Developers using Google's Gemini 3.8 TTS can now build voices from natural-language prompts or by providing short audio samples, then reference those voices repeatedly through a stored ID in the Gemini API. The company announced Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS today via the Gemini API and Google AI Studio, marking a shift from the traditional text-to-speech landscape where developers were confined to pre-existing voice options.

The workflow centers on creating a voice once and reusing it across an application. Developers describe the voice they want or supply a short recording, then save the result and invoke it by ID on subsequent requests. Google manages the voice profile internally, eliminating the need to resend the original recording or description with each new synthesis call.

Turning recordings into voice IDs

Voice replication uses a new Voices endpoint (POST /v1beta/voices) and requires two recordings from the same speaker. Both samples must be clean audio between 10 and 30 seconds in length, plus a separate consent recording where the speaker confirms ownership of the voice and agrees to synthetic replication. Google verifies that the consenting speaker matches the reference clip before proceeding.

Upon approval, Google returns a voice_… ID stored in the developer's project for one year, alongside any voices created through the voice-design tools. Each project can hold up to 200 voices total, and developers can retrieve, list, or delete them through the API like any other stored resource.

A project can hold up to 200 voices in total, and developers can retrieve, list, or delete them through the API just as they would other stored resources.

Voice replication also supports a temporary mode. Setting store=False returns an encrypted voicekey_… instead, which remains with the application and must be supplied again when the voice is needed. Since the key expires after seven days, this approach suits short-lived tasks.

Google marks all Gemini-generated audio with SynthID, and replicated voices carry C2PA content credentials for tracing audio origins. However, voice replication is unavailable through AI Studio in Illinois, Texas, the European Economic Area, the U.K., Switzerland, and India.

Prompting a voice from scratch

The voice-design feature generates a persona from natural-language descriptions of role, accent, and character, supporting more than 100 languages and dialects. Flash TTS documentation lists 130 supported languages, while Flash-Lite supports 101. Google also maintains a library exceeding 2,000 production-ready voices.

Developer documentation describes 30 prebuilt studio voices plus hundreds more in an extended library, filterable by language, accent, pitch, and use case through GET /v1beta/voices. A remixing feature for adjusting timbre, pitch, pace, and accent of library voices using prompts is listed as coming soon.

Google recommends creating a voice once and reusing its ID rather than redescribing the same persona in every request. According to the documentation, repeatedly sending lengthy persona descriptions is the leading cause of voice drift. After creation, subsequent requests need only a brief style instruction, if any.

Gemini 3.8 treats input text strictly as a verbatim transcript—a breaking change for users who embedded stage directions in prompts to the 3.1 preview model. Sustained direction for a turn, such as whispering, sarcasm, or rapid speech, now belongs in a speech_metadata annotation, while momentary sounds like <sigh>, <cough>, and <short pause> appear inline in angle brackets. In two-speaker scripts, listener reactions wrapped in pipes, such as |mhm|, generate backchannels and overlapping speech without splitting the script into additional turns.

Gemini 3.8 sees input text strictly as a verbatim transcript, a breaking change for anyone who embedded stage directions in prompts to the 3.1 preview model.

Two-speaker scripts have limits

Native two-speaker generation carries one important constraint. A single request can handle up to two speakers using prebuilt voices, but dialogue between designed or replicated voices must be generated turn by turn and assembled from 24 kHz PCM output.

Unary requests return WAV by default, while streaming requests return raw 16-bit PCM. Mu-law and A-law encodings are available for telephony applications. Google states that Flash TTS preserves voice quality and timbre across hours of continuous audio, targeting audiobook and podcast production.

Flash for performance, Flash-Lite for volume

Both models share an identical API schema, making it a one-parameter change to switch between them. Both support voice design and replication.

Google positions Flash TTS for demanding acting work, including complex dialogue, heavy vocal tag usage, difficult pronunciations, regional dialects, and extended narration. Flash-Lite TTS is the faster, more cost-effective option and the direct successor to gemini-3.1-flash-tts-preview, optimized for bulk production, read-aloud features, and cascaded voice agents that combine a text model with a separate speech step.

For such agents, Google recommends one TTS call per turn as the language model's text arrives, with the stored voice maintaining identity throughout the conversation.

Plugging into voice agent frameworks

A speech model represents only one layer of a production voice application. A real-time agent still requires transport, speech recognition, turn detection, interruption handling, and session state. Google directs developers toward frameworks that already manage those layers, citing Agora, LiveKit, Pipecat, and Vercel's AI Gateway as platforms supporting Gemini speech generation through the Gemini API.

This approach allows teams to integrate Gemini as the speech layer without rebuilding their audio pipeline. However, anyone planning to depend on a replicated voice should verify that their framework passes custom voice_… IDs through before committing. API access through Gemini Enterprise is listed as coming soon.

How OpenAI's approach compares

OpenAI provides custom voices as well, but with tighter restrictions. Customers must engage with sales, face a limit of 20 voices per organization, and must supply a consent recording alongside a voice sample of up to 30 seconds. The resulting voice ID functions across its speech endpoint, Realtime API, and Chat Completions.

OpenAI lacks Google's prompt-based voice design capability, which can synthesize a voice from a written description alone. Its 13 built-in voices can be adjusted for tone or speed, and applications must disclose that the speech is AI-generated.

Google's advantage is that it's giving developers more ways to create the voice they want before the first line of text ever reaches it.