# 6 most human-sounding AI voices for phone calls in 2026

By Samuel Oram-Jones

On a phone line a voice gets one short turn to sound like a person, and Design Arena built a category around exactly that. Its Audio Realism Bench sorts prompts into Phone Agents, Explainers, and Conversations, and says most prompts run "up to roughly 40 words (about 15 seconds), though phone-agent prompts are single turns and tend to be shorter".

Among text to speech APIs that document 8 kHz telephony output, six took the highest places in Design Arena's results of 4 August 2026, and Bland Speech v3 took first. All rival facts are as of November 2026.

## What are the most human-sounding AI voices for phone calls?

The most human-sounding AI voice for phone calls in Design Arena's results of 4 August 2026 was Bland Speech v3, ahead of Microsoft's MAI-Voice-2, xAI's Grok TTS, Google's Gemini 2.5 Pro TTS, Cartesia's Sonic 3.5, and MiniMax's Speech-02 HD.

- **Bland Speech:** Best overall for a phone line, 1st at 1365
- **Azure:** Best for Microsoft's MAI-Voice-2 on Azure, 2nd at 1224
- **Grok TTS:** Best for a pay as you go rate of $15 per million characters, 3rd at 1156
- **Google:** Best for a Gemini model on Google Cloud, 5th at 1109
- **Cartesia:** Best for the lowest first-audio claim on this list, 6th at 1096
- **MiniMax:** Best for 40 languages on one API, 7th at 1093

| API | Voice scored on 4 August 2026 | Place and score | Phone audio | Published price |
|---|---|---|---|---|
| Bland Speech | Bland Speech v3 | 1st of 13, 1365 | Mu-law at 8 kHz, PCM from 8 to 48 kHz | $0.04 per 1,000 characters |
| Azure | MAI-Voice-2 | 2nd, 1224 | Mu-law and A-law at 8 kHz | $15 per million characters for Neural voices |
| Grok TTS | Grok TTS | 3rd, 1156 | Mu-law and A-law at 8 kHz | $15 per million characters |
| Google | Gemini 2.5 Pro TTS | 5th, 1109 | MULAW and ALAW in Gemini-TTS streaming | $1.00 in and $20.00 out per million tokens |
| Cartesia | Sonic 3.5 | 6th, 1096 | Mu-law and A-law at 8 kHz | 1 credit per character, plans from $5 a month |
| MiniMax | Speech-02 HD | 7th, 1093 | Mu-law at 8 kHz | $100 per million characters (speech-02-hd) |

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

- **Voice scored on 4 August 2026:** the model Design Arena's blind test scored. Where a company has released newer models since, the newer ones weren't scored.
- **Place and score:** the place among 13 AI voices and the score from Design Arena's chart. The human recordings scored 1500.
- **Phone audio:** the 8 kHz telephony formats each company documents. US phone networks carry mu-law.
- **Published price:** each company's published rate for its main model or the scored one. Google bills Gemini-TTS by token, with 25 audio tokens for each second of audio.

</details>

## What should a buyer check in an AI voice for phone calls?

A buyer should check how the score was earned, what phone audio the API returns, and how the voice behaves when a caller talks over it.

Start with the score. Design Arena's listeners hear two clips of the same transcript, blind and in random order, and pick the one that sounds more like a real person, with real human recordings in the same pool. For each model the bench uses one female and one male voice, "chosen as its most naturally conversational American-English voice", so a score describes two voices and not a whole catalog. A buyer should listen to the voice that will take the calls.

Then phone audio. Bland Speech, Azure, Grok TTS, Cartesia, and MiniMax return 8 kHz mu-law, and Google's Gemini-TTS streams MULAW and ALAW. OpenAI's GPT Realtime 2 ranked 4th at 1119 in the same results, but it runs on OpenAI's Realtime API rather than its text to speech endpoint, and that endpoint's tts-1, tts-1-hd, and gpt-4o-mini-tts models shut down on 6 January 2027, so OpenAI sits outside these six.

Last, interruption, because callers talk over agents. Bland Speech's WebSocket treats a new `context_id` as an immediate preemption of the turn that is playing, which makes interruption a normal part of the protocol, and MiniMax's bidirectional WebSocket takes a `task_cancel` message and lets a team continue afterward. Speed is each company's own claim: Bland publishes 199ms to first audio, Cartesia says under 90ms, and Microsoft says less than 300ms for its standard and HD voices.

## How do the 6 most human-sounding AI voices for phone calls compare?

These six voices took the top places among text to speech APIs that document 8 kHz output, and Bland Speech v3 took first place by 141 points over Microsoft's MAI-Voice-2.

### 1. Bland Speech: Best overall for a phone line, 1st at 1365

**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. It streams over HTTP at `POST /v2/tts` or over a WebSocket at `/v2/tts/ws` that takes text as a language model writes it, and returns mu-law at 8 kHz or PCM from 8 to 48 kHz. Bland publishes 199ms to first audio. Performance tags such as [laughs] and [say warmly] shape delivery, and a first-party Pipecat service and a LiveKit Agents plugin for Python connect it to voice agent frameworks.

**Who is it for?** Bland Speech suits teams putting a voice on phone calls and voice agents, where callers should hear a person.

<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 in Audio Realism Bench's results of 4 August 2026, at 1365</li><li>Mu-law at 8 kHz and a WebSocket with barge-in built into the protocol</li><li>199ms to first audio, as Bland publishes it</li><li>One standard rate, $0.04 per 1,000 characters, 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 50,000 characters free for a new account. The voices can be heard in the [Bland Speech](/speech) studio before a call goes through the [text to speech API](/text-to-speech-api).

### 2. Azure: Best for Microsoft's MAI-Voice-2 on Azure, 2nd at 1224

**What does it do?** Microsoft's MAI-Voice-2 ranked 2nd in the results. Azure offers the newer MAI-Voice-2.1 and MAI-Voice-2.1-Flash through the same Speech API and SDK as its other voices, in 23 languages. Azure's neural and HD voices return mu-law and A-law at 8 kHz, and its Speech SDK streams text in over a WebSocket.

**Who is it for?** Azure suits teams already on Microsoft's cloud that want a high-scoring voice under their existing agreement.

<div class="cmp-proscons" data-rival="azure"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>MAI-Voice-2 ranked 2nd of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>Neural voices at $15 per million characters, with 0.5 million characters free a month</li><li>Microsoft says HD and standard voices take less than 300ms</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>MAI-Voice is in public preview, "without a service-level agreement", and Microsoft says it "is not recommended for production workloads"</li><li>MAI-Voice-2.1, the version Azure sells, wasn't in the results of 4 August 2026</li></ul></div></div>

**Pricing:** $15 per million characters for Neural and Neural HD Flash voices, with 0.5 million characters free a month.

### 3. Grok TTS: Best for a pay as you go rate of $15 per million characters, 3rd at 1156

**What does it do?** Grok TTS is the text to speech API in the Grok API from xAI, now named SpaceXSI. It streams over HTTP at `POST /v1/tts` and over a bidirectional WebSocket, returns mu-law and A-law at 8 kHz in 20 languages, and xAI's own example request asks for mu-law at 8 kHz.

**Who is it for?** Grok TTS suits teams that want a high-scoring voice on a flat per-character rate, with no plan to manage.

<div class="cmp-proscons" data-rival="grok-tts"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Grok TTS ranked 3rd of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>$15 per million characters, pay as you go</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>For HIPAA, xAI points a buyer to a BAA questionnaire rather than stating compliance</li></ul></div></div>

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

### 4. Google: Best for a Gemini model on Google Cloud, 5th at 1109

**What does it do?** Google's Gemini 2.5 Pro TTS ranked 5th, Gemini 3.1 Flash TTS 8th at 1037, and Gemini 2.5 Flash TTS 10th at 985. Gemini-TTS runs through the Cloud Text-to-Speech API, which supports bidirectional streaming, and Gemini-TTS streams PCM, ALAW, MULAW, or Opus. Google's Cloud Text-to-Speech documents 380+ voices across 75+ languages and variants.

**Who is it for?** Google suits teams already on Google Cloud that want a Gemini voice and one bill.

<div class="cmp-proscons" data-rival="google"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Gemini 2.5 Pro TTS ranked 5th of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>MULAW and ALAW streaming</li><li>Text to speech is covered by Google Cloud's HIPAA BAA</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Gemini-TTS has no free usage and bills per audio token</li><li>Gemini 3.1 Flash TTS is in preview</li></ul></div></div>

**Pricing:** Gemini 2.5 Pro TTS costs $1.00 per million text tokens in and $20.00 per million audio tokens out, and Gemini 2.5 Flash TTS costs $0.50 and $10.00.

### 5. Cartesia: Best for the lowest first-audio claim on this list, 6th at 1096

**What does it do?** Cartesia's Sonic 3.5 ranked 6th in the results, and Sonic 3.6, released on 27 August 2026, speaks 44 languages and came out after them. Cartesia says first audio arrives in under 90ms. Sonic streams over HTTP, server-sent events, and a WebSocket, and returns mu-law and A-law at 8 kHz, a format Cartesia names for North American telephony and Twilio.

