Now live New models for transcription and text-to-speech. AssemblyAI-STT-3.7 AssemblyAI-TTS-1
Insights & Use Cases

OpenAI TTS alternatives for real-time voice agents (before the January 2027 shutdown)

OpenAI retires tts-1 and gpt-4o-mini-tts on January 6, 2027. Compare TTS alternatives for real-time voice agents and what changes when you switch.

Abstract green cylinder illustration

Written by

Kelsey Foster

Published on

18 November 2026

On October 1, 2026, OpenAI posted a deprecation notice for its text-to-speech models. tts-1, tts-1-hd and both gpt-4o-mini-tts snapshots stop working on January 6, 2027. The replacement OpenAI names, gpt-realtime-2.1-mini, doesn't run on the endpoint those models use. OpenAI's own TTS guide puts it plainly: "This is a change of API workflow."

So every team with a voice agent on /v1/audio/speech has a decision to make before January. You can rebuild on OpenAI's Realtime API, or you can keep a separate text-to-speech stage and move it to another provider.

This post is for that decision. It covers what OpenAI is retiring, the five things a voice agent needs from a replacement, a side-by-side table of the main OpenAI TTS alternatives, and what actually changes when you leave tts-1. If you've already decided to move to AssemblyAI and want the code, go straight to how to migrate from OpenAI TTS to AssemblyAI.

What OpenAI is retiring, and what it recommends instead

OpenAI's deprecations page lists four text-to-speech models, all with the same shutdown date and the same recommended replacement:

Model Shutdown date OpenAI's recommended replacement
tts-1 January 6, 2027 gpt-realtime-2.1-mini
tts-1-hd January 6, 2027 gpt-realtime-2.1-mini
gpt-4o-mini-tts (gpt-4o-mini-tts-2025-03-20) January 6, 2027 gpt-realtime-2.1-mini
gpt-4o-mini-tts (gpt-4o-mini-tts-2025-12-15) January 6, 2027 gpt-realtime-2.1-mini

Two details matter more than the date.

The replacement isn't a TTS endpoint. OpenAI's TTS guide says gpt-realtime-2.1-mini "isn't a supported model for /v1/audio/speech." OpenAI describes it as a distilled reasoning model for realtime voice interactions, taking audio and text in over WebRTC, WebSocket or SIP connections. You don't swap a model name. You move from one HTTP request per utterance to a stateful Realtime session.

Your existing code keeps working until the shutdown. The guide says existing integrations can keep using the Speech API until January 6, 2027. That's under eight weeks away, which is enough time to evaluate alternatives properly, but not enough to leave it until the holidays.

What a voice agent needs from the replacement: five checks

A narration script and a phone agent need very different things from a TTS API. For a real-time voice agent, these five checks narrow the field fast. For a deeper look at how each provider fits LiveKit, Pipecat and other orchestrators, see our guide to the best TTS API for voice agents.

  1. Streaming text input. Your LLM writes its reply a token at a time. A WebSocket that accepts text as it arrives and starts speaking once it has a full sentence saves you from waiting for the whole reply. OpenAI's /v1/audio/speech takes a complete input string per request.
  2. Cancel for barge-in. When the caller talks over the agent, you need to stop synthesis and drop the text you've already queued, without tearing down the connection.
  3. Telephony encodings. Phone carriers expect 8 kHz G.711, either μ-law or A-law. Getting that straight from the API saves a resampling and encoding step. If you're unsure which one your carrier uses, our explainer on μ-law vs A-law for phone calls covers it.
  4. An OpenAI-compatible request shape. If your code, or your framework's OpenAI plugin, already calls /audio/speech, a provider that answers the same request shape lets you move by changing a base URL and a key.
  5. How concurrency is counted. Providers count different things: generations running at once, active contexts, or new sessions per minute. For a voice agent that holds one connection per call, the counting method decides how many calls you can run at peak. Check it before you sign anything.

Latency belongs on this list too, but it's the one number you should measure yourself, on your network, with your text. Our guide to TTS latency explains what to measure, and how to evaluate TTS for voice agents covers accuracy and naturalness testing.

OpenAI TTS alternatives at a glance

Every row below comes from that provider's own documentation, checked on November 12, 2026. "Not documented" means we didn't find it in the provider's docs, not that it can't exist. There's no price column: pricing changes often, so each provider's pricing page is linked in the next section.

