# OpenAI TTS vs Azure text to speech (2026): Which one outputs phone-line audio?

By Samuel Oram-Jones

A phone line carries 8 kHz audio in a compressed format called mu-law or A-law, and one of these two speech endpoints returns it directly while the other does not. As of November 2026, Azure's Speech service outputs raw 8 kHz mu-law and A-law, and OpenAI's speech endpoint returns mp3, opus, aac, flac, wav, or pcm at 24 kHz. That difference matters less than a second one: OpenAI says `tts-1`, `tts-1-hd`, and `gpt-4o-mini-tts` shut down on 6 January 2027.

Bland Speech outputs 8 kHz mu-law and bills a standard $0.04 per 1,000 characters ($40 per million), pay as you go with no monthly plan, and Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026.

## Which is better, OpenAI TTS or Azure text to speech?

Azure text to speech is the better pick for a team that needs phone-ready audio, a long language list, and a Microsoft contract, and OpenAI's speech endpoint is the better pick only for a short-lived prototype, since its models end on 6 January 2027. A phone agent that has to sound like a person is a job for Bland Speech more than for either.

| | Bland Speech | OpenAI TTS | Azure text to speech |
|---|---|---|---|
| Status | Bland Speech v3 is current | `tts-1`, `tts-1-hd`, and `gpt-4o-mini-tts` shut down 6 January 2027 | Neural, HD, and custom voices; MAI-Voice in public preview with no SLA |
| Price per million characters | $40, one standard rate | `tts-1` $15, `tts-1-hd` $30; `gpt-4o-mini-tts` by token | Neural and Neural HD Flash $15; custom voice $24 |
| Free allowance | 50,000 characters once | Pay per use | 500,000 neural characters a month on the F0 tier |
| Phone audio | mu-law at 8 kHz, PCM 8 to 48 kHz | mp3, opus, aac, flac, wav, 24 kHz pcm | 8 kHz mu-law and A-law, PCM up to 48 kHz |
| Streaming | HTTP chunks and WebSocket | HTTP chunked transfer; WebSocket only on the Realtime API | REST playback as it transfers; WebSocket v2 text streaming in the Speech SDK |
| Published latency | 199ms to first audio | OpenAI says `tts-1` has lower latency than `tts-1-hd` | Under 300 ms for HD and standard voices; over 500 ms for Azure OpenAI HD voices |
| Cloning | Instant from about 10 seconds, one clean sample | Eligible customers, samples of 60 seconds or less | Personal voice from 5 to 90 seconds, by limited-access intake form |
| Realism on 4 August 2026 | 1st of 13 at 1365 | GPT-4o Mini TTS 9th at 1011; GPT Realtime 2 4th at 1119 | Microsoft MAI-Voice-2 2nd at 1224 |

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

- **Phone audio:** Mu-law and A-law are compressed 8-bit formats used on phone lines. A voice that outputs them directly needs no conversion before a call carries it.
- **Published latency:** Each company's own statement, measured by its own method. Microsoft's figures are for its HD voices, standard voices, and Azure OpenAI HD voices, and none of the three is a race Bland ran.
- **Realism on 4 August 2026:** Position and Elo score in Design Arena's Audio Realism Bench results of that date. Elo is a rating built from head-to-head votes, here from listeners picking the clip that sounds more like a person.

</details>

## Where do OpenAI TTS and Azure text to speech fall short for a voice agent?

OpenAI TTS falls short on its clock and its formats, and Azure text to speech falls short on access, since its cloning and its newest voices sit behind a review or a preview.

### OpenAI TTS

OpenAI announced on 1 October 2026 that its three speech models were deprecated, and the replacement, `gpt-realtime-2.1-mini`, lives on the Realtime API, which OpenAI describes as "a change of API workflow". A voice agent built on `/v1/audio/speech` has a rewrite scheduled for the new year.

Delivery is steered with a plain-language `instructions` field on `gpt-4o-mini-tts`, so the steering goes away with that model. OpenAI's usage policies require a clear disclosure to end users that the voice is AI-generated, and custom voices from recordings are limited to eligible customers, with a consent recording in an approved phrase and a sample of 60 seconds or less.

