kingdom, Switzerland, India, Illinois or Texas. That limits a prominent capability for many developers and reflects the legal sensitivity around biometric voice data.

Long-form quality needs practical testing. Google claims the models can maintain voice identity and pacing across hours of audio, but creators should test whole chapters rather than short samples. Drift, pronunciation, emotion consistency and editing time can determine whether a nominally strong model reduces production work.

Cost is another missing comparison in the launch post. Flash-Lite is described as cost-efficient, but the article does not state a per-character or per-minute price. Builders should compare the total cost of accepted audio, including regeneration and human review, once pricing and service limits are clear.

The next useful evidence will be independent multilingual listening tests and production measurements across long scripts. The release expands Google’s audio tools now; it does not yet prove that one model is best for every voice product.

Verification

  1. VERIFIED — Google announced Gemini 3.8 Flash TTS and Flash-Lite TTS on 23 September 2026. Primary source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/
  2. VERIFIED — Google says the models support more than 100 languages and dialects and line-by-line direction. Primary source: the Google launch post above.
  3. VERIFIED — The launch describes consent verification, SynthID watermarking and C2PA credentials for voice replication. Primary source: the Google launch post above.
  4. VERIFIED AS REPORTED EVALUATION — Benchmark and preference rankings are cited by Google and were not independently reproduced here. Primary source: the Google launch post above.
  5. VERIFIED — Google lists regional limits for AI Studio voice replication. Primary source: the Google launch post above.

Glossary candidates

  • Text-to-speech: Technology that converts written text into spoken audio.
  • Voice replication: Generating speech that preserves the characteristics of a reference speaker.
  • Watermarking: Embedding a detectable signal that identifies generated content.

Cold-reader sentence: Google released two multilingual Gemini speech models with directed performance and consent-checked voice replication, while pricing and independent long-form tests remain outstanding.