Azure Text to Speech: Voices, Pricing and Limits

Azure text to speech is the speech synthesis feature of Azure AI Speech, which Microsoft now lists as Azure Speech in Foundry Tools. As of October 2026, standard neural voices list at $15 per 1 million characters and HD voices at $22, in East US and in USD. The free F0 tier includes 0.5 million neural characters per month.

This guide covers the voice tiers, how billing works, the SSML differences between tiers, the quotas on each tier and the regions where each tier runs. Prices come from Azure's Retail Prices API and Microsoft's quota documentation, both checked on 2026-10-17.

What Azure text to speech includes

Microsoft's text to speech overview describes standard voices as "high-quality neural voices available out of the box in 100+ languages and locales." Each standard voice model is available at 24 kHz and at high-fidelity 48 kHz.

The overview lists three main options:

Standard voices, called Neural on the pricing page.
Custom voices, which Microsoft describes as professional voice fine-tuning or personal voice.
Batch synthesis, for audio longer than 10 minutes, such as audiobooks or lectures.

Visemes, the mouth-shape events used for lip sync, are supported only for en-US neural voices.

Azure voice types compared

The table below combines the published prices with Microsoft's voice tier descriptions. The list prices are for East US in USD.

| Voice tier | List price per 1M characters | Status | Notes | |---|---|---|---| | Standard (Neural) | $15 | Generally available | 100+ languages and locales, full SSML, cloud, embedded, hybrid and containers | | Neural HD (Dragon HD, Dragon HD Omni) | $22 | Generally available, with some Preview voices | Cloud only, SSML subset, bracket tags for style | | Custom neural, real-time | $24 | Limited access | Training billed per compute hour, with a 96-hour cap | | Custom neural HD synthesis | $48 | Limited access | Listed in the Retail Prices API | | Personal voice | $24, plus storage | Limited access | Consent required | | MAI-Voice-2.1 and Flash | Not readable on the pricing page | Public preview | 23 languages, gated voice cloning |

Microsoft's HD voices page describes DragonHD as about 30 fine-tuned voices, and Dragon HD Omni as a larger set that includes AI-generated voices. Microsoft's own figures for the Omni set differ between sections of the page, so this guide does not give a count for Omni.

Azure text to speech pricing and billing rules

Azure bills per character of each successfully processed request. Several billing rules change what counts:

SSML markup counts as billable, except the speak and voice elements.
Spaces and punctuation count.
Each Chinese character counts as two characters, including kanji used in Japanese, hanja used in Korean and hanzi used in other languages.
You pay for a request even if it produces no audio.

Microsoft's quotas and limits page, updated 2026-09-09, states "the unit price of $15 per million characters" for standard voices. Other list prices from the Retail Prices API:

Neural HD: $22 per 1M characters.
Custom neural real-time: $24 per 1M characters.
Custom neural HD synthesis: $48 per 1M characters.
Custom long-audio: $100 per 1M characters.
Personal voice synthesis: $24 per 1M characters.

Prices differ by region and currency, and the figures above are list prices in USD for East US only. Azure also offers commitment tiers for neural voices at 80 million, 400 million and 2,000 million characters per month. The pricing page names those tiers, but this guide does not give their prices.

The free F0 tier includes 0.5 million neural characters per month. Some third-party articles quote a 5-million-character free tier. Microsoft's pricing page gives 0.5 million for F0, and the larger figure applies to older plans for existing customers only.

SSML and style control

Support for SSML and style control differs by tier.

Bracket tags: HD voices accept style and paralinguistic tags in the text, for example an excited tag before a sentence. Paralinguistic tags include laughter, coughing, throat clearing, breathing, sighing and yawning.
Express-as on Dragon HD Omni: the mstts:express-as element is supported on Dragon HD Omni only, not on DragonHD, according to Microsoft's SSML support table.
Omni tuning: Dragon HD Omni exposes temperature (default 0.7, range 0.3 to 1.0), top_p, top_k (default 22) and cfg_scale (default 1.4).

HD voices support only a subset of SSML. The prosody, emphasis, audio, bookmark and silence elements are not supported on either DragonHD or Omni. The break and phoneme elements are supported on DragonHD only.

If your SSML depends on prosody or emphasis, test it on the tier you plan to use before you commit to HD voices.

Quotas and limits

Microsoft's quotas page gives these real-time limits:

| Limit | F0 (free) | S0 (standard) | |---|---|---| | Real-time TTS transactions | 20 per 60 seconds, not adjustable | 30 transactions per second by default, adjustable up to 1,000 | | Maximum audio per request | 10 minutes | 10 minutes | | Distinct voice and audio tags per SSML | 50 | 50 | | WebSocket SSML message size per turn | 64 KB | 64 KB | | Batch synthesis | Not available | Available |

Batch synthesis has its own limits:

JSON payload up to 2 MB.
Up to 10,000 text inputs per job.
REST limit of 100 requests per 10 seconds.
Jobs are kept for up to 31 days.

