# 6 best text to speech APIs for streaming LLM output in 2026

By Samuel Oram-Jones

A language model writes a reply a few tokens at a time, and a text to speech API has to decide when a fragment is enough to speak. Speak too early and the voice can restart its intonation at every break, which is what Speechify says its sessions do at each segment. Wait too long and a caller sits in silence while the model finishes its sentence.

Six APIs here take text from a language model as it is written. They are ranked on how they handle that decision on a live phone call, who makes it, how a turn is closed, and whether a caller can cut one off.

## What is the best text to speech API for streaming LLM output?

Bland Speech is the best text to speech API for streaming LLM output on a phone call, because its server does the deciding: its WebSocket takes a language model's tokens as they arrive and buffers them into speech segments on the server, an end-of-turn message releases whatever is left, a new context ID or a cancel message stops a turn that a caller interrupts, 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 LLM output on a phone call
- **Inworld:** Best for sentence detection switched on with one setting
- **MiniMax:** Best for text sent at any granularity, down to single characters
- **Smallest AI:** Best for fragments joined by a shared context ID
- **Murf:** Best for sending tokens straight in with no client-side tokenizer, at $10 per million characters
- **Speechify:** Best for an explicit flush, with a documented cost in prosody

| API | Who decides when text is spoken | How the client controls it | Price per million characters |
|---|---|---|---|
| Bland Speech | The server buffers and finds speech boundaries | An end-of-turn message sends the tail | $40 |
| Inworld | The server, at sentence boundaries, with auto mode on | The auto mode setting | $15 to $25 on pay as you go |
| MiniMax | The server buffers text into sentences | A cancel message stops a turn, and sending can continue | $60 to $100 |
| Smallest AI | The server joins fragments at sentence boundaries | A shared context ID | $19.50 on Lightning v3.1 Pro |
| Murf | Falcon splits sentences itself | Tokens go straight in, with no client tokenizer | $10 |
| Speechify | The server, at a sentence end followed by whitespace | A flush or close message sends the tail | $10 to $15 at list price |

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

- **Who decides when text is spoken:** A tokenizer is code that splits text into pieces. When a server buffers incoming tokens, the client sends them as they come. When a client buffers, the client decides what counts as a sentence.
- **How the client controls it:** The message or setting a developer uses to close a turn or stop one. A flush message tells the server that no more text is coming for now.
- **Price per million characters:** Each company's published rate with the plan or model named. Speechify's figure is its standing list price for Simba 3.5 on its Starter, Pro, and Scale plans, and Smallest AI prices per 10,000 characters, shown here per million.

</details>

Bland Speech doesn't lead on price. Murf's $10 per million characters is a quarter of Bland's $40, and Inworld's TTS-2 Flash is $15. What Bland brings against those is the 4 August 2026 result of 1365, while MiniMax's Speech-02 HD ranked 7th at 1093 and the models from Inworld, Smallest AI, Murf, and Speechify were either not in those results or ranked lower.

## Who decides when a fragment of LLM text is ready to speak?

On all six APIs here the server does the buffering, and they differ in how much a client can steer it. Bland accumulates incoming deltas, detects speech boundaries, and releases text to synthesis when it has enough context for natural output, and Bland's guidance is that a client needn't wait for complete sentences, build a character threshold or a timer, or send a flush after every fragment. One catch is that very short or unfinished text can stay buffered without producing audio, which is why Bland asks for an end-of-turn message when the model finishes. Another is that each text delta is appended exactly as sent, so a client keeps the spaces and punctuation from its language model's stream.

Other APIs leave more to the client. Neuphonic closes a piece of text with a stop token, NVIDIA's Riva realtime API collects text in a buffer until a commit event processes it, Sarvam exposes a minimum buffer length for flushing, and Gradium recognizes a flush tag inside the text. Speechify's server speaks one segment at a time, usually a sentence or a few, and a client sends a flush or close message to have the last unfinished sentence spoken.

The cost of a split is audible. Speechify states that the voice's intonation and pacing start afresh at each segment, so a pause or pitch change can fall between two sentences that a single request would have joined smoothly.

