# 6 best WebSocket text to speech APIs in 2026

By Samuel Oram-Jones

A WebSocket text to speech connection is state a team has to look after: what each message can say, how a turn ends, how long a session lives, and what audio comes back. Gradium caps a session at 3,000 seconds, or 7,200 on its largest plan, and IBM closes its socket after every request, which are the details that decide whether a voice agent runs cleanly across a thousand calls.

These six APIs take text over a WebSocket and are ranked on how well a socket fits a phone conversation: one connection per call, interruption inside the protocol, and audio a phone line carries. A socket that multiplexes several streams is a different job, and Gradium and Murf are the entries built for it.

## What is the best WebSocket text to speech API?

Bland Speech is the best WebSocket text to speech API for phone calls as of November 2026, on a stated basis: a socket carries one conversation, a new context ID or a cancel message interrupts the turn that is playing, audio returns as mu-law for phone lines, and the voice ranked first of 13 AI voices in Design Arena's [Audio Realism Bench results of 4 August 2026](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/).

- **Bland Speech:** Best overall for one phone conversation per socket
- **Cartesia:** Best for a socket organized in contexts, with a first-audio claim under 90ms
- **ElevenLabs:** Best for a range of models split across two sockets
- **Gradium:** Best for several requests over one socket
- **Murf:** Best for a $10-per-million rate, with word timestamps on the socket
- **Smallest AI:** Best for sentence buffering through a shared context ID

| API | How a socket is organized | Phone audio | Price per million characters |
|---|---|---|---|
| Bland Speech | One conversation per socket, one active turn, a new context ID interrupts | Mu-law at 8 kHz, PCM from 8 to 48 kHz | $40 |
| Cartesia | Contexts, with partial text sent as continuations | Mu-law and A-law at 8 kHz | $37.38 to $50 in a plan |
| ElevenLabs | A stream-input socket, with v3 and v4 routed through Text to Dialogue | ulaw_8000 and alaw_8000 | $40 on v4 Turbo, $80 on v4 |
| Gradium | Several requests over one socket, matched by request ID | Mu-law and A-law at 8 kHz | $36 to $58 per million credits |
| Murf | Several contexts per socket, one slot each | Mu-law and A-law on HTTP streaming, WAV and PCM on the WebSocket | $10 on Falcon 2.1 |
| Smallest AI | Fragments joined by a shared context ID | Mu-law at 8 kHz | $17.50 to $19.50 |

<details><summary>What these rows mean</summary>

- **How a socket is organized:** A WebSocket is a connection that stays open in both directions. A context is a label that groups the text for one spoken turn, so a client can start a new turn or cancel the old one.
- **Phone audio:** The formats each company documents for speech it returns, where mu-law and A-law are the two standard ways a phone line compresses a voice.
- **Price per million characters:** Each company's published rate with the plan or model named. Gradium and Cartesia bill credits inside monthly plans, and Smallest AI prices per 10,000 characters, shown here per million.

</details>

Cartesia publishes a first-audio claim under 90ms against Bland's published 199ms, and Murf's rate of $10 per million characters is a quarter of Bland's $40. What Bland Speech v3 brings against those is its score of 1365 in the 4 August 2026 results, while Cartesia's Sonic 3.5 ranked 6th at 1096 and Murf's voices weren't in them.

## What should a buyer check in a WebSocket text to speech API?

A buyer should check the lifetime of the connection first. Bland's socket stays open for a whole conversation, across turns, and Gradium's Free plan caps a session at 1,500 characters, about two minutes of speech.

The message vocabulary comes next. Bland's socket takes an init message that fixes the voice and audio format for the life of the connection, then speak messages carrying text deltas under a context ID, an end-of-turn message that flushes buffered text, a cancel message, and a close message. Bland caps one turn at 4,000 characters.

Browsers need a plan for keys. Bland's guidance is that a backend mints a short-lived stream token, valid for five minutes, so an API key never reaches the page.

Phone audio is the last check, because a socket that returns only PCM or WAV leaves a team converting audio. Bland's socket returns mu-law at 8 kHz or PCM from 8 to 48 kHz.

## Which WebSocket text to speech APIs work differently?

Several APIs use a WebSocket in ways that change how a team builds. IBM Watson Text to Speech takes one request message of up to 5 KB of text over its WebSocket and closes the connection when the audio is done, so each request gets its own connection, with word timings available. Neuphonic closes each piece of text with a stop token, and its WebSocket has a 60 second timeout. Azure's text streaming runs over WebSocket v2 through its Speech SDK in C#, C++, and Python. Google documents bidirectional streaming through its client libraries' StreamingSynthesize call, and OpenAI keeps WebSocket and WebRTC on its Realtime API instead of its speech endpoint.

