Google's Flash TTS Models Let You Design an AI Voice From a Text Description — Here's What That Changes
Agentic AIEvery major AI voice product has required you to pick from a library of pre-built personas. Flash TTS flips that: you describe the voice you want in text, and the model generates it. That’s a different creative and product interface than any TTS system has offered before.
The Detail
Google launched two models on September 23: Gemini 3.8 Flash TTS and Flash-Lite TTS. Flash TTS targets creative use cases (podcasts, audiobooks, narrative content); Flash-Lite handles high-volume, cost-sensitive applications like dubbing pipelines and voice agents.
Voice design from text: You write a prompt describing accent, role, and vocal characteristics — “a warm, mid-Atlantic accent with slight gravel, like a veteran radio journalist” — and the model generates a voice that matches. The library of 2,000+ preset voices covers regional variants (Mexican Spanish, Quebec French, Scottish English) for teams that want consistency without custom design.
Voice cloning is available separately: 30 seconds of audio creates a voice profile. The person being cloned must record a consent statement that matches the audio sample — an explicit consent mechanism baked into the workflow.
Additional capabilities: 100+ languages, stage directions on individual dialogue lines, two-voice dialogue generation, nonverbal sounds (laughter, sighs, breath), and minimal speaker drift across hours of audio.
Pricing: Flash TTS $0.50/M text tokens + $9.00/M audio tokens; Flash-Lite $0.50/M text tokens + $6.00/M audio tokens. One hour of audio equals approximately 90,000 tokens, making Flash TTS roughly $0.81 per hour of generated audio at scale. Available through Gemini API, Google AI Studio, Vercel, and LiveKit.
The So What
Two things are worth separating here: the voice-design-from-text capability and the pricing.
The pricing is straightforwardly competitive with ElevenLabs and OpenAI TTS. What’s new is the text-prompt voice design interface. Building a custom brand voice for an AI product previously required recording sessions, vendor contracts, or selecting the least-bad option from a voice library. Text-prompt design means a developer can iterate on a voice persona in the same workflow as any other prompt engineering.
The practical constraint is that consistent voice identity across a product’s lifetime requires either locking to a preset or carefully preserving the design prompt — text-described voices will drift if prompts change. The consent architecture for voice cloning is also meaningfully different from competitors that accept any uploaded audio sample. Whether that’s a limitation or a feature depends on your use case.
For teams building voice-forward AI products: Flash TTS’s entry into the Vercel and LiveKit ecosystems is the distribution story. This is positioned to become the default TTS option for teams already building in those stacks, which covers a significant share of the current voice agent development market.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →