# Voice agent text to speech in 2026: Deepgram, Cartesia, ElevenLabs, and Bland Speech

By Samuel Oram-Jones

Four text to speech companies publish four first-audio numbers for voice agents, 80ms from Deepgram, under 90ms from Cartesia, about 100ms from ElevenLabs, and 199ms from Bland, and the four aren't counting the same thing. Sorting by that column ranks each company's own claims, and a voice agent breaks in production in other places: streaming a language model's words in, stopping when a caller interrupts, and sounding like a person after both.

Bland Speech is the slowest of the four on the published number and the first AI voice in Design Arena's 4 August 2026 listening test, and a buyer has to price both facts. Bland's [text to speech comparison hub](/compare/text-to-speech) covers the other providers.

## Which text to speech is best for a voice agent: Bland Speech, Deepgram, Cartesia, or ElevenLabs?

For a voice agent that takes phone calls, Bland Speech is the pick of the four as of November 2026, because Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026 and was built for calls. Deepgram is the pick for the smallest published first-audio figure and for speech to text and text to speech from one vendor, Cartesia for a first-audio figure under 90ms, and ElevenLabs for language range and for builds on Twilio ConversationRelay.

| Feature | Bland Speech | Deepgram | Cartesia | ElevenLabs |
|---|---|---|---|---|
| Audio Realism Bench, 4 August 2026 | 1st of 13 AI voices, 1365 | Not in those results | Sonic 3.5: 6th, 1096 | Eleven v3: 12th, 938 |
| First audio, as published | 199ms | As low as 80ms on Flux TTS | Under 90ms | Median about 100ms on v4 Turbo, before application and network latency |
| Streaming text in | WebSocket, where a new context interrupts the turn that is playing | WebSocket on /v2/speak | WebSocket with contexts | WebSocket, with v3 and v4 routed through Text to Dialogue |
| Price per million characters | $40, one standard rate | $45 on Flux TTS, $30 on Aura-2 | $37.38 to $50 inside a plan | $40 on v4 Turbo, $80 on v4 |
| Phone audio | Mu-law at 8 kHz, PCM at 8 to 48 kHz | Mu-law and A-law at 8 or 16 kHz | Mu-law, A-law, and PCM | ulaw_8000, alaw_8000, and pcm_8000 |
| Pipecat and LiveKit | Both, with a Python LiveKit plugin | Both | Both | Both |

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

- **Audio Realism Bench:** Design Arena's blind listening test, where listeners pick which of two clips of the same line sounds more like a real person, and the votes become an Elo rating, a score that rises each time a voice is picked.
- **First audio, as published:** The time from sending text to hearing the first sound. Each number is the company's own claim, and none is a race Bland ran.
- **Streaming text in:** A WebSocket is a connection that stays open so a language model's words can arrive in pieces and the voice starts before the sentence ends.
- **Pipecat and LiveKit:** Open-source frameworks for building voice agents, each with a list of text to speech providers.
- **Phone audio:** Mu-law is 8-bit compressed audio at 8 kHz, used on North American and Japanese phone lines and by Twilio. A-law is the other common phone encoding.

</details>

## How do Deepgram, Cartesia, ElevenLabs, and Bland Speech handle first audio and streaming?

On the published numbers Deepgram, Cartesia, and ElevenLabs start faster than Bland Speech, and all four stream a language model's words in over a WebSocket. Speed is a real advantage for the three, and it doesn't show up on a listening test.

### What do the four first-audio numbers count?

Deepgram says Flux TTS responds in as low as 80ms at full 24kHz fidelity, even under production load. Cartesia says Sonic delivers first audio in under 90ms and calls that model latency. ElevenLabs puts the median for Eleven v4 Turbo at about 100ms before application and network latency, and its own pages give two figures for Flash v2.5, about 100ms and about 75ms. Bland publishes 199ms on a streamed request, counted to the first audio chunk rather than the finished file.

A voice agent's reply waits on speech recognition and a language model before any voice is generated, so text to speech is one part of the delay a caller hears. The gap between 199ms and 80ms is 119 milliseconds.

### How does each stream text in and handle an interruption?

All four take text over a WebSocket. Cartesia says only its WebSocket accepts text in pieces, through contexts. ElevenLabs' standard WebSocket doesn't support Eleven v3 or v4, which run through Text to Dialogue. Deepgram's Flux TTS WebSocket is turn-based, and tone carries forward across turns.

Bland Speech's WebSocket at /v2/tts/ws is for text that arrives as a language model writes it, and a new context_id immediately preempts the active turn, which makes interruption and barge-in a normal part of the protocol. Pipecat's Bland service uses a cancel message for interruptions, and LiveKit's Bland plugin cancels the current turn when a caller cuts in.