Option OpenAI-compatible speech endpoint Streaming text input Barge-in cancel 8 kHz μ-law / A-law Languages LiveKit + Pipecat
AssemblyAI-TTS-1 Yes; accepts OpenAI voice names, refuses tts-1 as a model name Yes (WebSocket) Cancel frame pcm_mulaw / pcm_alaw 6 Native plugins for both
ElevenLabs (Eleven v4 Turbo) Not documented Yes (Text to Dialogue WebSocket) close_context on the multi-context socket ulaw_8000 / alaw_8000 (HTTP) 90+ Documented as Speech Engine guides
Cartesia (Sonic 3.6) Not documented Yes (WebSocket) cancel: true on a context pcm_mulaw / pcm_alaw 44 Documented for both
Inworld (TTS-2, TTS-2 Flash) Yes; refuses OpenAI model and voice names Yes (WebSocket) cancelContext MULAW / ALAW 470, in three support tiers Linked from Inworld docs to each framework's guide
Rime (Coda) Not documented Yes (WebSocket) cancel on a context μ-law (audio/pcmu); A-law not documented 10 (cloud API) Documented for both
Stay on OpenAI: gpt-realtime-2.1-mini No; Realtime API only Realtime session response.cancel μ-law (audio/pcmu) Not stated on the model page See each framework's docs
Self-host an open-weight model Depends on the server Depends on the server Depends on the server Depends on the server Depends on the model Depends on the model
Try AssemblyAI-TTS-1 on Your Own Script

Paste the lines your agent actually says, including phone numbers and confirmation codes, and hear how AssemblyAI-TTS-1 reads them before you change any code.

Try playground

The alternatives, one by one

Each entry answers one question: if you leave tts-1 for this, what changes?

AssemblyAI-TTS-1

AssemblyAI-TTS-1 (assemblyai-tts-1) is AssemblyAI's streaming text-to-speech model, trained on millions of hours of audio, including real conversational data. It speaks six languages: English, Spanish, French, German, Italian and Portuguese.

If you leave tts-1 for it, here's what changes:

  • The base URL and key. AssemblyAI answers OpenAI's POST /audio/speech request shape at https://tts.us.assemblyai.com/v1/openai (and https://tts.eu.assemblyai.com/v1/openai in the EU). Point the OpenAI SDK, or the OpenAI plugin in LiveKit or Pipecat, at that base URL with an AssemblyAI API key. The OpenAI compatibility docs cover the setup.
  • The model name. model must be an AssemblyAI name: assemblyai-tts-1, or a language model name such as assemblyai-tts-en. OpenAI names such as tts-1 are refused with 400 and the code model_not_found.
  • Voices. OpenAI voice names still work. Every one of them, alloy included, speaks with the default voice of the request's language: aai-kyla in English. To pick a different voice, send its library slug.
  • Formats and fields. mp3, pcm, wav, mulaw and alaw are accepted. opus, aac and flac are refused. instructions is ignored, and speed is accepted only at 1.

For a new build, the native WebSocket API is the better fit for a live agent. Text can arrive a token at a time, a Cancel frame handles barge-in on the same connection, and 8 kHz pcm_mulaw or pcm_alaw comes straight off the socket for telephony. Native LiveKit and Pipecat plugins run on it, with word-timed transcripts.

Concurrency. Streaming TTS sessions are counted the way Streaming Speech-to-Text sessions are: by new sessions per minute. Sessions open at once aren't counted, so calls already in progress don't count against your limit.

Benchmarks. In AssemblyAI's own evals (dated October 28, 2026, published on the AssemblyAI-TTS-1 launch post), AssemblyAI-TTS-1 returned its first byte of audio in 47 ms at the median in a single-client lab test. The production figure to plan around is about 90 ms to first audio. It read 93.7% of the hard targets in a 580-sentence pronunciation set correctly: phone numbers, dates, URLs and confirmation codes. On Cekura's text-to-speech benchmark, as of November 10, 2026, it ranks #1 on phone numbers and codes.

Price. $15 per 1M characters. TTS is also reachable through OpenRouter's OpenAI-compatible speech API.

Pairing it with speech-to-text. For the listening side of the same agent, AssemblyAI-STT-3.7-Realtime-Agent (assemblyai-stt-3-7-realtime-agent, $0.45/hr) is the streaming speech-to-text model built for voice agents.

Point Your OpenAI SDK at AssemblyAI

Keep your /audio/speech code. Change the base URL, the key and the model name, and your OpenAI voice names keep working.

Read the docs

ElevenLabs

ElevenLabs' docs describe Eleven v4 Turbo (eleven_v4_turbo) as available through the Text to Dialogue WebSocket. If you leave tts-1 for it:

  • No OpenAI-compatible speech endpoint is documented. The OpenAI compatibility ElevenLabs documents is for OpenAI's Realtime wire protocol on its agents product, not for /audio/speech. You rewrite the TTS call against ElevenLabs' own API or SDK.
  • Streaming text input goes over WebSocket: wss://api.elevenlabs.io/v1/text-to-dialogue/stream-input, with a multi-context version at /multi-stream-input.
  • Barge-in is handled on the multi-context socket with close_context. ElevenLabs' docs describe it as flushing the context's remaining audio.
  • Telephony formats ulaw_8000 and alaw_8000 are listed on the HTTP Create speech endpoint.
  • Languages: the docs say the Eleven v4 family supports 90+ languages.
  • Concurrency limits are set per plan.

