{"schemaVersion":"1.0","type":"Article","types":["Article"],"slug":"gemini-3-8-flash-tts-arrives-with-voice-design-and-line-by-line-direction-blt32","url":"https://zyvop.com/gemini-3-8-flash-tts-arrives-with-voice-design-and-line-by-line-direction-blt32","title":"Gemini 3.8 Flash TTS Arrives With Voice Design and Line-by-Line Direction","subtitle":"Two new Gemini API models ship with a same-day docs and pricing update, a new speech_metadata prompting shape, and rates that undercut the outgoing preview model.","tldr":"Google shipped Gemini 3.8 Flash TTS and Flash-Lite TTS today, with the docs and pricing page updated same-day. Includes a working code example using the new speech_metadata prompting shape, benchmark results, and real per-token pricing versus the outgoing preview model.","keywords":["SynthID","text-to-speech","Google AI Studio","Gemini API","Gemini 3.8","AI News"],"entities":["Samod Alex","SynthID","text-to-speech","Google AI Studio","Gemini API","Gemini 3.8","AI News","ZyVOP"],"keyTakeaways":["Google shipped two new speech models today: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, described in the official announcement as the company's most expressive audio generation models yet.","Both take text in and return audio out through the existing Gemini API, but they add generative voice design, a much larger voice library, and finer line-by-line control over delivery.","The fastest way to see what changed is to make a call, so start there."],"headings":["Try It: A Working Example","Two Models, Two Jobs","How It Benchmarks","Safety and Consent","What to Verify Before You Ship"],"outboundLinks":["https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/","https://ai.google.dev/gemini-api/docs/speech-generation","https://aistudio.google.com/generate-speech?model=gemini-3.8-flash-tts","https://notebook.google.com/","https://aistudio.google.com/generate-speech?model=gemini-3.8-flash-lite-tts","http://vids.new/","https://docs.cloud.google.com/gemini-enterprise-agent-platform","https://ai.google.dev/gemini-api/docs/pricing","https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-live-3-5-translate/","https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/","https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/","http://docs.agora.io/en/ai/models/tts/gemini","https://docs.livekit.io/agents/models/tts/gemini/","https://docs.pipecat.ai/api-reference/server/services/tts/google#geminittsservice","https://vercel.com/docs/ai-gateway/modalities/text-to-speech","https://www.hume.ai/rw-voice-eq","https://voicearena.com/tts-leaderboard/us-english","https://deepmind.google/models/synthid/"],"contentText":"Google shipped two new speech models today: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, described in the official announcement as the company's most expressive audio generation models yet. Both take text in and return audio out through the existing Gemini API, but they add generative voice design, a much larger voice library, and finer line-by-line control over delivery. The fastest way to see what changed is to make a call, so start there. Try It: A Working Example Gemini 3.8 TTS changed the prompting model from what earlier previews used. The input text is now a strict verbatim transcript: sustained delivery instructions like tone, pace, or emotion go in a speech_metadata annotation's style field instead of being written as prose stage directions, per the updated TTS guide. Momentary sounds like a cough or a pause still go inline, but as angle-bracket tags such as <sigh> or <short pause>, not square brackets. Here is a single-speaker call in Node using that shape: import * as fs from \"node:fs\"; import { GoogleGenAI } from \"@google/genai\"; const client = new GoogleGenAI({}); async function main() { const interaction = await client.interactions.create({ model: \"gemini-3.8-flash-tts\", input: [ { type: \"user_input\", content: [ { type: \"text\", text: \"Here is the interesting part. The model now treats this text as a verbatim transcript, not a script with stage directions baked in.\", annotations: [ { type: \"speech_metadata\", style: \"warm, unhurried, technical confidence\", }, ], }, ], }, ], response_format: { type: \"audio\" }, generation_config: { speech_config: [{ voice: \"Kore\" }], }, }); const audioBuffer = Buffer.from(interaction.output_audio.data, \"base64\"); fs.writeFileSync(\"nadia-walkthrough.wav\", audioBuffer); } await main();Unary calls like this one now return a complete WAV file with a RIFF header by default, so the response bytes write straight to disk; add stream: true and you get headerless raw PCM chunks instead for real-time playback. Earlier preview models returned headerless PCM by default even for unary calls, which needed manual WAV wrapping, one less dependency if you're migrating. The pipeline looks the same whichever model you pick; only the model string and the prompt content change: flowchart TD A[Pick a use case] --> B{Creative direction or scale} B -->|Acting cues, dialects, long-form stability| C[gemini-3.8-flash-tts] B -->|High-volume, cost-efficient scale| D[gemini-3.8-flash-lite-tts] C --> E[Verbatim transcript plus speech_metadata style] D --> E E --> F[client.interactions.create with speech_config] F --> G[WAV bytes in output_audio.data] G --> H[Write straight to disk or stream as raw PCM]Multi-speaker dialogue now takes a speech_config.speakers array with mode: \"conversational\", and every turn needs its own speaker name inside that turn's speech_metadata, as the same guide explains. A newer addition is backchannel syntax: wrapping a reaction in pipes, like \"So the launch is Thursday |oh hmm| are we ready?