## Where do Deepgram, Cartesia, and ElevenLabs fall short next to Bland Speech?

Deepgram falls short on its two-generation lineup, Cartesia on what happens when credits end mid-call, and ElevenLabs on getting its newest models into a live stream.

### Where Deepgram falls short

Deepgram sells two generations at two prices. Flux TTS is the current model, which Deepgram recommends for all new builds, at $0.045 per 1,000 characters on pay as you go and on Growth. Aura-2, a previous-generation model by Deepgram's own description, costs $0.030 on pay as you go and $0.027 on Growth. Flux TTS ships with built-in text normalization in English plus 10 additional languages, and Deepgram says SSML is stripped from input rather than supported. A delivery dial from calm to animated runs from -2 to 2 and is marked beta.

<div class="cmp-proscons" data-rival="deepgram"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>First audio as low as 80ms at 24 kHz under production load, as Deepgram publishes it</li><li>Speech to text and text to speech from one vendor, with Twilio, Pipecat, and LiveKit guides</li><li>Self-hosting, private cloud, and on-premise deployment under Deepgram's terms</li><li>A $200 free credit with no card and no expiry</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Flux TTS costs $45 per million characters, above Bland Speech's $40</li><li>Aura-2, the cheaper voice, is previous generation</li><li>SSML is stripped, and the delivery dial is beta</li></ul></div></div>

Use Deepgram when:

- A team wants speech recognition and speech synthesis from one vendor and one invoice
- A security team will host text to speech itself, in a private cloud or on-premise
- A team will run Flux TTS at $45 per million characters or Aura-2 at $30 and decide which generation it needs

### Where Cartesia falls short

Cartesia bills Sonic 3.6 in credits at 1 per character from monthly plans, and with overages switched off, new requests are blocked once credits run out. For a voice agent that means new requests fail partway through a call when the month's allowance ends, unless overages are on, in which case they are billed at renewal at $65, $45, or $38 per million credits. Each break tag costs a credit.

<div class="cmp-proscons" data-rival="cartesia"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>First audio in under 90ms, as Cartesia publishes it</li><li>44 languages from one model, and pronunciation dictionaries</li><li>Word-level timestamps on its WebSocket</li><li>Mu-law and A-law output 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, and Sonic 3.6 was released after those results</li><li>New requests are blocked at zero credits when overages are off</li><li>Emotion controls are beta and English only</li></ul></div></div>

Use Cartesia when:

- A team will watch a monthly credit balance and turn overages on for production
- A team wants 44 languages and word timing from one WebSocket

### Where ElevenLabs falls short

ElevenLabs' real-time model is Eleven v4 Turbo, and its route into a live call is the catch: the standard text to speech WebSocket doesn't support Eleven v3 or v4, so a voice agent on v4 Turbo uses the Text to Dialogue WebSocket, a different endpoint. Price comes in two steps, $0.08 per 1,000 characters on v4 and $0.04 on v4 Turbo. Free accounts are limited to non-commercial use, and the commercial license starts on Starter.

<div class="cmp-proscons" data-rival="elevenlabs"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>90+ languages across the Eleven v4 family</li><li>One of the three text to speech providers Twilio ConversationRelay offers, with Google and Amazon</li><li>Audio tags such as [whispering] and [laughing] on Eleven v4 and v3</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, and Eleven v4 was released after</li><li>The standard text to speech WebSocket doesn't support Eleven v3 or v4</li><li>Two prices for the newest models</li><li>Free accounts are non-commercial</li></ul></div></div>

Use ElevenLabs when:

- A team builds a voice agent on Twilio ConversationRelay, where ElevenLabs is one of three text to speech options
- A team ships in dozens of languages and works from one vendor's 90+ language family

## What do Deepgram, Cartesia, ElevenLabs, and Bland Speech cost?

One million characters of agent speech costs $40 on Bland Speech, $45 on Deepgram's Flux TTS, $40 or $80 on ElevenLabs depending on the model, and $49 on Cartesia's Startup plan, which covers 1.25 million credits.

### Bland Speech pricing

Bland Speech costs $0.04 per 1,000 characters, which is $40 per million, pay as you go and the same in the studio and the API, with no monthly plan. A new account gets 50,000 characters free once, and credits don't expire. Enterprise pricing is custom, set with Bland's sales team.