Pricing: elevenlabs.io/pricing/api. For a phone-agent-specific view, see ElevenLabs alternatives for phone agents.

Cartesia

Cartesia's current model is Sonic 3.6 (sonic-3.6). If you leave tts-1 for it:

  • No OpenAI-compatible speech endpoint is documented. You rewrite the TTS call against Cartesia's API or SDK.
  • Streaming text input goes over wss://api.cartesia.ai/tts/websocket, sending each input on a context with continue: true.
  • Barge-in uses {"context_id": ..., "cancel": true}, which Cartesia's docs say stops further messages for that context.
  • Telephony formats: pcm_mulaw and pcm_alaw at an 8000 Hz sample rate.
  • Languages: 44, per Cartesia's model docs.
  • Concurrency is measured as the number of unique active contexts, with limits per plan.

Pricing: cartesia.ai/pricing. For a head-to-head on voice agent workloads, see AssemblyAI vs Cartesia for voice agents.

Inworld

Inworld's docs list inworld-tts-2 as its flagship model and inworld-tts-2-flash as its lowest-latency model. If you leave tts-1 for it:

  • An OpenAI-compatible speech endpoint is documented, at POST /v1/audio/speech under the base URL https://api.inworld.ai/v1. OpenAI names aren't mapped, though: per Inworld's docs, OpenAI model names such as tts-1 return 400, and OpenAI voice names such as alloy return 404. You change both the model and the voice in every request.
  • Streaming text input goes over wss://api.inworld.ai/tts/v1/voice:streamBidirectional.
  • Barge-in uses cancelContext, which aborts the synthesis in progress and discards the text.
  • Telephony formats: MULAW and ALAW, defaulting to 8,000 Hz.
  • Languages: Inworld's multilingual page lists 470 languages across three support tiers, with the third tier marked best-effort.
  • Concurrency is set per plan as concurrent generations.

Pricing: inworld.ai/pricing.

Rime

Rime's docs call Coda (coda) its flagship TTS model. Requests that don't set a model go to Mist v3, so set coda explicitly. If you leave tts-1 for it:

  • No OpenAI-compatible speech endpoint is documented. Rime's OpenAI migration guide is for Prism, its speech-to-speech product, and moves from OpenAI's Realtime API, not from /audio/speech.
  • Streaming text input goes over wss://coda.api.rime.ai/ws, with start, text and end messages.
  • Barge-in uses a cancel message on the turn's context, answered with cancelled.
  • Telephony formats: μ-law as audio/pcmu at 8000 Hz. A-law isn't documented.
  • Languages: ten in the cloud API.
  • Concurrency depends on your plan, separate from Rime's limit on open WebSocket connections.

Pricing: rime.ai/pricing.

Stay on OpenAI: gpt-realtime-2.1-mini

This is OpenAI's recommended path, and it's a bigger change than the name suggests. You move off /v1/audio/speech and onto the Realtime API, where the model runs as a session over WebRTC, WebSocket or SIP. The Realtime API does have voice-agent features: audio/pcmu output for telephony, and a response.cancel event to stop an in-progress response.

What you give up is the separate TTS stage. In a pipeline of speech-to-text, then your LLM, then TTS, each stage is a part you can choose, log and swap. gpt-realtime-2.1-mini is itself a reasoning model with tool use, so moving to it means either letting it generate the replies or reworking how your own LLM's text reaches it. If your agent's logic lives in another LLM, plan for that work. For other options on this side of the decision, see our roundup of OpenAI Realtime API alternatives.

Self-host an open-weight model

If audio and text can't leave your infrastructure, the remaining option is an open-weight TTS model on your own GPUs, sometimes behind a server that exposes an OpenAI-style /audio/speech route so your client code barely changes. Every column in the table then depends on the model and server you pick: streaming input, cancel, G.711 output and languages are all yours to build or verify. So are latency under load, scaling and on-call.

What changes when you switch from tts-1

Whichever provider you pick, the same handful of things break or behave differently. This is the summary; the migration guide has before-and-after code for the AssemblyAI move.

