OpenAI Prompt-to-Voice and GPT-Live Multilingual Voices Guide

Learn how to generate custom voices from text prompts, configure GPT-Live audio sessions, and use nine new multilingual natural voices in production.

By ailog · Published Oct 17, 2026 · 7 min read · How we work
In short
  • OpenAI introduced prompt-to-voice on 2026-10-16, enabling developers to build custom voices from text descriptions via /v1/audio/voices without audio reference files.
  • Prompt-generated voices operate specifically with gpt-live-1 and are capped at 100 voices per organization, with free creation and standard session runtime rates.
  • Nine new prebuilt natural voices were added to GPT-Live spanning German, French, Hindi, Japanese, and Korean.
  • Custom voices created from descriptions differ significantly in scope, API limits, and target endpoints from recording-based custom voices.

Overview of the 2026-10-16 GPT-Live Voice Update

On 2026-10-16, OpenAI updated its audio generation capabilities across two endpoints: /v1/audio/voices and /v1/live/sessions. The core addition is prompt-to-voice, a feature generally available to all API customers that generates reusable voice profiles strictly from descriptive natural language prompts. This setup eliminates the previous prerequisite of providing acoustic reference samples or signed consent voice recordings when creating custom voices for real-time applications.

Alongside prompt-to-voice, the release expanded the built-in voice roster for the gpt-live-1 model by introducing nine natural voices covering five international languages. Developers building conversational interfaces now have access to localized out-of-the-box acoustic profiles without needing dedicated voice engineering. While prompt-generated voice profiles cost nothing to generate and store, active audio streaming during live interactions is billed under standard GPT-Live session rates. Teams evaluating overall architecture costs across conversational pipelines can project token expenses using the /tools/llm-api-cost-calculator/.

  • German voices: Nuri (feminine, natural), Juni (feminine, natural), and Flitz (masculine, natural).
  • French voices: Brise (feminine, natural) and Sillage (masculine, natural).
  • Hindi voice: Nira (feminine, natural).
  • Japanese voices: Harema (feminine, natural) and Shida (masculine, natural).
  • Korean voice: Noeul (feminine, natural).

Creating Voices from Text Prompts with the Voices API

The prompt-to-voice workflow creates an audio profile by submitting a descriptive text prompt to the /v1/audio/voices endpoint. When calling this endpoint, developers set type to prompt, define a human-readable name, pass vocal attributes inside prompt, and specify the model target as auto. An optional field named script_hint accepts up to 2,000 characters of representative spoken dialogue to guide delivery; if omitted, the API produces a sample script internally. The API will reject requests if script_hint contains blank text or is excessively short.

curl --fail-with-body https://api.openai.com/v1/audio/voices \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "prompt",
    "name": "Support assistant",
    "prompt": "A calm, reassuring adult voice for a customer support assistant. Speak with an American English accent, a mid-range pitch, a smooth, warm tone, and a steady, conversational pace.",
    "model": "auto"
  }' \
  -o voice.json

A successful API call produces a JSON response containing an object of type audio.voice. The payload includes a unique voice identifier (such as voice_123), the integer Unix timestamp created_at, the specified name, and the descriptive text prompt. The creation endpoint does not return binary audio data or preview playback files. Organizations must record and store the returned id along with the original generation prompt in their backend databases to reuse the voice profile across subsequent sessions.

{
  "object": "audio.voice",
  "id": "voice_123",
  "name": "Support assistant",
  "created_at": 0,
  "type": "prompt",
  "prompt": "A calm, reassuring adult voice for a customer support assistant. Speak with an American English accent, a mid-range pitch, a smooth, warm tone, and a steady, conversational pace."
}

Crafting prompt descriptions requires defining explicit acoustic dimensions rather than behavioral instructions. Developers should describe specific vocal qualities like regional accent, perceived adult age, presentation, pitch, timbre, texture, loudness, speech rate, and distinctiveness. Behavioral constraints—such as avoiding interruptions, formatting answers, or initiating handoffs—must be reserved for session-level system prompts rather than the voice definition.

Configuring Prompt Voices in GPT-Live Sessions

Prompt-generated voices are engineered specifically for gpt-live-1 and cannot be targeted through the Text to Speech API, Realtime API (gpt-realtime-2), or standard Chat Completions audio outputs. When initializing a GPT-Live session over WebRTC or WebSockets, developers reference the voice by passing an object containing the voice identifier within session.audio.output.voice.id. This contrasts with standard built-in voices like quartz, marin, or nuri, which are configured directly as plain string values.

{
  "model": "gpt-live-1",
  "instructions": "You are a helpful voice assistant. A calm, reassuring adult voice for a customer support assistant. Speak with an American English accent, a mid-range pitch, a smooth, warm tone, and a steady, conversational pace.",
  "audio": { "output": { "voice": { "id": "voice_123" } } }
}

To achieve optimal vocal adherence and consistency, documentation prescribes appending the original descriptive voice prompt directly into session.instructions alongside the conversational instructions. When connecting through WebRTC, this configuration is supplied in the JSON session object alongside transport. In WebSocket architectures, the configuration must be included in the initial session.start payload. Voices are fixed at session startup; switching voice IDs requires establishing a new session.