<div class="cmp-proscons" data-rival="openai"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>tts-1 at $15 per million characters</li><li>Plain-language instructions on gpt-4o-mini-tts</li><li>SOC 2 Type 2 for its API, and a Business Associate Agreement offered</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>All three speech endpoint models shut down on 6 January 2027</li><li>No mu-law on the speech endpoint</li><li>The replacement is a different API</li><li>Custom voices limited to eligible customers</li></ul></div></div>

OpenAI TTS is a poor fit for:

- A product that has to keep running past 6 January 2027 without a rewrite
- A telephony stack that wants 8 kHz mu-law straight from the speech endpoint

### Azure text to speech

Azure's catch is that its most interesting voices are gated. Custom voice is a Limited Access feature available by registration only, and Microsoft says only customers managed by Microsoft, meaning those working directly with Microsoft account teams, are eligible. Personal voice needs an intake form and takes a 5 to 90 second sample plus a verbal consent statement. MAI-Voice-2.1 and MAI-Voice-2.1-Flash are in public preview, and Microsoft says that preview carries no service-level agreement and "is not recommended for production workloads".

Microsoft prices Neural HD and MAI-Voice as N/A in the default region, and Microsoft's Retail Prices API puts Neural HD at $22 per million characters in East US. Text streaming over WebSocket works in the C#, C++, and Python versions of the Speech SDK.

<div class="cmp-proscons" data-rival="azure"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>100+ languages and locales on standard voices</li><li>Raw 8 kHz mu-law and A-law output</li><li>Styles and style degree from 0.01 to 2 through SSML on a subset of neural voices</li><li>A HIPAA Business Associate Agreement through the Microsoft Product Terms by default</li><li>Speech containers for self-hosting, as Microsoft offers them</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Custom voice limited to customers managed by Microsoft</li><li>MAI-Voice in public preview with no SLA</li><li>Neural HD and MAI-Voice priced N/A in the default region</li><li>HD voices do not support every SSML tag</li></ul></div></div>

Azure text to speech is a poor fit for:

- A team that wants to clone a voice today without a Microsoft account team or an approved intake form
- A production line that needs MAI-Voice under an SLA
- A buyer who needs one published price for the best Azure voices

[Bland Speech vs Azure text to speech](/compare/bland-speech-vs-azure-tts) sets Azure against Bland, and [OpenAI TTS alternatives](/blog/openai-tts-alternatives) covers the ways off OpenAI's endpoint.

## What do OpenAI TTS, Azure text to speech, and Bland Speech cost?

At published per-character rates, `tts-1` and Azure Neural both cost $15 per million, `tts-1-hd` costs $30, and Bland Speech costs $40.

Ten million characters a month is $150 on `tts-1` or Azure Neural, $300 on `tts-1-hd`, and $400 on Bland Speech. Azure also sells commitment tiers for neural voices: $960 for 80 million characters with overage at $12 per million, $3,900 for 400 million at $9.75, and $15,000 for 2,000 million at $7.50. HD, Azure OpenAI, custom, and personal voices are not in those tiers. At 80 million characters a month, the first tier costs $960 against $1,200 on `tts-1` and $3,200 on Bland Speech.

Azure's custom voice costs $24 per million characters, plus $52 per compute hour for training up to $936 per training run and $4.04 per model per hour for endpoint hosting. Bland Speech's professional clones are fine-tuned on 30 minutes or more of audio after a consent and rights review. Phone minutes, taxes, and the language model that writes the text come on top of every figure here.

Bland Speech is the most expensive of the three per character, and a team that prices only on that line will choose one of the others.

## Which sounds more human, OpenAI TTS, Azure text to speech, or Bland Speech?

In Design Arena's Audio Realism Bench results of 4 August 2026, Bland Speech v3 ranked first of 13 AI voices at 1365, against a human recording at 1500, Microsoft's MAI-Voice-2 ranked 2nd at 1224, OpenAI's GPT Realtime 2 ranked 4th at 1119, and GPT-4o Mini TTS ranked 9th at 1011.

