Gemini 3.8 Flash TTS: Google's New Text-to-Speech Models Speak 130+ Languages — With a Free Tier
On September 23, 2026, Google announced two new text-to-speech models in the Gemini family: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS — described on the official blog as the company's "most expressive audio models yet." The headline detail for much of the world: Flash TTS covers over 130 languages and Flash-Lite over 100, and both are usable free of charge on the Gemini API free tier before you ever pay a cent. This guide walks through what launched, how the two models differ, what it costs, and how to start building with it — written for developers, creators, and product teams worldwide.
Source: Google's official blog — Gemini 3.8 text-to-speech announcement
The short answer on capabilities
Both models accept text-only input and produce audio-only output — single-speaker or two-speaker conversations in one request. Control happens through structured speech_metadata plus inline vocal tags: style, accent, pace, and tone at the turn level, and point-in-time direction inside the text itself. You can write <short pause> for a beat of silence, <sigh> for a sigh, or use |reaction| pipe syntax for backchannels where a listener reacts while the other speaker continues.
In plain terms: you are not choosing a voice and hitting "read aloud." You are directing a performance — placing breaths, hesitations, and reactions the way a voice director coaches a human actor.
Two models, two jobs
Google drew the roles clearly, and the right choice depends on your workload:
| Dimension | Gemini 3.8 Flash TTS | Gemini 3.8 Flash-Lite TTS |
|---|---|---|
| Official positioning | "Maximum acoustic fidelity, nuanced acting, expressive control" | "Fast, cost-efficient workhorse" |
| Best for | Complex dialogue, long-form narration, flagship production | Bulk generation, voice agents, high-volume apps |
| Languages | 130+ | 100+ |
| Output audio price per 1M tokens (until Dec 31, 2026) | $9.00 | $6.00 |
| Price from Jan 1, 2027 | $18.00 | $12.00 |
| Text input per 1M tokens | $0.50, then $1.00 in 2027 | $0.50, then $1.00 in 2027 |
| Free tier | Standard + Priority free of charge | Standard + Priority free of charge |
| Switching | One parameter change — same API schema | One parameter change — same API schema |
The engineering detail that matters: both models share one API schema, so moving from the premium model to the cheap one is literally a one-parameter change. Prototype with Flash, scale with Flash-Lite, no rewrite.
Voices: 30 studio presets, design, and cloning
Out of the box you get 30 prebuilt studio voices with names like Zephyr, Puck, Charon, Kore, Fenrir, Leda, Aoede, Achernar, and Sulafat. Beyond that, an Extended Voice Library is queryable via GET /v1beta/voices, filterable by language, region, accent, gender, pitch, persona, and context — with free-text search (requires google-genai 2.25.0+ or @google/genai 2.24.0+).
The bigger news is that the 3.8 generation supports both voice design and voice replication — two features the older preview models (3.1-flash-tts-preview, 2.5-pro-preview-tts) did not have. Voice design generates a custom voice from a text description; voice replication lets you build a consistent voice for your projects.
Know the limits before you architect around them: 200 stateful voices per project with a 1-year TTL, and stateless voicekey_... voices live just 7 days — a real constraint for long-running production pipelines.
Source: Official Gemini API speech generation docs
Pricing, worked out in real numbers
From the official pricing page, per million tokens — and note the official footnote: audio tokens are 25 tokens per second of audio, so one minute ≈ 1,500 tokens.
- Standard tier: text input $0.50, audio output $9.00 (Flash TTS) or $6.00 (Flash-Lite TTS) — through December 31, 2026.
- From January 1, 2027: input doubles to $1.00, output doubles to $18.00 / $12.00.
- Batch/Flex mode: half price ($4.50 / $3.00 output today) for non-urgent bulk jobs.
- Priority tier: $16.20 / $10.80 output for latency-sensitive loads.
- Free tier: the pricing page lists both models as free of charge on Standard and Priority for the free tier (Batch/Flex not available there).
A concrete calculation: one hour of audio = 3,600 seconds × 25 tokens = 90,000 audio tokens. At Flash-Lite TTS's current $6.00 per million, that is about $0.54 per finished hour of speech — before prices double in 2027. For audiobook producers, e-learning teams, and app builders in any emerging market, that number changes what is economically viable.
One planning note: these are launch prices, and Google has put the increase in writing. Budget multi-month projects against the 2027 rates, not today's.
How to start building
Access runs through the Gemini API itself: a Google account, an API key from the developer platform, and requests via the Interactions API or the official SDKs for Python (google-genai), JavaScript (@google/genai), and Go.
Format behavior worth knowing before you write code: standard unary requests return WAV files; streaming (stream: true) returns headerless raw PCM — 24 kHz, mono, 16-bit little-endian — which is exactly what real-time voice agents want. Additional formats (audio/mulaw, audio/alaw) and configurable sample rates are supported.
Multi-speaker requests support up to two speakers with prebuilt voices; custom voices are synthesized turn by turn. Language detection is automatic — you do not pin a language in the basic request.
Your first request: five practical steps
What does the road from zero to your first audio clip actually look like? The standard path, per the official docs:
- Create an API key on Google's developer platform — the key is your request identity.
