Hands-On with Gemini 3.8 TTS: 2,000+ Voices and Multi-Character Dialogue Tested

Two New Models, One Big Leap

Google's two newly released text-to-speech models — gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts — push TTS capabilities to a new level. Developer Simon Willison built a bring-your-own-key TTS Playground on top of these models using vibe coding, giving us a hands-on way to experience their real-world performance.

What the Two New Models Bring

Google released two speech synthesis models at once: the higher-performance gemini-3.8-flash-tts and the more affordable gemini-3.8-flash-lite-tts. Their standout feature is a built-in voice library of over 2,000 voices spanning multiple languages and timbres.

Beyond that, both models support custom voice cloning. According to the official documentation, all it takes is "a 30-second audio sample" — whether it's your own voice or one you have the rights to use — and it can be replicated. This means TTS is no longer confined to preset voice profiles, and the barrier to personalized voice generation has dropped significantly.

A Ready-to-Use Playground

Simon Willison took advantage of the CORS-friendly Gemini API and used GPT-4o (via vibe coding) to build a fully frontend Playground. The design is intentionally minimal: users bring their own Gemini API key, which is stored only in page memory and sent directly to Google — never written to browser storage or retained by the tool itself.

The interface is split into two panels. The left Compose panel handles content arrangement and supports two modes: "Single Voice" and "Conversation." The right Connection panel displays a "DIRECT API" indicator and notes that 2,089 voices are loaded and searchable. There's also an "Under the hood" panel that expands to show the full Request JSON and Response details — a nice touch for developers who want to understand the API call structure.

Notably, Compose settings are saved into the URL for easy bookmarking and sharing, while the API key is excluded — a smart balance of convenience and security.

What is "vibe coding"? It's a popular AI-assisted development approach where developers describe what they want in natural language, and an AI model (like GPT or Claude) generates working code directly. The developer focuses on direction and validation rather than writing every line. This approach dramatically reduces prototype development time and is especially well-suited for quickly testing API capabilities. The CORS policy support is what makes this purely frontend approach possible — Google allows browsers to make direct requests to the Gemini API without a middleman server, enabling zero-backend deployment.

Multi-Character Dialogue: The Real Killer Feature

One of the API's standout capabilities is the ability to define complete multi-character conversations, with each character assigned a distinct voice and delivery style instruction.

Simon demonstrated this with a script featuring two pelicans debating whether to move to Pacifica Pier. Character Gus uses the Puck voice with an "excited and gossipy" delivery; character Pearl uses the Kore voice with a "calm and unimpressed" tone. The script was generated by Claude Opus 4.5 and output directly as a renderable URL — the entire pipeline from script to audio was almost entirely AI-driven.

This ability to assign different voices and emotions to multiple characters in one pass has enormous potential for podcast production, audio content, game voiceover, and more.

Why this matters technically: Traditional TTS solutions handling multi-character dialogue typically require separate API calls for each line, manual audio stitching, and painful effort to match speaking pace and pauses between characters. Gemini 3.8 TTS's native multi-character support means the model perceives the full conversational context in a single inference pass — character transitions, rhythm, and emotional flow are all managed by the model rather than patched together in post-processing. The "delivery style" instruction is another key dimension: developers can describe a character's emotional state in natural language — such as "anxious and whispering" or "boisterous and loud" — and the model maps those descriptions to acoustic properties like speaking rate, pitch variation, and pause timing, with no manual audio parameter tuning required.

Speed and Cost: Real Numbers

The performance data is straightforward: using gemini-3.8-flash-tts (not the cheaper Flash-Lite), generating 1 minute and 18 seconds of audio took roughly 20 seconds and cost approximately $0.0274.

That works out to just under 3 cents and less than half a minute for a multi-character audio clip over a minute long. This kind of price-to-performance ratio transforms high-quality speech synthesis from an occasional experiment into a scalable production tool. Switching to Flash-Lite would bring costs down even further.

Takeaway

Gemini 3.8 TTS delivers a compelling combination: 2,000+ voices, 30-second voice cloning, native multi-character dialogue, and low cost — raising the practical bar for text-to-speech by a meaningful notch. Simon Willison's bring-your-own-key Playground provides a zero-backend, shareable, and fully transparent entry point for anyone who wants to get hands-on with this API quickly.