ElevenLabs Text to Speech Streaming: Live Audio API

Yes. ElevenLabs text to speech streaming sends audio back in chunks as it is generated, so playback can start in about 75 milliseconds instead of waiting for the full file. The official docs cover it in the Streaming text to speech how-to guide and through the WebSocket API. Plans start at $6/month for Starter with 30,000 credits (checked 5 Nov 2026).

Streaming powers live use cases like the Scribe V2 realtime guide for transcription; if chunks fail, the troubleshooting guide helps; for finished files see the WAV output guide, and for video the Dubbing V2 guide.

How streaming text to speech works

Standard text to speech waits until the entire audio file is ready, then returns it. Streaming changes the delivery: the model emits audio in small chunks, and the API sends each chunk the moment it exists. Your player starts playback with the first chunk while later chunks are still being generated. The official docs describe this in the Streaming text to speech how-to guide, and the result is latency as low as 75 milliseconds from request to first audio – fast enough for conversational voice agents where a half-second pause feels broken.

Two interfaces support streaming. The WebSocket API at the stream-input endpoint keeps one connection open and accepts text inputs over time, which suits continuous conversation. The REST endpoint also supports chunked transfer for single requests. Model choice matters: Eleven v4 Turbo is the low-latency option at around 100 milliseconds, while the full Eleven v4 trades a little speed for maximum quality. Our model comparison breaks down which generation to pick for speed versus quality.

How to stream audio, step by step

Streaming is a developer feature, so these steps assume you have an API key. Credit costs are identical to non-streaming generation: you pay per character generated, not per chunk or per connection.

1. Get an API key. Create one in your ElevenLabs profile under API keys. The same xi-api-key header authenticates streaming endpoints. Cost: free to create; keys do not cost credits.
2. Pick a low-latency model. Choose Eleven v4 Turbo for the fastest first-chunk time (around 100ms) or Eleven v4 for top quality. Cost: billed per character at the model rate – about 1 credit per character on v4, about 0.5 on Turbo and Flash models.
3. Open the WebSocket. Connect to the stream-input endpoint with your key, voice id and voice settings. Send a JSON payload with your text. Cost: the connection itself is free.
4. Play chunks as they arrive. Audio chunks stream back as base64-encoded MP3. Feed them to your player immediately instead of buffering the whole response. A 500-character request costs about 500 credits on v4 regardless of chunking.
5. Close or reuse the socket. For one-shot requests, close after the final chunk. For agents, keep the socket open and send new text inputs as the conversation continues. Cost: only the characters you generate.

Latency numbers that matter

The headline figure is 75 milliseconds from request to first audio chunk on the WebSocket path, with Eleven v4 Turbo documented around 100 milliseconds. Those numbers assume a nearby region and a short prompt; longer text and distant regions add time. What matters for user experience is time-to-first-audio, not total generation time: a 30-second narration still takes 30 seconds to fully generate, but the listener hears the first word almost instantly. That is why streaming is the default choice for agents, live dubbing and interactive stories, and overkill for rendering a podcast episode.

To keep latency low in production, send text in sentence-sized pieces rather than one giant block, pick the closest API region, and use v4 Turbo unless you have a quality reason not to. Player-side buffering of one or two chunks smooths out network jitter without adding noticeable delay. Measure time-to-first-audio in your own setup because your network path dominates the numbers once the model is fast enough.

Streaming use cases in detail

Conversational voice agents are the flagship use case. A user speaks, your app transcribes, a language model replies in text, and streaming TTS speaks the reply starting in about 75 milliseconds. Without streaming, every turn would add a full render delay and the conversation would feel sluggish. Customer support bots, phone agents and in-game characters all depend on this tight loop.

Live narration covers sports commentary, event hosting and real-time translation, where the script is written as the event unfolds. Streaming lets the voice keep pace with the action instead of batching commentary after the fact. Accessibility tools are a third pillar: screen readers and reading assistants that stream text as the user scrolls feel dramatically more responsive than ones that pause to render each paragraph.

Streaming versus standard generation

What streaming costs

Streaming does not change the price of audio. You pay the same per-character rate as standard generation on the model you choose: about 1 credit per character on Eleven v4, about 0.5 credits per character on Turbo and Flash models. A 2,000-character agent response costs about 2,000 credits on v4 whether it streams or not. The plans below show monthly allowances; the free tier 10,000 credits is enough to prototype a streaming integration.

Who should NOT use streaming

Skip streaming if your end product is a file. Podcast episodes, audiobook chapters, video voiceovers and training modules are all rendered once and edited afterward – standard generation gives you one clean MP3 with less code. Skip it too if you have no developer resources: streaming requires handling WebSocket connections, base64 chunks and playback buffering, which is real engineering work. And if your audience is on unreliable networks, the chunked delivery can stutter; a pre-rendered file plays smoothly everywhere.

Frequently asked questions

How fast is ElevenLabs text to speech streaming?

About 75 milliseconds from request to first audio on the WebSocket path, with Eleven v4 Turbo documented around 100 milliseconds. That is time-to-first-audio; the full generation still takes as long as the audio itself. Actual latency varies with text length, region and network.

Does streaming cost more credits?

No. Streaming bills at the same per-character rate as standard generation on the same model: about 1 credit per character on Eleven v4 and about 0.5 on Turbo and Flash models. You pay for characters generated, not for chunks, connections or speed.

What is the WebSocket endpoint for streaming?

The documented endpoint is the stream-input path under /v1/text-to-speech/{voice_id}/ on the WebSocket host. It accepts JSON text inputs over a persistent connection and returns base64 audio chunks. Authenticate with your xi-api-key, the same key used for REST generation.

Which model is best for streaming?

Eleven v4 Turbo is the best default for streaming because it is tuned for low latency at around 100ms. Use the full Eleven v4 when quality matters more than shaving milliseconds, such as narrated content delivered live. Flash models are the cheapest per character if cost dominates.

Can I stream from the ElevenLabs website?

Streaming is an API feature, not a website button. The web interface generates complete files. To stream, use the WebSocket or chunked REST endpoints with an API key. Non-developers who need live audio should look at Conversational AI agents, which handle streaming internally.

Does streaming work on the free plan?

Yes, within API access limits. The free tier includes 10,000 credits per month, and streaming consumes credits at the normal per-character rate. That is enough to build and test a streaming integration before upgrading to Starter at $6/month for 30,000 credits.

How do I handle interruptions in a voice agent?

Because the WebSocket stays open, you can stop sending text and start a new input when the user interrupts. Discard buffered chunks for the old reply and begin streaming the new one. Good agents also send a stop signal so the player does not finish the stale audio.

Is streaming available in all regions?

The API is global, but latency depends on your distance to the serving region. Pick the closest region your integration supports and measure time-to-first-audio from your actual deployment. If your users are far from the region, a small player-side buffer prevents stutter.

Sources checked

• elevenlabs.io/docs – Streaming text to speech how-to guide, WebSocket API reference and latency figures, checked 5 Nov 2026.
• elevenlabs.io/pricing – plan prices and monthly credit allowances, checked 5 Nov 2026.