Mistral's Voxtral TTS offers a realtime WebSocket session with two modes, responsive for lower latency and long-form for quality on longer text, and one session can run several generations in sequence without reconnecting.

## How do the 6 best WebSocket text to speech APIs compare?

The six differ in how a socket is organized and how many calls it can carry, and Bland Speech was built for one phone call per socket and ranked first of 13 AI voices in the 4 August 2026 results.

### 1. Bland Speech: Best overall for one phone conversation per socket

**What does it do?** Bland Speech is the text to speech model inside Bland, a voice AI provider that runs its own speech model for enterprise phone calls. Bland's [text to speech API](/text-to-speech-api) keeps one connection open for a conversation over WSS /v2/tts/ws. A speak message carries a text delta under a context ID, and Bland buffers the deltas into speech segments and returns binary audio before the turn is complete. Sending speak with a new context ID preempts the active turn, and a cancel message ends it with no replacement. Pipecat's Bland service uses this socket and keeps it open across turns, and LiveKit Agents' Bland plugin streams over it by default.

**Who is it for?** Bland Speech suits teams running a voice on a phone call, one socket per caller.

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>Interruption is part of the protocol, through a new context ID or a cancel message</li><li>Mu-law at 8 kHz and PCM from 8 to 48 kHz</li><li>Billing follows delivered audio, so buffered text that never produced sound isn't charged</li><li>$0.04 per 1,000 characters, one standard rate</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>More than a hobby project needs, because Bland Speech is built for production calls</li></ul></div></div>

**Pricing:** $0.04 per 1,000 characters, which is $40 per million, with no monthly plan and 50,000 characters free for a new account. Enterprise pricing is custom, set with Bland's sales team. The [Bland Speech](/speech) page has a studio to try a voice.

### 2. Cartesia: Best for a socket organized in contexts, with a first-audio claim under 90ms

**What does it do?** Cartesia sells Sonic, and Sonic 3.6 was released on 27 August 2026. Its WebSocket works in contexts, and Cartesia says that when the full transcript isn't known up front a developer should use the WebSocket with contexts, not its bytes or server-sent events endpoints. Cartesia says first audio arrives in under 90ms.

**Who is it for?** Cartesia suits teams that want a context-based socket and a small published first-audio figure.

<div class="cmp-proscons" data-rival="cartesia"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>A WebSocket organized in contexts</li><li>Cartesia says first audio arrives in under 90ms</li><li>Mu-law and A-law at 8 kHz</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Sonic 3.5 ranked 6th of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>Credits from a monthly plan</li></ul></div></div>

**Pricing:** 1 credit per character. Pro is $5 a month for 100,000 credits, Startup $49 for 1.25 million, and Scale $299 for 8 million, with overage at $65, $45, or $38 per million.

### 3. ElevenLabs: Best for a range of models split across two sockets

**What does it do?** ElevenLabs sells text to speech, transcription, and voice agents. Its text to speech WebSocket sits at /v1/text-to-speech/{voice_id}/stream-input and doesn't accept Eleven v3 or v4, which run through Text to Dialogue instead. Eleven v4 Turbo, released on 28 September 2026, has a median latency of about 100ms excluding application and network time.

**Who is it for?** ElevenLabs suits teams that want ElevenLabs' range of models and will manage which model runs on which endpoint.

<div class="cmp-proscons" data-rival="elevenlabs"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>v4 Turbo at $0.04 per 1,000 characters</li><li>90+ languages on the Eleven v4 family</li><li>ulaw_8000 and alaw_8000 output</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Eleven v3 ranked 12th of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>Eleven v3 and v4 need a different route than the standard WebSocket</li></ul></div></div>

**Pricing:** $0.08 per 1,000 characters for Eleven v4, v3, and Multilingual v2, and $0.04 for v4 Turbo and Flash, against plans from Free to Business at $990 a month.

### 4. Gradium: Best for several requests over one socket

**What does it do?** Gradium sells real-time voice AI, with text to speech, speech to text, and translation. Its WebSocket takes text token by token, and several requests can run over one socket, matched by a client request ID. Gradium's engine recognizes a flush tag and break tags in the text, and output includes mu-law and A-law at 8 kHz.

**Who is it for?** Gradium suits teams that want to run several streams over one connection and are willing to manage session lifetimes.

