Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS is Google’s flagship creative text-to-speech model, launched on September 23, 2026 next to a cheaper sibling, Gemini 3.8 Flash-Lite TTS. Google frames the pair as moving voice generation “from static presets into a dynamic creative studio”: you can design a brand-new voice from a text prompt, direct each line like a script, and keep the voice steady across long recordings.

The model code is gemini-3.8-flash-tts. It is text-in, audio-out, billed per audio token at 25 tokens per second — which is where the headline $0.00225 per 10 seconds comes from. That price is an introductory rate that ends on December 31, 2026.

Flash TTS vs Flash-Lite TTS

The two models share the same API schema and prompt structure, so switching is a model-name change.

What’s new

• Generative voice design. Describe the role, accent and voice characteristics in natural language across more than 100 languages — Google’s demos include a Melbourne DJ, a tinny monotone robot and a Japanese dragon. Designed voices can be saved and reused with minimal drift.
• A very large voice library. Beyond about 30 original voices, Google lists 2,000+ production-ready voices, including regional varieties like Mexican Spanish, Quebec French and Scots English. Voice remixing — adjusting timbre, pitch, pace and accent of a library voice — is still “coming soon”.
• Voice replication. A 30-second sample can recreate a voice you have the rights to use, gated by a verbal consent check, with SynthID watermarking and C2PA credentials.
• Line-by-line performance control. Style and speaker instructions go in structured speech_metadata; short non-speech events such as <laugh>, <sigh>, <cough>, <breath> and <short pause> go inline. Google also demos backchanneling cues like |mhm|.
• Native two-speaker scenes with distinct voices and natural turn-taking from one script, plus hours-scale long-form generation with minimal speaker drift.
• Regional accents and minority dialects, with automatic input-language detection across 130+ languages.

Benchmarks

Google reports that Gemini 3.8 Flash TTS takes #1 overall on Hume AI’s Voice Design Benchmark (71.4) and leads on accent modeling (60.8). It also places Flash and Flash-Lite TTS #1 and #2 on Hume’s Overall Quality Index, and cites top positions on Voice Arena’s blind human-preference evaluations in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. These are Google-reported results — check the live leaderboards before treating them as settled, especially for languages you ship in.

Independent blind votes are less flattering. On the Artificial Analysis Speech Arena (October 2026, each provider’s own voices), Gemini 3.8 Flash TTS is #5 at Elo 1274, behind Eleven v4 Turbo (1330), Eleven v4 (1326), Qwen-Audio 3.1 TTS-Plus (1298) and Cartesia Sonic 3.6 (1286); Flash-Lite TTS is #8 (1246) and Gemini 3.1 Flash TTS #13 (1214). Its edge is price: $16.50 per 1M characters for Flash TTS and $11.00 for Flash-Lite, against $40–$80 for Eleven v4. Those figures are consistent with Google’s introductory rates, which double on January 1, 2027, so budget for about twice that from next year.

On Artificial Analysis’s cloned-voice board, where every model reads the same eight cloned voices, Flash TTS drops to #17 (Elo 1050) and Flash-Lite TTS is #13 (1061), against 1161–1185 for Eleven v4 and Qwen-Audio 3.1 TTS-Plus. Gemini’s strengths are designed and library voices and long-form stability; if cloning a specific voice is your main need, test it against the leaders first.

Pricing

For comparison, the previous Gemini 3.1 Flash TTS Preview costs $1 per 1M text input and $20 per 1M audio output, so 3.8 Flash TTS is cheaper until the 2027 step, then roughly 10% cheaper on output. There is a free tier (where Google may use your content to improve its products), and Batch, Flex and Priority inference are supported.

Specifications and migration

• Input: text only, up to 8,192 tokens. Output: audio, up to 16,384 tokens (Gemini API service limit).
• Not supported: function calling, structured outputs, context caching, code execution, search grounding, thinking, URL context and the Live API — for live conversation, look at Google’s separate Gemini Live models.
• Default output is WAV (audio/wav) with a standard RIFF header for unary requests, instead of the headerless PCM of earlier Gemini TTS models. Remove any code that wraps PCM in a WAV header, or set response_format.mime_type to audio/l16, audio/mulaw or audio/alaw if you need raw audio.
• Prompting changes: the model reads text verbatim, so inline directions such as “Say cheerfully:” can be spoken aloud. Put style and speaker in speech_metadata, name the speaker on every turn of multi-speaker requests, and use Voice Design IDs instead of long legacy director-notes blocks.
• Deadline: gemini-3.1-flash-tts-preview, gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts shut down on November 17, 2026.
• Where to use it: Gemini API and Google AI Studio now; Gemini Enterprise via API “coming soon”; Gemini Notebook for everyone. Flash-Lite TTS also powers Google Vids.

How it compares

• vs Eleven v4: ElevenLabs leads blind votes (#1 and #2 against Gemini’s #5) and clones voices much better, with 90+ languages. Gemini 3.8 Flash TTS covers more languages (130) and costs a fraction — $16.50 versus $40–$80 per 1M characters on Artificial Analysis’s figures, or about $0.0135 per minute of audio at Google’s introductory rate.
• vs MiniMax Speech: MiniMax is a cloning-first option with fast voice cloning from about ten seconds of audio, at $60–$100 per 1M characters; Gemini ranks higher in blind votes (#5 versus #19), costs far less, and adds prompt-based voice design, scene-style two-speaker direction and long-form stability.
• vs MAI-Voice-2.1: Microsoft’s model is a public-preview alternative on Azure and Foundry — one voice across 23 languages with a native accent, gated cloning from a 5–60 second clip, $22 per 1M characters. Gemini 3.8 Flash TTS counters with 130 languages, voice design and a free tier.
• vs Qwen-Audio 3.1: Alibaba’s family bundles speech recognition, TTS and realtime models into one lineup at very low prices; Gemini concentrates on creative direction and voice design for text-to-speech.

Limitations

• Introductory pricing. Budget for $18 per 1M audio tokens from 2027.
• Narrow API surface. No tool calling or caching, and text-only input.
• Vendor-reported rankings. Hume and Voice Arena results come from Google’s launch post; the independent Artificial Analysis board has it #5, and #17 for cloned voices.
• Migration work. Existing prompts and PCM-handling code need updating.

See the audio models directory for more speech and music models.

Best for

• Audiobooks, podcasts and long-form narration with a stable voice
• Game, animation and dubbing characters designed from a text description
• Two-person scenes and dialogue scripts with explicit speaker turns
• Localized voiceovers in regional accents and dialects across 130 languages

Pros & cons

Frequently asked questions

Sources

• Google — Gemini 3.8 text-to-speech says hello
• Google AI for Developers — Gemini 3.8 Flash TTS model page
• Google AI for Developers — Gemini Developer API pricing
• Google AI for Developers — Deprecations
• Artificial Analysis — Text to Speech leaderboard (provider voices)
• Artificial Analysis — Text to Speech leaderboard (controlled voice)