Intel / article

Gemini 3.8 text-to-speech says hello

Source: deepmind.google
Date: Published Sep 23, 2026

From the Daily Brief

Google released two text-to-speech models, gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. Voices are built from a natural-language prompt, and developers can direct emotion and pacing line by line or script directed scene dialogue between characters. Simon Willison's writeup lists more than 2,000 voices, custom voice cloning from a 30-second sample and multi-speaker scripting, and he published a bring-your-own-key playground for testing them. Google says output carries built-in watermarking. Distribution is wide from day one: AI Studio, the Gemini API, Enterprise, Notebook and Vids, and Vercel added both models to its AI Gateway with long-form narration, two-speaker dialogue and support for more than 100 languages. For product teams, the change is that expressive, multilingual speech now sits behind the same gateway many already use for text models, which removes one reason to sign a separate voice vendor. Cloning from a short sample also raises consent questions that teams should settle before shipping it to users.

Read the 2026-09-24 Brief →

Context

Google's Gemini 3.8 Flash TTS voice replication requires a recording of the voice owner's consent that matches the reference speaker, according to its announcement. Generated audio also carries a SynthID watermark. These serve different purposes: consent verification governs voice creation, while watermarking supports identifying the resulting audio as AI-generated.

Key takeaways · AI-distilled

Google released two TTS models: Gemini 3.8 Flash TTS for creative direction and custom voice design, and Flash-Lite TTS for high-volume, lower-cost uses like dubbing and voice agents.
Flash TTS can design new voices from natural-language prompts across 100+ languages, and replicate a voice from a 30-second sample, but only after a verbal consent recording from the matching speaker.
Delivery is scriptable line by line, including native two-speaker scenes and inline cues such as <laughs> or |mhm| for vocal bursts and backchanneling; Google says voice quality holds across hours of long-form audio.
Google cites Hume AI benchmarks: Flash TTS ranks #1 on the Voice Design (71.4) and the two models take #1 and #2 on Hume's Overall Quality Index. All output is watermarked with SynthID.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source deepmind.google