For HTTP 429 responses on standard voices, Microsoft says most come from limited backend capacity for a voice in the selected region rather than from your quota. Use retries and use the voice in its native region.

Batch synthesis for long audio

Batch synthesis suits audio longer than 10 minutes. You send the request, poll for status and download the result when the job completes. The quotas above apply, and the job is not available on F0. If your project is an audiobook or a long lecture, use batch rather than sending many real-time requests.

Regions and deployment options

Microsoft's regions page lists 33 regions or geographies for neural TTS. HD voices run in a subset of those regions, including East US, East US 2, West Europe, Sweden Central, Southeast Asia, Central India, Canada Central, France Central and West US 2. Check the page for the current list before you pick a region.

Deployment differs by tier. Standard voices can run in the cloud, in embedded or hybrid setups and in containers. HD voices are cloud only and real-time only.

MAI-Voice-2.1 is accessible globally and served from 14 regions, according to its documentation.

Custom, personal and instant voices

Custom and personal voices are gated. Microsoft describes custom voice as "limited access for responsible use," and you apply through the Limited Access Review. Microsoft says professional voice fine-tuning usually takes 20 to 40 compute hours for a single style and about 90 hours for multi-style, with a 96-compute-hour billing cap. Personal voice is billed for profile storage and synthesis.

MAI-Voice-2.1 offers instant voice cloning from a consented clip of 5 to 60 seconds. That feature is gated, and the MAI models are in public preview. Microsoft says they are "not recommended for production workloads." The Flash variant targets low-latency agents, and the non-Flash model targets long-form audio. Microsoft lists 23 languages for MAI-Voice-2.1.

Getting started in five steps

Create an Azure AI Speech resource and note its key and region.
Install the SDK for your language, or call the REST API directly.
Pick a voice name for the tier you need, using the names listed on Microsoft's HD voices page for HD voices.
Send a short test request with plain text, then add bracket tags or SSML if you need style.
Check your character count, quotas and region before you move to production.

Azure compared with other clouds

Azure is a cloud service you provision. You create a resource, manage keys and regions, and tune behavior with SSML. It suits teams already running on Azure who want their speech billing on an existing Azure account. Our best text to speech API comparison puts Azure next to Google, OpenAI, Amazon Polly, ElevenLabs and Deepgram on the same published-price basis. Amazon Polly alternatives covers a similar comparison for teams leaving AWS.

Where OndaLabs fits

OndaLabs is an AI voice studio. Where Azure asks you to provision a resource and write SSML, OndaLabs gives you a set of tools for finished audio:

Text to speech for narration and voiceovers.
Two-speaker dialogue for conversations.
Custom voices from a recording you have permission to use, or designed from a text description.

OndaLabs offers 30 prebuilt voices usable in every supported language, with dedicated pages for 37 languages. Delivery is directed with plain words and supported expressive tags. Credits are priced per character, with Onda Lite at 1 credit per character and Onda Flash at 1.5. The Free plan costs $0 and includes 2,000 credits per month, about 2 minutes. The Starter plan costs $9 per month for 170,000 credits, and it includes commercial use.

These are different billing models, so do not treat the Azure and OndaLabs prices as directly comparable. Review the pricing page for current plan details.

Frequently asked questions

Does Azure text to speech have a free tier?

Yes. The F0 tier includes 0.5 million neural characters per month, with a limit of 20 transactions per 60 seconds. F0 is not available for batch synthesis. Microsoft's pricing page gives this figure, and it supersedes older third-party claims of a 5-million-character free tier.

How much does Azure text to speech cost per million characters?

Standard neural voices list at $15 per 1 million characters, and Neural HD voices at $22, both in East US and in USD as of October 2026. Prices vary by region and currency, and commitment tiers are priced separately. Check the pricing page for your region before you budget.

What are Azure HD voices and MAI voices?

HD voices include Dragon HD and Dragon HD Omni. They are cloud only, real-time only and support a subset of SSML, with style controlled through bracket tags and, on Omni, the express-as element. MAI-Voice-2.1 and Flash are in public preview, support 23 languages and are not recommended for production workloads.

Which Azure regions support HD voices?

Microsoft's regions page lists a subset of regions for HD voices, including East US, East US 2, West Europe, Sweden Central and Southeast Asia. Standard neural voices are offered in a wider set of regions. Check the regions page before you choose a deployment region, because the list can change.

Can I create a custom or personal voice in Azure?

Yes, but access is gated. Custom voice and personal voice require limited-access approval through Microsoft's Limited Access Review. Personal voice also requires consent. Plan for the approval step before you design a product around a custom voice.

How long can one real-time Azure text to speech request be?

A real-time request can produce up to 10 minutes of audio. For longer content, Microsoft recommends batch synthesis, which handles up to 10,000 text inputs per job and keeps jobs for up to 31 days.