Create and deploy custom audio with our new text-to-speech models: 🔵 Gemini 3.8 Flash TTS: Design unique voices with distinct accents and characteristics. 🔵 Gemini 3.8 Flash-Lite TTS: Built for efficiency and scale, choose from your created styles or our expansive production-ready library.

Fine-tune the delivery line by line, shaping pacing, emotion, and cues like laughs or pauses. All generated audio is watermarked with SynthID so it can be reliably identified as AI-generated. Start building with the Gemini API via @GoogleAIStudio. Find out more →

Just drop Gemini 4 Pro guys...

When is Gemini next pro model coming out?

can’t wait to accidentally make my app speak in a 1920s transatlantic accent

It’s funny when an apology, explanation, and a customer resolution all sounded the same 😁 No more! Massive kudos to the Google Gemini team for the much needed Gemini 3.8 Flash and Flash-Lite TTS updates. Our thoughts on how this helps voice-AI builders:

the accent customization is interesting. how much control do developers get over prosody and speaking style beyond just accent? can you dial in specific emotional tones or does it stay neutral?

Using my own voice is what I need! Will my voice be used for improving your models by default? Or is that opt-in (or just not an option)?

要点をまとめて日本語で紹介しました Key points summarized in Japanese:

For a custom voice, pronunciation rules belong next to accent controls. The product name shouldn’t sound different in onboarding and support.

This is great, i see a lot of creativity coming out of this

Coming soon: a tidal wave of advertisements in voices you didn't ask for but will hear anyway

Flash 重表达、Flash-Lite 重规模，这条产品分割很清晰。做中文口播更关心：口音/风格可定制之后，音色漂移和长段落稳定性怎么样。

Custom watermarked voices are one product surface. Multi-agent game seats are another place models actually have to act. Free four-seat expedition on TapeKit — captain human or external AI, teammates mix AIs + tactical chips:

Yap. 👍

the demo gets attention but the real test is who owns the stop rule when the agent loops

Exciting update. Great to see more flexibility for creating unique voices at scale.

The useful split is Flash for fidelity and acting, Flash-Lite for throughput and latency. For a voice agent, Lite is the sensible default to benchmark first. SynthID helps identify generated audio, but it does not answer consent or impersonation questions for replicated voices.

Getting closer to giving every agent its own voice.

Custom voices only help if the controls stay simple and predictable.

The two-model split is the key point: Flash TTS for designing unique voices with distinct accents, Flash-Lite for efficiency and scale. Line-by-line control over pacing, emotion, laughs and pauses targets TTS's weakest link — prosody.

TTS 这一步挺实用，尤其是把“声音风”变成可复用的生产资产。真正难的是长对话里保一致，不然开像助手，聊十句就换了个人。

Custom voices become product-grade when teams can version their style, test pronunciation and pacing, and preserve provenance across edits. Audio quality is not only realism, it’s repeatability.

我在本地 CPU 上跑 IndexTTS 克隆自己的声音，RTF 约 36，一分钟音频要算半个多小时。 托管 TTS 替没显卡的人省掉的，是这台机器，不是那点调用费。

great audio for the best agents!! machine + humans harmony

Line-by-line pacing control is what makes TTS usable past a one-off demo.

The delivery controls are the interesting part

Flash-Lite just turned the autocomplete tab into a casting department.

I strongly hope to see Google Meet's transcription accuracy improved, and a lite transcribe model cheaper than the current one would make integrating it into Meet economically viable.