OpenAI Text to Speech
OpenAI Text to Speech: voice controls, pricing and migration
OpenAI Text to Speech creates multilingual narration through an API, with delivery prompts, stock voices and streaming. Its Speech models retire in January 2027.
Paid plansAccount required

- Access
- Paid access
- Usage pricing
- USD 15 / 1,000,000 charactersTTS-1 — deprecated
- Available on
- Web, API
Narration for applications built on OpenAI
OpenAI Text to Speech turns a written script into narration for articles, learning materials and spoken application responses. The Speech endpoint returns an audio file or a stream, so playback can begin before the complete file is ready. There is an important deadline for existing integrations: TTS-1, TTS-1 HD and the March and December 2025 GPT-4o Mini TTS releases were deprecated on 1 October 2026 and are scheduled for removal on 6 January 2027. OpenAI recommends GPT-Realtime-2.1 Mini through the Realtime API for migration.
GPT-4o Mini TTS accepts plain-language directions for accent, emotion, intonation, pace and whispering. Delivery instructions are separate from the script being spoken. The Speech API lists 13 built-in voices, with a smaller set of nine for TTS-1 and HD; those older models do not accept the instructions parameter. Multilingual narration follows the language of the input text, while the stock voices are optimised for English. The endpoint exposes prose instructions and a speed control rather than an SSML parameter.
Each Speech request accepts at most 4,096 characters. Mini TTS separately has a 2,000-input-token limit, so longer articles and chapters need multiple requests. MP3 is the default output, with Opus, AAC, FLAC, WAV and raw PCM also available. Raw PCM uses 16-bit little-endian samples at 24 kHz without a WAV header. Audio streaming and server-sent events are distinct options: Mini TTS supports both, while TTS-1 and HD do not support SSE.
Custom voices require eligibility and approval from OpenAI. Voice creation uses two recordings: the speaker reads a prescribed consent statement, then supplies a matching voice sample of no more than 30 seconds. The documented limit is 20 voices per organisation. Custom-voice use is also subject to the supplemental agreement. Applications must clearly tell listeners that the speech is AI-generated. API content is excluded from model training by default unless the customer opts in; ordinary abuse-monitoring logs typically retain content for up to 30 days, with legal and safety exceptions and approved retention controls for eligible customers.
Billing is based on usage. TTS-1 charges USD 15 per million input characters and HD charges USD 30; 100,000 characters therefore cost USD 1.50 or USD 3 respectively at list rates. Mini TTS charges separately for text input and generated audio tokens, so its audio-output rate is not the full cost or a monthly subscription. The recommended Realtime replacement has a different connection model and separate input/output charges. For a new integration, its supported API and voice behaviour deserve attention before building around the retiring Speech endpoint.
Best for
- Developers maintaining existing Speech API integrations
- Publishers generating narration before the retirement date
- Teams planning a move to OpenAI Realtime speech
Limitations
The listed Speech models retire on 6 January 2027. Realtime 2.1 Mini uses a different API and billing model. Speech requests are limited to 4,096 characters, with a separate 2,000-input-token limit for Mini TTS. Custom voices require approval and recorded speaker consent. Listeners must be told that the voice is AI-generated.
OpenAI Speech API features
Script-to-audio narration
Turn written articles, lesson scripts and application responses into audio files or a stream. Playback can start before the full file is generated.
Plain-language delivery controls
Guide accent, emotion, pace, intonation and whispering with GPT-4o Mini TTS instructions. TTS-1 and HD do not support that parameter; the Speech speed control spans 0.25–4×.
Built-in multilingual voices
Choose from 13 Speech voice names, with nine available on TTS-1 and HD. Speech follows the input language, while the stock voices are optimised for English.
Six audio output formats
Export MP3, Opus, AAC, FLAC, WAV or raw PCM. Mini TTS supports audio or SSE streaming; TTS-1 and HD do not support SSE.
Consent-based custom voices
Approved customers can create a voice from a matching speaker sample and separate prescribed consent recording. Samples are limited to 30 seconds, with up to 20 voices per organisation.
Clear script limits
The Speech endpoint accepts up to 4,096 characters per request, and Mini TTS has a separate 2,000-input-token limit. Longer narration needs to be split into multiple requests.
A published migration deadline
OpenAI schedules removal of the listed Speech models for 6 January 2027 and recommends GPT-Realtime-2.1 Mini. Migration requires a Realtime integration rather than a model-name swap.
OpenAI Text to Speech pricing
The retiring Speech models remain usage-priced: TTS-1 costs USD 15 per million characters and TTS-1 HD costs 30. GPT-4o Mini TTS bills USD 0.60 per million text input tokens plus 12 per million audio output tokens. These models shut down on 6 January 2027. The recommended Realtime replacement uses separate token rates and a different API.
Example plans and usage rates; additional charges may apply.
TTS-1 — deprecated
USD 15 / 1,000,000 characters
TTS-1
Plan details
Per million input characters; scheduled removal 6 January 2027.
