Most "AI voice" tools still work the same way they did five years ago: pick a voice from a dropdown, adjust a speed slider, hope it doesn't sound robotic. On September 23, Google shipped something that breaks that pattern. Gemini 3.8 Flash TTS and its cheaper sibling, Flash-Lite TTS, let you describe a voice in plain language — "a gravelly Melbourne DJ," "a monotone customer-service robot," "a Japanese dragon with a rasp" — and get a usable, original voice back. For agencies building anything that talks to a client's customers, that's a meaningfully different starting point than picking voice #14 from a menu.
The headline feature is generative voice design: instead of choosing from a fixed voice library, you write a natural-language description and the model generates a matching voice on the spot, across more than 100 languages and dialects. Google says the library has grown from 30 preset voices to over 2,000 production-ready ones, and custom-designed voices can be saved and reused with minimal drift across a project — useful if you're building a recurring character for a brand, not just a one-off clip.
The second big change is performance direction. You can script vocal bursts like laughs, sighs, and gasps, insert backchanneling ("mhm," "yeah") for two-speaker scenes, and give line-by-line stage directions that control pacing, accent shifts, and tone. Long-form generation is built to hold quality across hours of continuous audio, which matters for anyone producing audiobooks or long-form dubbing rather than 15-second clips.
There's also a voice-replication mode: clone a voice from a 30-second sample, gated behind a verbal consent recording that has to match the reference speaker. Every output, cloned or generated, carries an imperceptible SynthID watermark plus C2PA content credentials, and replicated voices are further restricted in Illinois, Texas, the EEA, UK, Switzerland, and India — a sign Google is trying to get ahead of the deepfake-voice problem rather than bolt on safety after the fact.
Google split the release into two models with genuinely different use cases rather than just a "fast/cheap" naming exercise:
Flash TTS is built for creative direction — games, immersive audiobooks, podcasts, anything where a human is directing performance and wants granular control over how a line is delivered. Flash-Lite TTS is built for volume and cost: dubbing pipelines, high-throughput content generation, and — most relevant for automation builders — real-time voice agents, where you're generating speech on every turn of a conversation and unit economics matter more than nuance. Reported pricing for Flash-Lite lands around $0.50/$6 per million tokens, and both models are API-only (no open weights, no self-hosting), available now through the Gemini API and Google AI Studio, with Gemini Enterprise access coming soon.On Hume AI's independent Voice Design Benchmark, Flash TTS ranked #1 overall (71.4) and led on accent modeling, with Flash-Lite close behind on overall quality — so this isn't just a features launch, Google is also claiming the top of an external leaderboard.
If you build voice agents, IVR replacements, dubbing pipelines, or narrated content for clients, the practical shift is that voice becomes a spec you write instead of an asset you license. A client wanting "a warm, unhurried voice for our wellness app, but definitely not generic-assistant-sounding" used to mean browsing a voice marketplace or hiring a voice actor. Now it's closer to a prompt-engineering problem — which is exactly the kind of task that folds into the workflows agencies already build around LLMs.
The integration list is also a signal of where Google expects this to land first: Figma, HeyGen, Agora, LiveKit, and Pipecat are all named partners, which points squarely at real-time voice-agent and video-content tooling rather than pure entertainment use cases. If you're already building on LiveKit or Pipecat for a client's voice bot, swapping in Flash-Lite TTS is likely to be a near-term, low-friction upgrade rather than a rebuild.
The caveats worth flagging before you pitch this to a client: it's API-only with no self-hosting option, so anyone with data-residency or offline requirements is out of luck for now, and voice cloning is unavailable in several major markets due to the consent and regulatory restrictions. Budget for that when scoping international projects.
Our cutting-edge features simplify collaboration and creativity, making your workflow intuitive and efficient. Transform your vision into reality effortlessly with Hadidiz Flow.
