Google shipped two text-to-speech models on September 23 that reframe what a voice API can do. Gemini 3.8 TTS — available today as gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts — brings over 2,000 production-ready voices, a voice design tool that generates custom voices from plain English, and voice replication from a 30-second audio sample to the Gemini API. If your mental model of TTS is still “pick from a dropdown of 30 options,” this changes that significantly.
Two Models, Two Jobs #
These are not the same model at different price points. They solve different problems, and choosing correctly matters for both quality and cost.
Gemini 3.8 Flash TTS is built for quality. It supports 130 languages, handles complex multi-character dialogue, and maintains voice consistency across long-form audio without drifting — the kind of stability you need for an audiobook or a game with hours of NPC dialogue. Use it when the audio output represents your product directly.
Gemini 3.8 Flash-Lite TTS is built for throughput. It covers 101 languages and is optimized for high-volume workloads — voice agents, dubbing pipelines, read-aloud features. Use it when cost-per-call matters more than maximum fidelity. Both are available now in Google AI Studio and the Gemini API.
What Actually Changed with Gemini 3.8 TTS #
Previous TTS APIs gave you a list of voices. Gemini 3.8 TTS gives you a voice studio, and the difference is substantial for production use cases.
Voice design lets you describe a voice in natural language and get back a persistent custom voice ID. “Warm baritone with a slight British accent” becomes a voice your app uses consistently across every request. You can store up to 200 custom voices per project — though voice IDs expire after a year, so archive your design prompts.
Voice replication reconstructs a vocal identity from a 30-second audio sample. This is the feature studios have been waiting for: consistent character voices without recording sessions for every script change. However, Google requires a verbal consent recording and restricts this feature in the UK, EEA, Switzerland, Illinois, and Texas, so verify your deployment region before building around it. All generated audio receives SynthID watermarking for AI provenance transparency.
Getting Started: API Quickstart #
The API uses a new Interactions endpoint at https://generativelanguage.googleapis.com/v1beta/interactions. Here is a minimal working example in Python:
from google import genai
import base64
client = genai.Client()
interaction = client.interactions.create(
model="gemini-3.8-flash-tts",
input=[{
"type": "user_input",
"content": [{
"type": "text",
"text": "Have a wonderful day!",
"annotations": [{
"type": "speech_metadata",
"style": "cheerful and friendly"
}]
}]
}],
response_format={"type": "audio"},
generation_config={"speech_config": [{"voice": "Kore"}]}
)
with open("out.wav", "wb") as f:
f.write(base64.b64decode(interaction.output_audio.data))
The speech_metadata annotation is where you direct performance — tone, pacing, emotion. You can also embed inline vocal tags in your text: <sigh>, <laugh>, <>, <breath>. For multi-speaker scenes, the API supports up to two speakers per request using prebuilt voices; anything larger requires stitching. Simon Willison benchmarked the API at roughly 2.74 cents to generate 78 seconds of audio — fast and cheap enough for production workloads.
The ElevenLabs Assumption Is Outdated #
The default assumption in developer circles has been that ElevenLabs is the quality choice and everything else is budget TTS. That assumption no longer holds for most workloads.
On the Artificial Analysis Voice Arena leaderboard, Gemini 3.8 Flash TTS ranks second with an Elo of 1,260. ElevenLabs Eleven v3 sits at rank 17 with an Elo of 1,167 — a 93-point gap — at roughly three times the price ($100 per million characters versus Google’s ~$33). ElevenLabs still holds real advantages: sub-100ms latency tiers and a mature voice ecosystem for products already built around specific voice identities. However, for new projects starting from scratch, the math has shifted clearly toward Gemini.
Pricing: The January 2027 Cliff #
Current introductory rates run through December 31, 2026. After that, prices double.
- Flash TTS : $0.50/M input tokens + $9.00/M audio output tokens (now)
- Flash-Lite TTS : $0.50/M input tokens + $6.00/M audio output tokens (now)
- After January 1, 2027 : Input doubles to $1.00, output doubles to $18.00/$12.00
- Batch mode : Half price for offline generation
If you are building anything that will run past this year, budget at the post-2027 rates now — not the introductory ones. One practical gap worth flagging: Google does not publish a token-to-audio-time conversion rate, so cost estimation requires testing with your actual scripts rather than back-of-envelope math. Note that Google recently also ended the Gemini Omni Flash preview, so the 3.8 TTS line is the current stable target for new development.
What to Build #
The practical target list is broad. Voice agents using Flash-Lite TTS for high-volume inference, audiobooks and long-form narration where Flash TTS earns its cost premium, game characters where voice design creates distinct NPC identities without recording sessions, and accessibility features where 130-language coverage closes gaps most TTS APIs cannot bridge. The Google Gemini API docs include a TTS quickstart notebook covering all four voice modes: prebuilt, extended library, voice design, and voice replication. Platform integrations are already live with Vercel, LiveKit, Pipecat, Agora, Figma, and HeyGen.