### What does a month of agent speech cost on each?

Worked from published prices, with no phone or language model charges counted:

- **1 million characters:** Bland Speech $40. Deepgram Flux TTS $45, or $30 on Aura-2. ElevenLabs $40 on v4 Turbo or $80 on v4. Cartesia Startup $49.
- **10 million characters:** Bland Speech $400. Deepgram Flux TTS $450, or $300 on Aura-2. ElevenLabs $400 on v4 Turbo. Cartesia Scale $375, which is $299 for 8 million credits plus 2 million at $38 per million.

Deepgram offers a Flux TTS promotion, $1 in credits for every $1 spent up to $500 through 31 December 2026, and Cartesia's Scale plan beats Bland Speech's flat rate at 10 million characters. Bland Speech's rate doesn't change with volume or plan.

## Which sounds more human, Bland Speech, ElevenLabs, Cartesia, or Deepgram?

Bland Speech v3 ranked first of 13 AI voices in Design Arena's [Audio Realism Bench results](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/) of 4 August 2026, at 1365, where the human recording scored 1500. Cartesia's Sonic 3.5 ranked 6th at 1096 and ElevenLabs' Eleven v3 ranked 12th at 938. Deepgram wasn't in those results. Cartesia released Sonic 3.6 on 27 August 2026 and ElevenLabs released Eleven v4 on 28 September 2026, after the test, so neither is scored in it.

The test used one American-English female voice and one male voice per model, chosen as its most naturally conversational voice, so a team working in other languages should listen for itself.

### Why is Bland Speech the pick for a voice agent on phone calls?

Bland Speech is the pick when callers should hear a person, because Bland Speech v3 was built for phone calls and Bland runs its own model and speech instead of wrapping a third party's. Performance tags in square brackets, such as [laughs], [sighs], or [say warmly], shape how the words after them are delivered, and a first-party Pipecat service and a LiveKit Agents plugin put Bland Speech into the frameworks voice agents are built on. The [text to speech API](/text-to-speech-api) page carries the endpoints, and [Bland Speech](/speech) has the studio.

Bland, the company behind the model, is SOC 2 Type II audited, HIPAA compliant with a signed BAA, and FedRAMP 20x Class C certified at Moderate, granted 17 November 2026, and its trust page says speech synthesis runs inside the FedRAMP boundary.

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>One standard rate of $0.04 per 1,000 characters, with no plan tiers</li><li>A WebSocket where a new context interrupts the turn that is playing</li><li>Mu-law and PCM output, from 8 to 48 kHz</li><li>A first-party Pipecat service and a LiveKit plugin</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>

The two-way versions are [Bland Speech vs Deepgram](/compare/bland-speech-vs-deepgram), [Bland Speech vs ElevenLabs](/compare/bland-speech-vs-elevenlabs), and [ElevenLabs vs Cartesia](/blog/elevenlabs-vs-cartesia), and the [rival comparisons](/compare/text-to-speech/rival-vs-rival) page has every pairing.

## Frequently asked questions

### Which text to speech has the lowest latency for a voice agent?

On the numbers each company publishes, Deepgram's Flux TTS is lowest at as low as 80ms, then Cartesia under 90ms, ElevenLabs' Eleven v4 Turbo at a median of about 100ms, and Bland Speech at 199ms. The figures count different things, and ElevenLabs excludes application and network latency.

### Can Bland Speech handle a caller interrupting the agent?

Yes. Bland Speech's WebSocket at /v2/tts/ws treats a new context_id as an immediate interruption of the turn that is playing, and Pipecat's Bland service and LiveKit's Bland plugin both cancel the current turn when a caller cuts in.

### Do they work with Pipecat and LiveKit?

Yes. Pipecat's list of text to speech services includes Bland, Cartesia, Deepgram, and ElevenLabs, each maintained by the Pipecat team. LiveKit Agents has plugins for all four.

### What does voice agent text to speech cost per million characters?

On published prices, one million characters costs $40 on Bland Speech, $45 on Deepgram's Flux TTS or $30 on Aura-2, $40 on ElevenLabs' Eleven v4 Turbo or $80 on v4, and $49 on Cartesia's Startup plan, which covers 1.25 million credits. Bland Speech bills pay as you go with no monthly plan, and a new account gets 50,000 characters free once.

### Which voice agent text to speech sounds most human?

In Design Arena's Audio Realism Bench results of 4 August 2026, Bland Speech v3 ranked first of 13 AI voices at 1365, Cartesia's Sonic 3.5 ranked 6th at 1096, and ElevenLabs' Eleven v3 ranked 12th at 938. Deepgram wasn't in those results, and Sonic 3.6 and Eleven v4 were released after them.
