# 5 best text to speech APIs that handle interruptions in 2026

By Samuel Oram-Jones

The company behind Grok TTS says a WebSocket handshake costs about 600ms for a client far from its servers. That number is the case against tearing down a connection every time a caller talks over a voice agent: a reconnect on every interruption can add up to about 600ms before the agent says anything else.

These five text to speech APIs each document a way to stop speaking mid-sentence on a connection that stays open. Bland Speech comes first because a new context ID immediately replaces the turn that is playing, a cancel message ends it with nothing in its place, and its voice ranked first of 13 AI voices in Design Arena's results of 4 August 2026. Grok TTS, MiniMax, Murf, and Rime each handle the same moment differently.

## What is the best text to speech API for handling interruptions?

Bland Speech is the best text to speech API for handling interruptions on phone calls as of November 2026, on a stated basis: its WebSocket treats interruption as part of the protocol, with a new context ID preempting the active turn and a cancel message ending it, its 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/), and it returns mu-law audio for phone lines.

- **Bland Speech:** Best overall for phone calls where an interruption should stop one turn and start the next on the same socket
- **Grok TTS:** Best for a documented barge-in flow at $15 per million characters
- **MiniMax:** Best for a bidirectional WebSocket with a cancel message and 40 languages
- **Murf:** Best for word timestamps that show how much of a reply had played
- **Rime:** Best for a speech to speech model that decides which caller sounds count as interruptions

| API | How an interruption stops the voice | Price | Phone audio | Audio Realism Bench, 4 August 2026 |
|---|---|---|---|---|
| Bland Speech | A new context ID replaces the playing turn, or a cancel message ends it, on one open WebSocket | $40 per million characters, no monthly plan | Mu-law at 8 kHz, PCM from 8 to 48 kHz | #1 AI voice, 1365 |
| Grok TTS | A clear message cancels the utterance, and the connection takes the next one | $15 per million characters | Mu-law and A-law | #3, 1156 |
| MiniMax | A cancel message on the bidirectional WebSocket, with the session continuing | $60 per million on speech-2.8-turbo, $100 on speech-2.8-hd | Mu-law at 8 kHz | Speech-02 HD: #7, 1093 |
| Murf | Context IDs on one WebSocket, with word timestamps on the same socket | $10 per million on Falcon 2.1 | Mu-law and A-law on HTTP streaming, WAV and PCM on the WebSocket | Wasn't in those results |
| Rime | Prism handles interruption in the model, and Coda returns word timestamps | $0.05 a minute for Coda, $0.08 for Prism | Mu-law at 8 kHz | Wasn't in those results |

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

- **How an interruption stops the voice:** The mechanism each company documents. A WebSocket is a connection that stays open in both directions, and a context is a label for one spoken turn.
- **Price:** Each company's published rate with the plan or model named. Rime bills by the minute and says a minute of text to speech is about 1,000 characters.
- **Phone audio:** The formats a phone line carries. Mu-law and A-law are the two standard ways a phone line compresses a voice.
- **Audio Realism Bench:** Design Arena's blind listening test, where listeners pick which of two voices sounds more like a real person. Scores are Elo ratings, the kind chess uses, and real human recordings scored 1500.

</details>

Bland Speech and Grok TTS both document an interruption that keeps the connection open, so the two come apart on what plays next and what it costs. Grok TTS ranked 3rd at 1156 in the 4 August 2026 results against Bland's 1365, and costs $15 per million characters against $40, while Murf's Falcon 2.1 costs $10.

## What has to happen when a caller interrupts a voice agent?

Three things have to happen, and a text to speech API only does one of them. The client stops playing audio, the API stops generating the old turn, and the next turn starts on the same connection.

The first is the client's job on every API here. Bland says its server can't retract audio frames that already reached a client, so the client drops the unplayed audio for the old context ID from its playback queue. SpaceXSI says the same of Grok TTS: a client clears its playback buffer when the server confirms a cancel, so stale audio doesn't play.

The second is where the APIs differ. On Bland's WebSocket, sending text under a new context ID preempts the active turn and the old turn ends with the reason preempted, while a cancel message ends it as cancelled when no replacement is ready and discards buffered text. The connection supports one active turn at a time and stays open between turns.