TTS-1 HD — deprecated
USD 30 / 1,000,000 characters
TTS-1 HD
Plan details
Per million input characters; scheduled removal 6 January 2027.
Mini TTS text input — deprecated
USD 0.6 / 1,000,000 input tokens
GPT-4o Mini TTS
Plan details
Per million text input tokens; generated audio charged separately; scheduled removal 6 January 2027.
Technical specifications
Voice generation
| Feature | OpenAI Text to Speech |
|---|---|
| Speech languages | Multilingual speech with voices optimised for English. Voice selection differs between Mini TTS and the older TTS-1 models.Deprecated models; retirement 6 January 2027 |
| Voice and pronunciation controls | Delivery instructions work with GPT-4o Mini TTS, not TTS-1 or HD. Speech speed accepts 0.25–4×. The Speech endpoint exposes no SSML parameter; built-in voice selection varies by model.Deprecated; model-specific instruction support |
| Voice cloning | Audio-sample custom voices require customer eligibility, separate prescribed recorded consent and a matching speaker reference of at most 30 seconds. Maximum 20 voices per organisation; supplemental agreement applies.Approved customers; recorded speaker consent |
| Audio exports and streaming | MP3, Opus, AAC, FLAC, WAV or raw PCM. Mini TTS supports SSE or audio streaming; TTS-1 and HD do not support SSE.Deprecated models; output and streaming selection |
| Commercial-use terms | Clearly disclose to end users that the voice is AI-generated. Custom voice creation requires speaker consent and the applicable supplemental agreement.Required disclosure; custom voice permission |
| Output sample rate (Hz) | 24000Raw 16-bit little-endian PCM at 24 kHz; not a maximum across every encoding |
| API access | SupportedAPI account and authentication; models retire 6 January 2027 |
OpenAI Text to Speech alternatives
These services offer other routes for hosted narration and application speech.
Google Cloud Text-to-Speech
Chirp stock voices and Gemini delivery prompts, with character or token billing by model.
Amazon Polly
AWS speech synthesis with engine-specific voices, SSML and pronunciation lexicons.
Cartesia
Sonic streaming voices and paid-plan custom voice creation through a managed speech API.
OpenAI Text to Speech FAQs
When will OpenAI’s current text-to-speech models shut down?
OpenAI deprecated TTS-1, TTS-1 HD and the March and December 2025 GPT-4o Mini TTS releases on 1 October 2026. Their scheduled removal date is 6 January 2027. The recommended migration model is GPT-Realtime-2.1 Mini through the Realtime API.
How much does OpenAI Text to Speech cost?
TTS-1 costs USD 15 per million input characters and HD costs USD 30. At these list rates, 100,000 characters cost USD 1.50 or USD 3 respectively. GPT-4o Mini TTS charges USD 0.60 per million text input tokens plus USD 12 per million audio output tokens. These are usage charges for the retiring models, with different billing units.
Is Realtime 2.1 Mini a direct replacement for the Speech endpoint?
It is OpenAI’s recommended migration model, but it uses the Realtime API with WebRTC, WebSocket or SIP connections and different integration and billing. Published standard rates are USD 0.60 per million text input tokens, USD 2.40 per million text output tokens, USD 10 per million audio input tokens and USD 20 per million audio output tokens. Session inputs and outputs determine the charges.
How long can an OpenAI Text to Speech script be?
The Speech endpoint caps input at 4,096 characters per request. GPT-4o Mini TTS also has a 2,000-input-token limit. Both limits matter; tokens are not the same as characters. Longer articles or chapters need multiple requests and their audio segments must be assembled by the application.
Can OpenAI Text to Speech use a custom voice?
Custom voices are available to eligible customers after approval. Creation requires a prescribed consent recording and a matching sample from the same speaker. The sample must be no longer than 30 seconds, with a maximum of 20 voices per organisation. The supplemental agreement applies, and listeners must receive clear disclosure that the audio is AI-generated.
Does OpenAI Text to Speech support SSML?
The Speech endpoint exposes a plain-language instructions parameter rather than an SSML parameter. Instructions work with GPT-4o Mini TTS, not TTS-1 or HD. A separate speed setting accepts 0.25–4×. Existing SSML controls cannot simply be passed as a supported markup option.
Which audio formats and streaming options are available?
MP3 is the default, with Opus, AAC, FLAC, WAV and raw PCM also supported. Raw PCM is 16-bit little-endian audio at 24 kHz without a header. Mini TTS supports audio or SSE streaming; TTS-1 and HD do not support SSE.
Does OpenAI train on text sent to the Speech API?
API data is not used to train or improve OpenAI models by default unless you opt in to share it. This does not mean the service stores nothing: ordinary abuse-monitoring logs typically retain content for up to 30 days, with legal and safety exceptions. Eligible customers can apply for additional retention controls; custom-voice recordings have their own endpoint and agreement requirements.
Is this your product? Claim this page to update your listing.