\", layers a listener's aside into the active speaker's turn without a separate turn. Voice design, generating a wholly new voice from a text description, is documented now too: POST /v1beta/voices with type=\"prompted\" returns a persistent voice_... ID and a sample preview, usable in AI Studio or the API directly. Voice replication works the same way with type=\"replicated\", gated by the consent recording described below. Two Models, Two Jobs Model Built for Where it runs today Gemini 3.8 Flash TTS Original character voices, line-by-line acting cues, dialect shifts, backchanneling Gemini API and AI Studio for developers; Gemini Notebook for everyone Gemini 3.8 Flash-Lite TTS High-volume, cost-efficient dubbing and voice agents at scale Gemini API and AI Studio for developers; Google Vids for everyone Enterprise access to both models is coming soon through the Gemini Enterprise Agent Platform. Google's own model-selection guidance is blunter than the launch framing: reach for Flash TTS when maximum acoustic fidelity and nuanced acting matter most, and reach for Flash-Lite TTS as the direct workhorse replacement for gemini-3.1-flash-tts-preview in high-volume, cost-sensitive work. Pricing backs that up. Flash TTS audio output runs $9 per million tokens through the end of 2026 ($18 after), Flash-Lite runs $6 ($12 after), and both undercut the outgoing preview model's flat $20 rate by a wide margin. Audio tokens work out to 25 per second, per the pricing page. Both new models join 3.5 Live Translate, 3.5 Transcribe, and 3.8 Live and 3.8 Live Extended Thinking in the Gemini Audio family. Developer platforms including Agora, LiveKit, Pipecat, and Vercel's AI Gateway already route through the Gemini API for TTS. Google also names Figma, HeyGen, Linguana, Wondercraft, 99.co, and Ollang as partners integrating the new models for dubbing and localization. How It Benchmarks Benchmark Result Hume AI Voice Design Benchmark, overall Gemini 3.8 Flash TTS: 71.4, #1 overall Hume AI, accent modeling Gemini 3.8 Flash TTS: 60.8, leading score Hume AI Overall Quality Index Flash TTS #1, Flash-Lite TTS #2 Voice Arena blind preference, key languages Top positions in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi The improvement over Gemini 3.1 Flash TTS, according to Google, comes mainly from long-form consistency and dual-speaker screenplay control, both areas where the earlier model was prone to drift. Voice coverage also grew from 30 named presets to a library of over 2,000 production-ready voices, with Flash TTS covering 130 languages and Flash-Lite TTS covering 101, per the updated docs, including regional variants like Mexican Spanish, Quebec French, and Scots English. Safety and Consent Every clip from Gemini's audio models carries an inaudible SynthID watermark, which Google says keeps the audio identifiable as AI-generated and helps curb misinformation. Cloning a voice additionally requires a verbal consent recording from the voice's owner, and Google attaches C2PA credentials to replicated output for provenance tracking. That replication feature is not available inside AI Studio in Illinois, Texas, the EEA, the UK, Switzerland, or India, per a footnote in the announcement, so confirm regional availability before building a feature around it. What to Verify Before You Ship The TTS documentation and pricing page both updated the same day as this launch: model IDs, per-token pricing, and the new speech_metadata prompting shape are now officially documented rather than inferred from AI Studio links. One rough edge remains. The Go code samples on the TTS guide still call gemini-3.1-flash-tts-preview in several places, even though the Python, JavaScript, and REST examples on the same page were updated to the 3.8 models. If you're working in Go, swap the model string manually until Google catches that page up.","contentHash":"sha256:cae026790c21983911528525be667b559f61963b6c7a6357f93c8f52d97c7870","authorName":"Samod Alex","authorUrl":"https://zyvop.com/author/samod","authorSameAs":[],"category":"AI News","tags":["SynthID","text-to-speech","Google AI Studio","Gemini API","Gemini 3.8"],"audience":"Software engineers and developers building applications with AI News","tone":"Practical and evidence-based engineering guidance","readingTimeMinutes":5,"wordCount":1058,"faqs":null,"primaryTopic":"AI News","publishedAt":"2026-09-24T08:18:11.358Z","updatedAt":"2026-09-24T08:18:11.358Z","canonicalUrl":"https://zyvop.com/gemini-3-8-flash-tts-arrives-with-voice-design-and-line-by-line-direction-blt32"}