<div class="cmp-proscons" data-rival="gradium"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Multiplexing over a single WebSocket</li><li>A median first-audio figure of 49ms on a Speko benchmark</li><li>Mu-law and A-law at 8 kHz</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>A session lasts up to 3,000 seconds, or 7,200 on the largest plan</li><li>The Free plan caps a session at 1,500 characters</li><li>Its voices weren't in Audio Realism Bench's results of 4 August 2026</li></ul></div></div>

**Pricing:** Credit plans from Free, with 45,000 credits, to XS at $12 a month, S at $39, M at $312, and L at $1,480, at $58 down to $36 per million credits.

### 5. Murf: Best for a $10-per-million rate, with word timestamps on the socket

**What does it do?** Murf sells Falcon 2.1 text to speech with HTTP streaming and a WebSocket. Murf says to switch off a framework's sentence tokenizer and send a language model's tokens straight to the socket, because Falcon splits sentences itself. A WebSocket can hold several contexts, each using one concurrency slot, and word timestamps came to the WebSocket on 9 October 2026. Murf publishes first audio under 100ms for Falcon 2.1.

**Who is it for?** Murf suits teams that want a low per-character rate and word timings on the same connection.

<div class="cmp-proscons" data-rival="murf"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>$0.01 per 1,000 characters for Falcon 2.1</li><li>Word timestamps on the WebSocket</li><li>Regional endpoints beside a global router</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Falcon 2.1 voices speak in a single style, Conversational</li><li>Its voices weren't in Audio Realism Bench's results of 4 August 2026</li></ul></div></div>

**Pricing:** $0.01 per 1,000 characters for Falcon 2.1 on pay as you go, with a $10 free trial credit.

### 6. Smallest AI: Best for sentence buffering through a shared context ID

**What does it do?** Smallest AI sells Lightning text to speech. Its WebSocket carries fragments tagged with a shared context ID, which the server buffers and joins at sentence boundaries into one continuous generation, and per-word timing events arrive with the audio. Smallest AI says sub-100ms latency and gives about 200ms time to first byte at 40 concurrent requests over a WebSocket. Output includes mu-law at 8 kHz.

**Who is it for?** Smallest AI suits teams that want sentence joining handled on the server at about $17.50 per million characters.

<div class="cmp-proscons" data-rival="smallest-ai"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Fragments joined at sentence boundaries by context ID</li><li>Word timing events on the WebSocket</li><li>Mu-law at 8 kHz among its output formats</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Lightning v3.1 Pro ranked 11th of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>Its sub-100ms claim sits beside about 200ms at 40 concurrent requests</li></ul></div></div>

**Pricing:** About $17.50 per million characters for Lightning V3.1 and $19.50 for V3.1 Pro, with $10 of free credit to start.

## How were these WebSocket text to speech APIs compared?

The six APIs were compared on how a socket is organized, how it handles interruption, its text and session limits, phone audio, and published price. Rival details come from each company's own pages, each latency figure is the company's own claim and not a test Bland ran, and voice scores come from Design Arena's published results of 4 August 2026.

Bland's [comparison of text to speech APIs](/compare/text-to-speech) covers the same vendors by job, and its [best lists](/compare/text-to-speech/best-lists) hold the other rankings. [Streaming text to speech over a WebSocket, explained](/blog/streaming-tts-websocket) walks through the connection itself, and [Best text to speech for streaming LLM output](/blog/best-tts-for-llm-streaming) covers how each socket handles a language model's tokens.

## Frequently asked questions

### What is the best WebSocket text to speech API?

Bland Speech is the best WebSocket text to speech API for phone calls. One socket carries one conversation, a new context ID or a cancel message interrupts the turn that is playing, and its voice ranked first of 13 AI voices in Design Arena's results of 4 August 2026. Gradium and Murf suit a team that wants several streams on one socket.

### How does a WebSocket text to speech API work?

A client opens a WebSocket, sends text as it is written, and receives audio back on the same connection before the text is finished. On Bland's socket a client sends an init message to set the voice and audio format, speak messages with text under a context ID, and an end-of-turn message when the language model finishes.

### Does Bland Speech support WebSocket streaming?

Yes. WSS /v2/tts/ws takes text as a language model writes it, buffers it into speech segments, and returns audio before the turn ends. Pipecat and LiveKit Agents both ship a Bland integration that uses it.

### Can one WebSocket carry several text to speech streams?

On some APIs. Gradium matches several requests on one socket by a client request ID, and Murf holds several contexts on one WebSocket.

### Does Bland Speech return phone-ready audio over the WebSocket?

Yes. The init message sets the encoding to mu-law at 8 kHz or to PCM at 8000, 16000, 24000, 44100, or 48000 Hz, and the setting holds for the life of the connection.
