Google Releases Gemini Speech Models With Custom Voices and Line-by-Line Direction

A script for an AI voice no longer has to specify only the words. Google released two Gemini text-to-speech models on September 23 that let users direct how individual lines sound, including their pacing and emotion. One also creates voices from descriptions or copies a permitted voice from a recording, putting the consent check at the center of a more flexible production tool.

First choose a voice, then direct the scene

Gemini 3.8 Flash TTS handles the most detailed voice design. A user can describe a character’s role, accent and vocal qualities in ordinary language instead of selecting only from preset voices. Google says that design feature works across more than 100 languages and dialects. It also offers a library of more than 2,000 ready-made voices, including regional varieties such as Quebec French and Mexican Spanish.

Flash TTS can also make a reusable vocal profile from a 30-second sample of the user’s voice or one they have the right to use. Google says creators can save custom voices for later work. Voice remixing—adjusting a library voice’s pitch, pace, accent or other qualities—is planned, not part of the launch.

Both new models accept directions for individual lines. A writer can specify delivery through stage directions or script cues, then add laughter, sighs or brief listening sounds. A single script can stage a two-speaker exchange with separate voices and conversational turns. Google also says the models can sustain character voices across hours of audio with little drift, a useful claim for audiobooks and podcasts that has not been independently tested here.

Where the voices are rolling out

The two versions are rolling out to developers through the Gemini API and Google AI Studio. Studio’s new audio workspace lets developers prompt a voice, use voice replication and direct a two-speaker scene. Outside those developer tools, Google is putting Flash TTS in Gemini Notebook and Flash-Lite TTS in Google Vids. API access through Gemini Enterprise is coming later.

Google describes Flash-Lite as the option for producing speech at scale, including dubbing and voice agents. It has not supplied a price comparison in this announcement. For teams choosing between the models, the stated split is therefore about intended workload, not a quantified cost advantage.

A consent check for voice copying

To replicate a voice, users must provide a verbal consent recording from the voice owner that matches the reference speaker, Google says. The company also says every clip generated by its Gemini Audio models receives an imperceptible SynthID watermark. It lists C2PA credentials among the protections for replicated voices. Those measures address two different points: permission to create a voice and a way to identify generated audio afterward.

The permission step matters because this is not limited to inventing fictional characters. Flash TTS can recreate a recognizable vocal profile from a short recording. Google describes the check, but its announcement does not measure how well it resists an attempt to misuse someone else’s voice.

What the quality claims show

Google says Flash TTS placed first on Hume AI’s Voice Design Benchmark, scoring 71.4 overall and 60.8 for accent modeling. It says Flash TTS and Flash-Lite TTS took the top two places on Hume AI’s Overall Quality Index. Those are company-reported results, rather than proof that either model will perform equally well on every script, speaker or language.

The more immediate test is the workflow Google has opened to developers: whether a designed voice stays consistent when a creator revises a long script, directs a conversation line by line and returns to that voice later. Google claims its tools can handle each part. The rollout now puts that claim in developers’ hands.