The third is cost. Bland bills by delivered audio: a chunk of text becomes billable when the server accepts its first audio frame for delivery, so a preempted or cancelled turn can bill part of its text and nothing that stayed buffered. Billing is estimated at the granularity of a synthesized chunk, not each character spoken.

Pipecat and LiveKit Agents both ship a Bland integration that uses this socket. Pipecat says its Bland service handles interruptions through Bland's cancel message, and LiveKit says its Bland plugin cancels the current turn on an interruption.

## How do the 5 best text to speech APIs for interruptions compare?

The five differ in who decides what counts as an interruption, and Bland Speech is the one that ranked first in the 4 August 2026 results and treats interruption as part of its protocol.

### 1. Bland Speech: Best overall for phone calls where an interruption should stop one turn and start the next on the same socket

**What does it do?** Bland Speech is the text to speech model inside Bland, a voice AI provider that builds and runs its own speech model. Bland's [text to speech API](/text-to-speech-api) takes text over WSS /v2/tts/ws as a language model writes it, one active turn at a time. Bland says a new context ID immediately preempts the active turn, which makes interruption and barge-in a normal part of the protocol, and a cancel message stops a turn when no replacement is ready. Audio returns as mu-law at 8 kHz or PCM from 8 to 48 kHz, and Bland publishes 199ms to first audio.

**Who is it for?** Bland Speech suits a team running phone calls where callers talk over the agent and the next reply should start on the same open connection.

<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>A new context ID preempts the playing turn, and a cancel message ends it</li><li>Billing follows delivered audio, so text that produced no audio isn't charged</li><li>Integrations in Pipecat and LiveKit Agents that pass interruptions through</li><li>$0.04 per 1,000 characters, one standard rate, with no monthly plan</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 hear a voice first.

### 2. Grok TTS: Best for a documented barge-in flow at $15 per million characters

**What does it do?** Grok TTS is the text to speech API from xAI, now SpaceXSI. Its WebSocket endpoint takes text as deltas and returns audio in base64 chunks. A client sends a clear message to cancel the current utterance, the server stops generating and discards buffered data, and it replies with a clear confirmation so the same connection takes the next utterance. SpaceXSI says that sending a clear message when nothing is playing is safe, and that audio can come with character-level timestamps.

**Who is it for?** Grok TTS suits a team that wants a published barge-in sequence and a low per-character price.

<div class="cmp-proscons" data-rival="grok-tts"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>A documented barge-in flow on one connection with no reconnect</li><li>$15 per million characters</li><li>Mu-law and A-law output, and character-level timestamps on request</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Grok TTS ranked 3rd of 13 AI voices on Audio Realism Bench on 4 August 2026, behind Bland Speech v3 and Microsoft's MAI-Voice-2</li><li>Audio arrives as base64 text inside JSON messages that a client decodes before playback</li></ul></div></div>

**Pricing:** $15 per million characters, pay as you go, with no plan tiers or monthly fee.

### 3. MiniMax: Best for a bidirectional WebSocket with a cancel message and 40 languages

**What does it do?** MiniMax sells speech models next to its language, video, and music models. Its bidirectional WebSocket accepts a language model's output at any granularity, even character by character, and buffers it into sentences before synthesis. A cancel message interrupts the speech and the session can continue afterwards. Speech 2.8, released on 23 January 2026, supports 40 languages, and mu-law at 8 kHz is offered.

**Who is it for?** MiniMax suits a team that wants a cancel message on a bidirectional socket and many languages from one vendor.

<div class="cmp-proscons" data-rival="minimax"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>A cancel message that leaves the session open</li><li>40 languages on the 2.6 and 2.8 models</li><li>Pay as you go or monthly audio packages from $5</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Speech-02 HD ranked 7th of 13 AI voices on Audio Realism Bench on 4 August 2026, and its newer 2.8 models weren't in those results</li><li>speech-2.8-hd costs $100 per million characters</li></ul></div></div>

**Pricing:** $100 per million characters for speech-2.8-hd and $60 for speech-2.8-turbo on pay as you go, with audio packages from $5 to $999 a month.