- Install the official SDK for your language:
google-genaifor Python or@google/genaifor JavaScript (2.25.0+ / 2.24.0+ respectively if you want the Extended Voice Library queries). - Pick your model by job:
gemini-3.8-flash-ttsfor premium output orgemini-3.8-flash-lite-ttsfor volume. Same schema — the decision is reversible with one parameter. - Set voice and style: one of the 30 studio presets (or from the voice library endpoint), then the overall style in
speech_metadataand inline tags in the text for pauses, sighs, and reactions. - Receive the audio: via the response field in the SDKs, or the response steps in raw REST — WAV for standard requests, raw PCM if you enabled streaming.
No approval queue or regional allowlist is documented for the TTS models — access runs through the same standard Gemini API available globally.
The cost of three real projects, worked out
Official numbers get useful when applied to actual work. Estimated math (using Google's stated conversion — one second of audio = 25 tokens, one minute ≈ 1,500 tokens) at pre-2027 prices:
| Project | Audio volume | On Flash-Lite TTS ($6/M) | On Flash TTS ($9/M) |
|---|---|---|---|
| Mid-length audiobook | 8 hours = 720K tokens | ≈ $4.32 | ≈ $6.48 |
| Course of 30 lessons × 5 min | 150 minutes = 225K tokens | ≈ $1.35 | ≈ $2.03 |
| Voice agent, 100K replies × 20s/month | ≈ 555 hours = 50M tokens | ≈ $300 | ≈ $450 |
The practical takeaway: for recorded content, choose Lite and save a third; reserve full Flash TTS for showcase moments — an ad, a podcast intro, a client demo. And remember every figure doubles after January 2027: anyone building a high-volume voice agent should model unit economics on $12/$18, not today's prices.
Applications you can ship today
- Audiobooks in any of 130+ languages: a small publisher can voice a novel for less than the cost of a lunch per finished hour — a revenue line that simply did not exist for text-only catalogs.
- Audio lessons for edtech: distinct voices per character (two speakers in one request) turn a dry lesson into something closer to radio drama.
- Arabic and regional-language support lines: streaming returns audio as it is generated — the foundation of a real-time phone agent.
- Accessibility materials: converting articles, handouts, and archives into speech for visually impaired readers, at costs low enough for volunteer initiatives.
Quick comparison: where this sits against what you use today
- Versus classic TTS libraries: the difference is direction. Traditional engines read text; this lets you stage it — pauses, sighs, overlapping reactions, two speakers in one call.
- Versus OpenAI's voice stack: OpenAI's recent voice work (like the GPT-Live-1 model we covered) targets live listening-and-speaking conversation. Gemini 3.8 TTS targets produced, directed audio. Teams building voice support agents may reasonably use both.
- Inside Google's own family: it slots next to the Gemini 3.8 Live lineup aimed at real-time conversation — TTS is for generated speech, Live is for live dialogue, and music generation is a separate track entirely.
What does this mean for you?
- Creators: studio-grade narration in 130+ languages with performance control, at fractions of a dollar per hour — and a free tier to experiment on before committing.
- Developers in emerging markets: the economics finally work for vernacular audio at scale. A startup in Lagos, Jakarta, or São Paulo can localize its product voice without studio budgets.
- E-learning and edtech teams: auto-generate lesson audio with distinct voices per character, in the same languages your learners speak. Combined with modern study tools, full audio-first courses are now cheap to produce.
- Teams planning past December: lock in your unit economics knowing prices double January 1, 2027 — Batch mode at half price is your friend for anything not latency-sensitive.
Honest limitations
These models are text-in, audio-out — no speech understanding inside the TTS models themselves (that is Live territory). Multi-speaker is capped at two voices, prebuilt only in a single request. The 7-day TTL on stateless voicekeys means you will manage voice lifecycle deliberately. And the current pricing is explicitly temporary. None of these are dealbreakers; all of them belong in your design doc.
Frequently asked questions
Does Gemini 3.8 Flash TTS support multiple languages?
Yes. Flash TTS supports over 130 languages and Flash-Lite TTS over 100, with automatic language detection in the request. Coverage spans widely spoken and regional languages, and both models support single-speaker and two-speaker generation with prebuilt voices.
What is the difference between Flash TTS and Flash-Lite TTS?
Flash TTS is built for maximum acoustic fidelity, nuanced acting, and expressive control — complex dialogue and long narration. Flash-Lite TTS is positioned as a fast, cost-efficient workhorse for bulk production and voice agents. They share one API schema, so switching between them is a single parameter change.
How much does Gemini TTS cost per million tokens?
Until December 31, 2026: $0.50 per million text-input tokens on both models, and $9.00 (Flash TTS) or $6.00 (Flash-Lite TTS) per million audio-output tokens. From January 1, 2027 these rise to $1.00 input and $18.00 / $12.00 output. Batch/Flex mode is half of standard pricing.
Is there a free tier for Gemini 3.8 TTS?
Yes. Google's pricing page lists both models as free of charge on the Standard and Priority tiers of the Gemini API free tier — with Batch and Flex modes excluded from the free tier. That makes it possible to test voices, languages, and workflows fully before paying.
Can I clone or design a voice with Gemini 3.8 TTS?
Yes — the 3.8 generation supports both voice design (generating a custom voice from a text description) and voice replication, features the older preview TTS models lacked. Limits: 200 stateful voices per project with a 1-year TTL, and stateless voicekey voices expire after 7 days.
Sources
- Google's official blog — Gemini 3.8 text-to-speech says hello — the launch announcement, September 23, 2026
- Gemini API docs — Speech generation — full specification: languages, voices, streaming, metadata
- Gemini API official pricing — per-tier pricing, free tier, and the 25-tokens-per-second footnote
- Gemini API on Google's developer platform — SDKs and quickstarts