During an ongoing session, applications can track live interactions via transcript deltas without interrupting streaming audio. The endpoint transmits session.input_transcript.delta for caller speech and session.output_transcript.delta for assistant speech, both carrying start_ms and end_ms millisecond offsets relative to session start. To run secondary validation or guardrail checks against output fragments without increasing primary conversational latency, lightweight text models such as gpt-5.6-luna can be incorporated into parallel worker routines. For broader context on token streaming events across multimodal providers, see /posts/streaming-llm-responses-event-schemas-for-claude-gpt-and-gemini/.

Comparing Prompt-Based and Recording-Based Custom Voices

OpenAI supports two separate mechanisms for establishing custom voices: prompt-to-voice generation and recording-based synthesis. These paths serve different operational requirements, impose distinct account restrictions, and integrate into different API interfaces.

Feature / Dimension Prompt-to-Voice Recording-Based Voice
Creation Input Text description (plain prompt) Consent recording + reference audio sample
Supported Runtimes GPT-Live (gpt-live-1) GPT-Live, Realtime API, Text to Speech, Chat Completions
Availability General availability to all API accounts Limited to eligible accounts (sales access)
Organization Limit Up to 100 voices per organization At most 20 voices per organization
Sample Length Limit Optional script_hint up to 2,000 characters Audio sample 30 seconds or less (minimum 5s speech)
Supported Audio Formats Not applicable (text prompt only) mpeg, wav, ogg, aac, flac, webm, mp4 (max 10 MiB)
API Endpoints Used POST /v1/audio/voices POST /v1/audio/voice_consents, POST /v1/audio/voices
Creation Cost Free Free (requires account eligibility approval)

Recording-based voices require an exact consent phrase recited by the voice owner across 16 supported languages, including German, English, Spanish, French, Hindi, Indonesian, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Russian, Ukrainian, Vietnamese, and Chinese. Furthermore, the audio sample must contain at least five seconds of actual spoken audio and at least 15 transcribed tokens. Prompt-to-voice bypasses these audio capture steps entirely, allowing teams to quickly generate conversational personas within the 100-voice organizational quota.

Operational Limits, Data Retention, and Safety Policies

Deploying prompt-generated voices requires strict adherence to OpenAI usage and safety policies. Voices must represent adult speakers; generating children voices is explicitly disallowed. Descriptions cannot prompt for the imitation or likeness of named individuals, public figures, or fictional characters. Because prompt-to-voice algorithms do not guarantee uniqueness, multiple developers submitting similar descriptive terminology may produce overlapping or comparable acoustic outputs.

Transparency rules mandate informing listeners that the voice output is synthetic and AI-generated. Developers must preserve any watermarks, provenance metadata, or indicators of synthetic origin. From a data privacy standpoint, voice generation via /v1/audio/voices is incompatible with Zero Data Retention (ZDR). However, the subsequent speech generation in gpt-live-1 sessions is eligible for ZDR if configured for the organization. Note that enabling ZDR disables stored session recordings, recording downloads, and session forking.

  • Voice quotas: A hard maximum of 100 prompt-generated voices per organization applies, inclusive of end-user programmatic voice creations.
  • Token limits: Startup session instructions accept up to 16,384 tokens, while initial session history supports up to 128 messages and 8,192 combined tokens.
  • Dynamic session appends: Context updates via session.instructions.append, session.thinking.append, and session.commentary.append must each stay within 500 tokens.
  • Input controls: Microphones can be toggled without dropping sessions by issuing session.input_audio.mute and session.input_audio.unmute commands.

Production Implementation and Migration Checklist

Engineering teams adopting prompt-to-voice or migrating conversational pipelines to the new GPT-Live natural voices can follow this verification sequence before releasing audio agents to end users.

  • Verify endpoint selection: Use /v1/audio/voices with type: prompt and model: auto for text-based voice generation; do not attempt to invoke these IDs within Text to Speech (/v1/audio/speech) or Realtime (/v1/realtime) calls.
  • Test prebuilt international voices: If supporting German, French, Hindi, Japanese, or Korean, evaluate the nine native natural voices (such as flitz, sillage, or noeul) before authoring prompt voices.
  • Maintain voice registries: Store the returned voice id along with the source prompt description in backend storage to allow repeatable session configurations.
  • Embed descriptions in instructions: Inject the identical descriptive text from voice creation into session.instructions at GPT-Live startup to maximize acoustic consistency.
  • Review quota ceilings: Ensure multi-tenant applications managing individual user voices do not breach the hard limit of 100 prompt voices per organization.
  • Audit safety and disclosures: Ensure client user interfaces provide clear disclosure of synthetic speech and do not attempt to emulate recognizable public figures or underage voices.
  • Verify retention policies: Confirm whether organizational compliance policies mandate Zero Data Retention, noting that voice creation retains metadata while live session execution can comply with ZDR.
Sources
  1. OpenAI API changelog — 2026-10-16
  2. Custom voices
  3. Managing GPT-Live sessions

Related