## Which text to speech APIs can't take LLM output in pieces?

OpenAI's speech endpoint takes the whole text in one request and streams audio back over chunked HTTP, and its WebSocket and WebRTC connections sit on the Realtime API instead. The speech endpoint's models, tts-1, tts-1-hd, and gpt-4o-mini-tts, are due to shut down on 6 January 2027. Speechmatics streams audio out and says it doesn't offer full bidirectional streaming yet, with support planned.

## How do the 6 best text to speech APIs for streaming LLM output compare?

The six differ in where the buffering decision lives and what a client can do about it.

### 1. Bland Speech: Best overall for LLM output on a phone call

**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. WSS /v2/tts/ws keeps one connection open for a conversation. A speak message carries each text delta under a context ID, Bland buffers the deltas into speech segments and returns audio before the turn finishes, and an end-of-turn message releases the tail. Pipecat's Bland service and LiveKit Agents' Bland plugin both stream to that socket by default.

**Who is it for?** Bland Speech suits teams feeding a language model's output into a voice on a phone call, where a caller can interrupt at any moment.

<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>Server-side buffering with no client threshold, timer, or flush per fragment</li><li>A new context ID or a cancel message stops the turn that is playing</li><li>Billing follows delivered audio, so buffered text that never produced sound isn't charged</li><li>Mu-law at 8 kHz and PCM from 8 to 48 kHz</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. Inworld: Best for sentence detection switched on with one setting

**What does it do?** Inworld sells Realtime TTS-2 and TTS-2 Flash. Its persistent WebSocket sends audio chunks over the same connection, and setting autoMode to true with a sentence-boundary strategy has the server detect sentences as a language model's text arrives. Inworld says TTS-2 Flash reaches first audio in as low as 50ms. Output includes mu-law and A-law at 8 kHz.

**Who is it for?** Inworld suits teams that want a documented setting for how text is cut into sentences and a low per-character rate.

<div class="cmp-proscons" data-rival="inworld"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Sentence detection through one setting on the WebSocket</li><li>TTS-2 Flash at $15 per million characters on pay as you go</li><li>Server-sent events on its OpenAI-compatible endpoint</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>TTS-1.5 Max ranked 13th of 13 AI voices on Audio Realism Bench on 4 August 2026, and TTS-2 wasn't in those results</li><li>2,000 characters per REST request</li><li>The lower rates need a monthly plan from $25 to $1,500</li></ul></div></div>

**Pricing:** TTS-2 at $25 per million characters and TTS-2 Flash at $15 on pay as you go, falling to $12.50 and $7 on the $1,500 Growth plan, with Enterprise custom.

### 3. MiniMax: Best for text sent at any granularity, down to single characters

**What does it do?** MiniMax sells speech-2.8 text to speech in HD and Turbo versions. Its bidirectional WebSocket accepts text at any granularity, even character by character, and buffers it into sentences on the server before synthesis. A cancel message stops a turn, and sending can continue afterward. HTTP streaming returns server-sent events, and output formats include mu-law at 8 kHz.

**Who is it for?** MiniMax suits teams that want to pipe a language model's raw token stream into speech and be able to cancel mid-turn.

<div class="cmp-proscons" data-rival="minimax"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Text accepted character by character</li><li>Cancel and continue on one connection</li><li>Speech-02 HD ranked 7th of 13 AI voices on Audio Realism Bench on 4 August 2026, at 1093</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>speech-2.8-hd costs $100 per million characters</li><li>speech-2.8-turbo costs $60 per million</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 monthly audio packages from $5 for 100,000 points.

### 4. Smallest AI: Best for fragments joined by a shared context ID

**What does it do?** Smallest AI sells Lightning text to speech. Its WebSocket takes fragments tagged with the same context ID, buffers them, joins them at natural sentence boundaries, and speaks them as one continuous generation, with per-word timing events interleaved with the audio. Smallest AI says sub-100ms latency, and gives about 200ms time to first byte at 40 concurrent requests.

**Who is it for?** Smallest AI suits teams that want the server to join fragments and want word timings on the same socket.