Azure's standard neural and HD voices weren't in those results, and neither were MAI-Voice-2.1 and MAI-Voice-2.1-Flash, the names Azure Speech uses for its preview voices. OpenAI's `gpt-realtime-2.1-mini` was not scored either. Listeners in Design Arena's method hear two recordings of one transcript, blind and in random order, and pick the one that sounds more like a real person, which is described in [Design Arena's article](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/).

The 1224 belongs to Microsoft's MAI-Voice-2, and a buyer on Azure Speech works with the MAI-Voice-2.1 line in public preview, so the score describes a sibling of the preview voices and not a production Azure voice.

## Why does Bland Speech suit a phone agent better than OpenAI TTS or Azure text to speech?

Bland Speech suits a phone agent better than OpenAI TTS or Azure text to speech because it has Azure's phone-ready output without a gated program and OpenAI's request shape without a shutdown date.

An instant clone needs one clean sample of about 10 seconds, and every clone is private to the customer's organization with the customer confirming the voice is theirs to use. A client built for OpenAI's `/audio/speech` changes its base URL, API key, and model to route through Bland, with `speed` limited to 1.0 and `instructions` ignored. Native streaming runs on a WebSocket at `/v2/tts/ws`, and Bland's [text to speech API](/text-to-speech-api) documents both endpoints. Bland publishes 199ms to first audio.

Bland is SOC 2 Type II audited, HIPAA compliant with a signed BAA, and FedRAMP 20x Class C certified at Moderate, and Bland says speech synthesis runs inside that boundary. Bland Speech has a first-party Pipecat service and a LiveKit Agents plugin for Python. The step-by-step move from OpenAI is in [how to migrate from OpenAI TTS](/blog/migrate-from-openai-tts).

<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 in Design Arena's results of 4 August 2026</li><li>8 kHz mu-law output and an OpenAI-compatible endpoint</li><li>One standard rate with no monthly plan and no hosting fee</li><li>Cloning from one clean sample of about 10 seconds</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>

More matchups sit in the [text to speech hub](/compare/text-to-speech) and the [rival vs rival family](/compare/text-to-speech/rival-vs-rival), and Bland Speech itself is at [bland.ai/speech](/speech).

## Frequently asked questions

### Is OpenAI TTS better than Azure text to speech?

For a build that has to last, Azure text to speech is better, because OpenAI's `tts-1`, `tts-1-hd`, and `gpt-4o-mini-tts` shut down on 6 January 2027. Azure also outputs 8 kHz mu-law and A-law directly, which OpenAI's speech endpoint does not.

### Is OpenAI's text to speech API being shut down?

Yes, its three speech endpoint models are. OpenAI deprecated them on 1 October 2026 and set 6 January 2027 for the shutdown, with `gpt-realtime-2.1-mini` on the Realtime API as the replacement.

### Which is cheaper, OpenAI TTS or Azure text to speech?

They match at $15 per million characters for `tts-1` and Azure Neural. OpenAI's `tts-1-hd` costs $30. Azure's neural commitment tiers work out to $12, $9.75, and $7.50 per million at 80, 400, and 2,000 million characters, and Bland Speech costs $40.

### Can OpenAI TTS or Azure clone a voice?

Both do, with gates. OpenAI offers custom voices to eligible customers from recordings of 60 seconds or less. Azure's personal voice takes 5 to 90 seconds through a limited-access intake form, and custom voice is limited to customers managed by Microsoft. Bland Speech clones from one clean sample of about 10 seconds.

### Does Azure Speech support HIPAA?

Microsoft makes a HIPAA Business Associate Agreement available through its Product Terms by default to covered entities and business associates. OpenAI also offers a BAA. Bland is HIPAA compliant with a signed BAA.

### Does Azure Speech store the text a call sends?

For real-time synthesis with prebuilt voices, Microsoft says neither input text nor output audio is stored in its logs. Bland states that zero-retention processing is available for sensitive workflows.

### Can Bland Speech replace OpenAI TTS with a code change?

For most `/audio/speech` calls, yes: the base URL, API key, and model change. Bland's endpoint accepts `speed` of 1.0 only, ignores `instructions`, and returns a 400 error for server-sent events.
