Google has released Gemini 3.8 TTS, a text-to-speech family that can create a fictional voice from a written description, direct delivery line by line, and run two-speaker scenes. Its more constrained voice-replication path requires both a 10–30 second voice sample and a matching adult consent recording, making the distinction between invented characters and recognizable people part of the product.
Key facts
- Gemini 3.8 Flash TTS supports 130 languages; Flash-Lite supports 101.
- Voice replication requires a 10–30 second recording plus a spoken consent recording.
- Google says all Gemini Audio clips carry SynthID watermarking.
- Primary source: Google's Gemini 3.8 TTS launch post.
Google's product has two easily confused paths. Voice Design lets a developer type a brief such as an older, warm astronomer with a particular accent, then receives a new voice identity and preview. Voice Replication is for a particular person. The owner supplies the target audio and a distinct English declaration: “I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model.” Google says users must own the voice or have the rights to use it.
The delivery controls push the service beyond a dropdown of preset voices. A script is treated as a verbatim transcript, while structured metadata carries sustained style, a speaker label, and short acting events such as <laugh>, <sigh> or <short pause>. Two configured speakers can alternate in one request. Think of it as handing a synthetic cast a marked-up director's script rather than asking a single narrator to sound “happy.” The speech-generation guide describes those controls; this endpoint is still text-to-audio, not a full duplex conversational model.
The cost anchors the commercial story. Google's pricing page defines one second of output as 25 audio tokens. Through 31 December 2026, Flash output is $9 per million audio tokens and Flash-Lite is $6; that is about $0.81 and $0.54 for an hour of output at standard rates, before input. Google describes Flash-Lite as a low-latency option, but it has not published a primary-source median time-to-first-audio, so a sub-second performance claim would be speculation.
Google says every clip is marked with SynthID, and the Gemini Audio page says replicated voices receive C2PA credentials. Those are useful provenance measures, not a complete solution to impersonation. A recording can be re-recorded, converted or cloned locally; consent verification also cannot resolve every jurisdiction's right-of-publicity or contract question. The model stores verified voices for one year by default, while a non-stored option returns a client-managed key that expires after seven days.
The strongest positive case is that expressive, low-cost voice production becomes practical for narration, dubbing, accessibility and agent output. The strongest caveat is quality evidence: Google's benchmark claims are vendor claims, and public leaderboards did not independently display the cited new entries in this research pass. Use the launch as a capability and distribution event, not proof that Gemini is objectively the best voice in every language.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)