Google opens custom voice creation to all developers; OpenAI still requires sales calls

Developers can now build voices into applications by either describing what they want or providing a short audio sample, then storing and retrieving that voice through the Gemini API. The capability arrives in two new models released today: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, accessible via the Gemini API and Google AI Studio. Until now, text-to-speech systems have forced developers to choose from a fixed set of pre-built voices, but Gemini 3.8 inverts that constraint by putting voice creation in the hands of the developer.

Turning recordings into voice IDs

Voice replication works through a new Voices endpoint at POST /v1beta/voices. The process requires two clean audio samples from the same speaker, each between 10 and 30 seconds long, plus a separate consent recording in which the speaker confirms ownership of the voice and permission for Google to synthesize it. Google verifies that the person in the consent clip matches the reference sample before creating the voice profile.

Once approved, Google assigns a voice_… ID that persists in the developer's project for one year, alongside any voices generated through the voice-design tools. A single project can store up to 200 voices total, and developers manage them through standard API operations: retrieve, list, or delete. Alternatively, developers can skip storage and set store=False to receive an encrypted voicekey_… that lives with the application itself; this key expires after seven days, making it suitable for temporary jobs.

Google marks all Gemini-generated audio with SynthID watermarking, and replicated voices also receive C2PA content credentials for traceability. The voice replication feature is not available through AI Studio in Illinois, Texas, the European Economic Area, the U.K., Switzerland, or India.

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.

Prompting a voice from scratch

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

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

Google recommends creating a voice once and reusing its ID across requests rather than re-describing the same persona each time. The documentation notes that repeatedly submitting lengthy persona descriptions is the primary driver of voice drift. After initial creation, subsequent calls need only brief style instructions, if any.

A significant change in Gemini 3.8 is that input text is now treated as a strict verbatim transcript—a breaking change for users who embedded stage directions in prompts to the 3.1 preview. Sustained vocal effects like whispering, sarcasm, or rapid speech now belong in a speech_metadata annotation, while momentary sounds such as <sigh>, <cough>, and <short pause> appear inline in angle brackets. For two-speaker scripts, listener reactions wrapped in pipes like |mhm| generate backchannels and overlapping speech without fragmenting the script into separate 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

Two-speaker generation carries a constraint worth noting. A single request can handle up to two speakers when using prebuilt voices, but dialogue between custom or replicated voices requires turn-by-turn generation and assembly from 24 kHz PCM output.

Unary requests return WAV by default; streaming requests return raw 16-bit PCM. Mu-law and A-law encodings are available for telephony systems. Google states that Flash TTS preserves voice quality and timbre across extended audio, positioning it for audiobook and podcast work.

Flash for performance, Flash-Lite for volume

Both models share an identical API schema, so switching between them requires only a parameter change. Both support voice design and replication.

Flash TTS targets demanding performance scenarios: complex dialogue, heavy vocal annotation, difficult pronunciations, regional accents, and extended narration. Flash-Lite TTS prioritizes speed and cost, serving as the successor to gemini-3.1-flash-tts-preview and optimized for high-volume production, read-aloud functionality, and cascaded voice agents that combine a text model with a separate speech generation step.

For such agents, Google recommends issuing one TTS call per turn as the language model produces text, with the stored voice maintaining consistency throughout the conversation.

Plugging into voice agent frameworks

A speech synthesis engine represents just one component of a production voice application. Real-time agents require transport, speech recognition, turn detection, interruption management, and session tracking. Google directs developers toward existing frameworks that handle these layers, citing Agora, LiveKit, Pipecat, and Vercel's AI Gateway as platforms offering Gemini speech generation integration through the Gemini API.

This approach allows teams to substitute Gemini as the speech layer without redesigning their audio infrastructure, though anyone planning to use a replicated voice should verify that their chosen framework passes custom voice_… IDs through before committing. API access via Gemini Enterprise is noted as forthcoming.

How OpenAI's approach compares

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

OpenAI lacks Google's ability to generate voices from text descriptions. Its 13 built-in voices can be adjusted for tone or speed, and applications must disclose that speech is artificially generated.

Google's edge lies in offering developers multiple pathways to construct the voice they envision before any text reaches the synthesis engine.

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.