<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>About $19.50 per million characters for Lightning v3.1 Pro</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><li>Lightning v3.1 retires on 30 November 2026, and clones made on it are affected</li></ul></div></div>

**Pricing:** About $0.195 per 10,000 characters for Lightning v3.1 Pro, which is $19.50 per million, with $10 of free credit to start.

### 5. Murf: Best for sending tokens straight in with no client-side tokenizer, at $10 per million characters

**What does it do?** Murf sells Falcon 2.1 text to speech with HTTP streaming and a WebSocket. Murf says to turn off a framework's sentence tokenizer or text aggregation and stream a language model's tokens straight to the WebSocket, because Falcon does its own sentence splitting. Murf publishes first audio under 100ms for Falcon 2.1, and word timestamps came to the WebSocket on 9 October 2026.

**Who is it for?** Murf suits teams that want the simplest client code and the lowest per-character rate in this group.

<div class="cmp-proscons" data-rival="murf"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Tokens go straight to the socket, with no client tokenizer</li><li>$0.01 per 1,000 characters for Falcon 2.1</li><li>Word timestamps on the WebSocket</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Murf's WebSocket returns WAV and PCM</li><li>Murf's 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 $10 of free API credit to start.

### 6. Speechify: Best for an explicit flush, with a documented cost in prosody

**What does it do?** Speechify sells Simba text to speech, and Simba 3.5 launched on 21 October 2026 with 48 languages under one model ID. GET /v1/audio/stream/ws opens a WebSocket for text that arrives in pieces, such as a reply a language model streams a few tokens at a time, and the session speaks one segment at a time. Speechify says a Simba 3.2 session reached first audible audio at a median of 47ms over the WebSocket from US East on warm connections, and 145ms from Zurich. Output includes ulaw_8000.

**Who is it for?** Speechify suits teams that want control over when the last sentence is spoken and accept a sentence-by-sentence delivery.

<div class="cmp-proscons" data-rival="speechify"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Flush and close messages that send the last sentence</li><li>Word timestamps through speech marks on the WebSocket</li><li>Simba 3.5 at a list price of $10 to $15 per million characters</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Intonation and pacing start afresh at each segment</li><li>A session ends after 60 seconds with nothing to speak and stops taking text after 5 minutes</li><li>Each segment is billed as one request</li></ul></div></div>

**Pricing:** Starter at $10 a month, Pro at $99, and Scale at $499, with Simba 3.5 at $15, $12, and $10 per million characters, Speechify's list price as of November 2026.

## How were these text to speech APIs compared for LLM streaming?

The six APIs were compared on who decides when streamed text is spoken, how a turn is closed, interruption, phone audio, and published price. Rival details come from each company's own pricing and documentation, 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, and [Best text to speech that handles interruptions](/blog/best-tts-interruptions-barge-in) covers what happens when a caller cuts in.

## Frequently asked questions

### What is the best text to speech API for streaming LLM output?

Bland Speech is the best text to speech API for streaming LLM output on phone calls. Its WebSocket takes a language model's tokens as they arrive, buffers them on the server, and stops a turn when a caller interrupts, and its voice ranked first of 13 AI voices in Design Arena's results of 4 August 2026. Inworld, MiniMax, and Murf suit a team that wants a different buffering model or a lower rate.

### Does Bland Speech take text from an LLM as it is written?

Yes. WSS /v2/tts/ws takes tokens exactly as the model produces them, and a single character is a valid delta, up to 4,000 characters a turn.

### Does a text to speech API need a sentence tokenizer?

Not on Bland, Inworld in auto mode, MiniMax, Smallest AI, or Murf, where the server buffers, and Murf tells developers to turn a framework's sentence tokenizer off. Neuphonic, Riva, Sarvam, Gradium, and Speechify leave more of the decision to the client.

### What happens when a caller interrupts the voice?

On Bland's socket, text under a new context ID preempts the turn that is playing, and a cancel message stops it without a replacement. Audio already sent can't be retracted, so the client discards what hasn't played.

### Does OpenAI's text to speech API accept streamed LLM text?

Not on its speech endpoint, which takes the whole text in one request. OpenAI's WebSocket and WebRTC connections are on its Realtime API, and tts-1, tts-1-hd, and gpt-4o-mini-tts are due to shut down on 6 January 2027.