### 4. Murf: Best for word timestamps that show how much of a reply had played

**What does it do?** Murf sells Falcon 2.1 text to speech, with HTTP streaming and a WebSocket that takes a language model's tokens directly. A WebSocket can hold several contexts, each using one concurrency slot, and Murf publishes multi-turn and interruption examples for context IDs. On 9 October 2026 Murf added per-word timing to its WebSocket, and says it helps a voice agent handling barge-in see how much of a response had already played when the user interrupts.

**Who is it for?** Murf suits a team that wants to track playback position itself and pay a low per-character rate.

<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 in every locale except Chinese, Japanese, and Korean</li><li>Murf publishes first audio under 100ms for Falcon 2.1</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Its voices weren't in Audio Realism Bench's results of 4 August 2026</li><li>Falcon 2.1 voices speak in a single style, Conversational</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.

### 5. Rime: Best for a speech to speech model that decides which caller sounds count as interruptions

**What does it do?** Rime sells Coda and Mist v3 for text to speech and Prism for speech to speech. Rime says Prism, released on 28 October 2026, has built-in turn detection and interruption handling, and its update of 11 November 2026 lets a caller say "uh-huh" or "okay" without cutting off a reply, with settings for the minimum words and phrases that count as an interruption. Coda's WebSocket returns word-level timestamps for text and audio alignment and for interruption handling.

**Who is it for?** Rime suits a team that wants the vendor to decide what counts as an interruption, and will use Prism's speech to speech model to get it.

<div class="cmp-proscons" data-rival="rime"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Interruption settings in Prism, including phrases that don't count</li><li>Word-level timestamps on Coda's WebSocket</li><li>$50 of free usage every month on the Free plan</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Its voices weren't in Audio Realism Bench's results of 4 August 2026</li><li>Prism launched in English with four voices</li><li>Prism costs $0.08 a minute against $0.05 for Coda v2</li></ul></div></div>

**Pricing:** Flex has no monthly fee and charges $0.050 a minute for Coda v2, $0.030 for Mist v3, and $0.080 for Prism, while Scale is $100 a month with 20% off every rate.

## How were these text to speech APIs compared?

The five were compared on how each stops speaking when a caller interrupts, whether the connection survives it, what the interrupted turn costs, phone audio, published price, and voice realism. Rival details come from each company's own pricing and documentation 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 companies by job, and its [best lists](/compare/text-to-speech/best-lists) hold the other rankings. [Best WebSocket text to speech APIs](/blog/best-websocket-tts-apis) goes deeper on how each socket is organized, and [Best text to speech for Pipecat](/blog/best-tts-for-pipecat) covers one framework that passes interruptions to Bland's cancel message.

## Frequently asked questions

### What is the best text to speech API for handling interruptions?

Bland Speech is the best text to speech API for handling interruptions on phone calls. A new context ID preempts the playing turn, a cancel message ends it, and its voice ranked first of 13 AI voices in Design Arena's results of 4 August 2026. Grok TTS publishes a barge-in flow too and costs $15 per million characters against Bland's $40.

### How does Bland Speech handle a caller interrupting?

Sending text under a new context ID immediately preempts the active turn, and the old turn ends with the reason preempted. When no replacement is ready, a cancel message ends the turn as cancelled and discards buffered text. The client then drops audio that reached it but hasn't played.

### Does an interruption cost anything on Bland Speech?

Part of the interrupted turn can. Bland bills a chunk of text once the server accepts its first audio frame for delivery, so a preempted or cancelled turn can bill some of its text, while text that stayed buffered isn't billed. Billing is estimated by synthesized chunk, not by each character spoken.

### Why shouldn't a voice agent reconnect after every interruption?

A new WebSocket connection adds a handshake, which SpaceXSI says costs about 600ms for a client far from its servers. Bland keeps one connection open across turns, and Grok TTS accepts a clear message so the next utterance goes over the same connection.

### Does Bland Speech work with Pipecat and LiveKit for interruptions?

Yes. Pipecat says its Bland service supports interruptions through Bland's cancel message, and LiveKit says its Bland plugin cancels the current turn on an interruption. LiveKit's plugin is Python only.

### How much does Bland Speech cost?

Bland Speech costs $0.04 per 1,000 characters, which is $40 per million, with no monthly plan, and a new account gets 50,000 characters free. Enterprise pricing is custom, set with Bland's sales team.