  • Model name. No provider in the table accepts tts-1. AssemblyAI and Inworld both refuse it with 400.
  • Voice names. OpenAI's tts-1 voices are alloy, ash, coral, echo, fable, onyx, nova, sage and shimmer. AssemblyAI accepts all of them and maps each to the language's default voice. Inworld returns 404 for them. Elsewhere, you're on a native API with the provider's own voice IDs.
  • Audio formats. OpenAI's default is MP3, and it also offers Opus, AAC and FLAC. AssemblyAI's compatible route refuses opus, aac and flac. For a voice agent, ask for raw PCM, or 8 kHz μ-law or A-law on a phone line, wherever you land.
  • speed and instructions. OpenAI's gpt-4o-mini-tts takes instructions to steer tone. AssemblyAI ignores instructions and accepts speed only at 1. Check how your new provider treats both before you rely on them.
  • Word timings. If you highlight text in sync with playback, check which route returns per-word timings. On AssemblyAI they come from the native WebSocket and the word-timings HTTP route, not the OpenAI-compatible route.
  • Connection keep-alive. The OpenAI SDKs close idle connections quickly: the Python SDK after 5 seconds, and the JavaScript SDK on Node after 4. A voice agent with gaps between replies then pays a fresh TCP and TLS handshake before almost every reply. Raise the keep-alive in your HTTP client, whichever provider you choose.

How to choose

Match your situation to a starting point:

If you need… Start with Why
The smallest code change An OpenAI-compatible route, such as AssemblyAI's Change the base URL, key and model name; OpenAI voice names keep working on AssemblyAI
A live phone agent A native WebSocket API with 8 kHz μ-law or A-law Streaming text in, cancel for barge-in, and carrier-ready audio on one connection
OpenAI's stack end to end The Realtime API with gpt-realtime-2.1-mini OpenAI's own replacement, but a different workflow with no separate TTS stage
Data that can't leave your infrastructure A self-hosted open-weight model You control where audio runs, and you own scaling and latency

For a broader shortlist that isn't framed around OpenAI, our list of the top text-to-speech APIs covers more providers.

The January 6 date is fixed, but the work it forces isn't the same for everyone. A team that only needs a base URL change can be done in an afternoon. A team moving to the Realtime API is changing its architecture. Find out which one you are this month, test your top two candidates on your own script, and leave time to load-test before the shutdown.

Replace tts-1 Before January 6

Sign up free, run AssemblyAI-TTS-1 against your own voice agent, and have your TTS stage moved well before the shutdown.

Sign up free

Frequently Asked Questions (FAQs)

When does OpenAI shut down tts-1 and gpt-4o-mini-tts?

OpenAI shuts down tts-1, tts-1-hd and the gpt-4o-mini-tts snapshots (2025-03-20 and 2025-12-15) on January 6, 2027. OpenAI announced the deprecation on October 1, 2026, and recommends gpt-realtime-2.1-mini as the replacement. Existing Speech API integrations keep working until the shutdown date.

What is the best alternative to OpenAI TTS for voice agents?

The best alternative depends on how much you want to change. If you want to keep your /audio/speech code, pick a provider with an OpenAI-compatible endpoint, such as AssemblyAI-TTS-1. If you're building a live phone agent, prioritize a WebSocket API with streaming text input, a cancel message for barge-in and 8 kHz μ-law output. If you want OpenAI end to end, the Realtime API with gpt-realtime-2.1-mini is OpenAI's own path.

Which text-to-speech APIs are OpenAI-compatible?

Among the providers compared here, AssemblyAI and Inworld document an endpoint that answers OpenAI's /audio/speech request shape. ElevenLabs, Cartesia and Rime don't document one for text-to-speech. If you self-host an open-weight model, whether you get an OpenAI-style route depends on the server you run it behind.

Can I keep my OpenAI voice names, like alloy and nova, after switching?

It depends on the provider. AssemblyAI's OpenAI-compatible route accepts every OpenAI voice name and speaks each one with the default voice of the request's language, aai-kyla in English. Inworld's compatible endpoint returns 404 for OpenAI voice names, and providers without a compatible endpoint use their own voice IDs. Either way, you'll pick a new voice by its provider ID if you want something other than the default.

Is OpenAI TTS free?

No. OpenAI's API pricing page lists tts-1, tts-1-hd and gpt-4o-mini-tts as paid, usage-billed models: tts-1 and tts-1-hd per million characters, and gpt-4o-mini-tts per token. After January 6, 2027, none of the three will be available at any price.

Does gpt-realtime-2.1-mini work with the /audio/speech endpoint?

No. OpenAI's TTS guide says gpt-realtime-2.1-mini isn't a supported model for /v1/audio/speech. It runs on the Realtime API over WebRTC, WebSocket or SIP, so moving to it is a change of API workflow, not a model-name swap.