**Who is it for?** Cartesia suits teams that put first-audio speed ahead of everything else and will manage a monthly credit allowance.

<div class="cmp-proscons" data-rival="cartesia"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Cartesia says first audio arrives in under 90ms</li><li>Mu-law and A-law at 8 kHz</li><li>Telephony on a Cartesia number at $0.014 a minute</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, at 1096</li><li>Credits come from a monthly plan, with overages opt-in</li></ul></div></div>

**Pricing:** 1 credit per character from monthly plans: Pro at $5 for 100,000 credits, Startup at $49 for 1.25 million, and Scale at $299 for 8 million, with overages at $65, $45, or $38 per million credits.

### 6. MiniMax: Best for 40 languages on one API, 7th at 1093

**What does it do?** MiniMax's Speech-02 HD ranked 7th in the results, and its current Speech 2.8 models, released on 23 January 2026, weren't in them. The 2.6 and 2.8 models cover 40 languages, and MiniMax returns mu-law at 8 kHz and takes text over a bidirectional WebSocket as a language model writes it.

**Who is it for?** MiniMax suits teams that want broad language coverage and a streaming API on a pay as you go rate.

<div class="cmp-proscons" data-rival="minimax"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Speech-02 HD ranked 7th of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>40 languages on the 2.6 and 2.8 models</li><li>A bidirectional WebSocket with a cancel message for interruptions</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>speech-02-hd costs $100 per million characters</li><li>Its current Speech 2.8 models weren't in the results of 4 August 2026</li></ul></div></div>

**Pricing:** $100 per million characters for `speech-02-hd` and `speech-2.8-hd`, and $60 for the turbo models, pay as you go, or monthly audio packages from $5 to $999.

## Which of these AI voices sounded most human in a blind test?

Bland Speech v3 sounded most human in Design Arena's Audio Realism Bench results of 4 August 2026, at 1365, and only the human recordings scored higher, at 1500. Microsoft's MAI-Voice-2 followed in 2nd at 1224, Grok TTS 3rd at 1156, OpenAI's GPT Realtime 2 4th at 1119, Google's Gemini 2.5 Pro TTS 5th at 1109, Cartesia's Sonic 3.5 6th at 1096, and MiniMax's Speech-02 HD 7th at 1093.

The gap from first to second place was 141 points. Newer models such as Cartesia's Sonic 3.6, MiniMax's Speech 2.8, and ElevenLabs' Eleven v4 weren't in those results, so a buyer comparing today's catalogs should listen to current voices as well. Design Arena's [published results](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/) explain how the votes were collected.

## How were these AI voices compared?

Samuel Oram-Jones ordered six APIs by their places in Design Arena's published results of 4 August 2026, then compared phone audio, streaming, interruption handling, published speed claims, and price per character. Only text to speech APIs that document 8 kHz telephony output and have a scored voice were listed.

Bland's [text to speech comparison](/compare/text-to-speech) sorts these vendors by job, and the [best lists](/compare/text-to-speech/best-lists) sort them by buying question. A team picking a voice for a phone line can read [best text to speech for phone calls](/blog/best-tts-for-phone-calls), and one weighing customer-facing voices can read [best AI voices for customer conversations](/blog/best-ai-voices-customer-conversations).

## Frequently asked questions

### What is the most human-sounding AI voice for phone calls?

Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, at 1365, and only real human recordings scored higher, at 1500.

### How does Design Arena test whether a voice sounds human?

Listeners hear two clips of the same transcript, blind and in random order, and choose the one that sounds more like a real person. Design Arena says its listeners are vetted native speakers, and it enters a human recording of each prompt into the same pool.

### What audio format does a phone call need?

US phone networks carry 8 kHz mu-law. Bland Speech returns mu-law at 8 kHz, and a request that names the mulaw encoding and a sample rate of 8000 gets telephony audio back.

### Can callers interrupt a Bland Speech voice?

Yes. Bland Speech's WebSocket lets a new `context_id` preempt the turn that is playing, which makes interruption and barge-in a normal part of the protocol.

### How fast is Bland Speech compared with other voices?

Each company publishes its own figure. Bland publishes 199ms to first audio on a streamed request, Cartesia says under 90ms, and Microsoft says less than 300ms for its standard and HD voices.

### Was OpenAI's voice in the results?

Yes. OpenAI's GPT Realtime 2 ranked 4th at 1119 and GPT-4o Mini TTS ranked 9th at 1011. OpenAI's speech endpoint returns mp3, opus, aac, flac, wav, or pcm, and its tts-1, tts-1-hd, and gpt-4o-mini-tts models shut down on 6 January 2027.
