Google released Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS today, introducing text-to-speech models that let developers build custom voices rather than choosing from a pre-set library. The new models allow users to either describe a voice using natural language or submit a brief audio sample to replicate an existing voice, then store and reuse that voice throughout an application. Google manages the voice profile on its end, so developers don't need to resend the original recording or description with each new request.

Voice replication requires two audio clips from the same speaker submitted through a new Voices endpoint: clean reference samples lasting between 10 and 30 seconds, plus a separate consent recording. In the consent clip, the speaker must read a statement confirming the voice is theirs and granting permission for Google to generate a synthetic version. Google verifies that the person providing consent matches the reference audio before proceeding. Once approved, the system generates a voice ID that's stored in the developer's project for one year, alongside any voices created through Gemini's voice-design tools. Each project can store up to 200 voices total, and developers can retrieve, list, or delete them through the API like any other resource. Alternatively, setting the storage parameter to false returns an encrypted voice key that remains with the application and must be supplied each time the voice is needed; because the key expires after seven days, this option suits short-term tasks.

For developers who prefer not to use recordings, voice design generates a persona from a text description of role, accent, and character. The feature works across more than 100 languages and dialects—Flash TTS supports 130 languages and Flash-Lite supports 101, according to the documentation. Google's announcement claims a library of over 2,000 production-ready voices. The developer documentation describes 30 prebuilt studio voices and hundreds more in an extended library that can be filtered by language, accent, pitch, and use case. Google recommends creating a voice once and reusing its ID rather than describing the same persona in every request, noting that repeatedly sending long persona descriptions is the most common cause of voice drift. Audio generated by Gemini carries SynthID watermarking, and replicated voices also include C2PA content credentials that can trace the audio's origin. Voice replication isn't available through AI Studio in Illinois, Texas, the European Economic Area, the U.K., Switzerland, or India.

The report explains that Gemini 3.8 treats input text strictly as a verbatim transcript, a breaking change for anyone who embedded stage directions in prompts to the 3.1 preview model. Sustained direction for a turn—such as whispering, sarcasm, or speaking rapidly—now goes in a speech metadata annotation, while momentary sounds like sighs, coughs, and short pauses sit inline in angle brackets. In two-speaker scripts, listener reactions wrapped in pipes produce backchannels and overlapping speech without breaking the script into extra turns. Single requests support up to two speakers using prebuilt voices, while dialogue between designed or replicated voices must be generated turn by turn and stitched together from the 24 kHz PCM output. Google positions Flash TTS for demanding acting work, including complex dialogue, heavy use of vocal tags, difficult pronunciations, regional dialects, and long narration; Flash-Lite TTS serves as the faster, less expensive option tuned for bulk production, read-aloud features, and cascaded voice agents that pair a text model with a separate speech step. For those agents, Google recommends one TTS call per turn as the language model's text arrives, with the stored voice carrying identity across the conversation.

Google points developers toward frameworks that already handle transport, speech recognition, turn detection, interruption handling, and session state, naming Agora, LiveKit, Pipecat, and Vercel's AI Gateway as platforms that support Gemini speech generation through the Gemini API. That lets teams drop Gemini in as the speech layer without rebuilding their audio pipeline, though anyone planning to rely on a replicated voice should confirm their framework passes custom voice IDs through before committing. API access through Gemini Enterprise is listed as coming soon. The report notes that OpenAI also offers custom voices but requires customers to go through sales, limits organizations to 20 voices, and mandates a consent recording alongside a voice sample of up to 30 seconds; what OpenAI doesn't have is Google's prompt-based voice design, which can create a voice from a written description. The report concludes that Google's advantage is giving developers more ways to create the voice they want before the first line of text ever reaches it. By opening voice creation to both natural-language design and quick replication, the platform shifts control from pre-recorded voice libraries to developer specification, potentially altering how conversational interfaces are built and personalized at scale.