Azure Speech

Published September 3, 2026 · This page indexes everything this site has verified about Azure and links to where each finding is set out in full. It is not a review — the review is. No listening test has been run, so nothing here says how Azure sounds.

On this page

What it costs

Microsoft’s own pricing page publishes no readable text-to-speech price: all 142 price cells render the literal string "$-", the only dollar-and-digits string in the 646KB document is the "$200" free-account credit boilerplate, and the page carries its own caveat that "Prices are estimates only and are not intended as actual price quotes." The site therefore sources rates from Microsoft’s first-party Azure Retail Prices API (Foundry Tools, East US), queried September 1, 2026: $15.00 per 1 million characters for the S1 Neural meter and $22.00 for Neural HD, the latter carrying an effective date of March 1, 2026 — with the $15 figure separately corroborated by a Microsoft Learn cost-estimation document stating "Multiply the result by the unit price of $15 per million characters to estimate the monthly cost." The free allowance is 0.5 million Neural characters a month on the recurring F0 tier, with unused models decommissioned after 7 days; an earlier check of the same Retail Prices API for the flagship ranking, on August 11, 2026, also recorded commitment tiers falling from $0.015 to $0.006 per 1,000 characters.

Every rate above is set against every other provider’s, normalized to one unit, in our pricing comparison.

Where Azure’s own material disagrees with itself

These are contradictions inside the vendor’s published documentation, not disagreements between us and the vendor. We report both readings rather than choosing one, and each links to the page where we set it out in full. We have published 12 in total; the 10 with the most bearing on a decision are below, and the rest sit in the pages linked underneath.

• Microsoft’s pricing page and its own price sources disagree about whether Azure Speech has a price at all. Every one of the 142 text-to-speech price cells renders as "$-", yet Microsoft’s own Retail Prices API returns $15.00 per million for the S1 Neural meter and $22.00 for Neural HD, and a Microsoft Learn cost-estimation document states the $15 rate in plain prose. As the site puts it, the number a buyer takes to a budget meeting cannot be sourced to the page their finance team will check. Where we published this.
• A much larger free allowance sits on the same pricing page as the current one. The current tier is stated as 0.5 million Neural characters free per month, while inside a block Microsoft labels "These features are being deprecated and only available for existing customers to use" a legacy free tier offers 5 million characters per month — ten times the current allowance, a scroll away, and unavailable to a new customer. Where we published this.
• Three official Microsoft pages give three different totals for preset voices and languages: "over 400 options in more than 140 languages and locales" on one, "more than 500" voices on another, and a third figure elsewhere. The site publishes all readings, and neither picks the most flattering number nor averages them. Where we published this.
• A single Microsoft page states three different totals for its high-definition voice count, in three places. Where we published this.
• Individual voice families are given conflicting language counts within a single page — one page gives two different numbers for the same family in its feature table and in its notes, and another does the same for personal voice. Where we published this.
• A default parameter value is stated one way in the prose and another way in the parameter table, on the same Microsoft page. Where we published this.
• Regional availability of HD voices differs between Microsoft’s regions page and its feature documentation. Where we published this.
• Azure’s advertised language figure matches neither reading of its own enumerated list. Its overview advertises "100+", while the site’s count of the enumerated text-to-speech voices table found 767 voice rows across 154 distinct locales, which reduce to 82 distinct base language subtags — so the advertised figure sits between the two and matches neither. Azure is at least explicit that it presents coverage in locales rather than languages, which the site credits. Where we published this.
• The cloning product’s naming contradicts its own documentation. "Custom neural voice" is now "custom voice": the documentation page is titled "What is custom voice?" and the phrase "neural voice" does not appear on it once, while its URL still contains custom-neural-voice. Separately, prebuilt voices are called "Standard voice" in the documentation and "Neural" on the pricing page. Where we published this.
• The product is two renames deep while its pre-rename identifiers remain live everywhere. SDK namespaces still read Microsoft.CognitiveServices.Speech, the container registry path still says azure-cognitive-services, the Azure resource type is unchanged at Microsoft.CognitiveServices/accounts, and the RBAC roles are still named "Cognitive Services Speech Contributor" and "Cognitive Services Speech User". Existing code keeps working; search results stay confusing. Where we published this.

Dated and time-sensitive

Facts with a clock on them — model sunsets, policy revisions, market status. These are the entries most likely to be out of date first, which is why they carry their dates.

• 2026-03-01. The Neural HD Text to Speech Characters meter, at $22.00 per 1 million characters, carries an effective date of March 1, 2026 in Microsoft’s Retail Prices API. Source on this site.
• 2027-04-01. Microsoft documents that "The Long Audio API will be retired on April 1st, 2027", so tutorials and integrations should be checked for which API they use; batch synthesis remains the documented path, accepting up to 10,000 text inputs per job with a job lifetime of up to 31 days. Read September 2, 2026. Source on this site.
• 2026-01-30. No rename date could be established. The documentation carrying the "Foundry Tools" name is stamped January 30, 2026, but that is page metadata rather than a rename announcement, and a full search of the Speech service release notes found no dated rebrand entry at all. The site declines to invent one and advises anyone citing a specific rename date to say where they got it. Source on this site.
• 2026-09-02. Published correction. The site had reported that Azure Speech publishes no speech-specific commercial-use grant; that was wrong. Microsoft’s Product Terms carry an express grant, verified live on this date: "TTS Service output use rights: For Customers of the paid tier TTS Service only, Customer may use the audio output of prebuilt neural voices generated using the TTS Service, including for commercial purposes." The grant covers the paid tier and prebuilt neural voices only, leaving free F0 output outside it. The site attributes its earlier error to the Product Terms document running to over 20MB and defeating its tooling, and corrected four pages. Source on this site.
• 2026-08-11. The newest models, MAI-Voice-2 and MAI-Voice-2-Flash, are public preview, carry no SLA, and Microsoft says they are not recommended for production. Recorded from the model documentation via the TTS overview on this check date. Source on this site.
• 2026-09-01. Two F0 free-tier limits quoted from the pricing page’s own footnotes: "Unused models will be automatically decommissioned after 7 days", and free speech-to-text hours are "shared between Standard and Custom" with batch unsupported. Source on this site.
• 2026-09-01. Product-naming lineage as Microsoft documents it: "Azure AI Studio → Azure AI Foundry → to Microsoft Foundry (current)" for the platform, and "Azure Cognitive Services → Azure AI Services → to Foundry Tools (current)" for the services portfolio, with the Azure resource type unchanged. Microsoft’s AI code of conduct states it "unifies and replaces the previous codes for Microsoft Generative AI Services, Azure Face in Foundry Tools, and Azure Speech in Foundry Tools text to speech" — which the site calls the clearest dated evidence it found that the older product name was in use as recently as 2025. Source on this site.
• 2026-09-02. Azure’s batch synthesis job time-to-live of up to 31 days is the only retention figure any of the ten vendors the site tracks publishes for async output; the batch unit itself, at up to 10,000 text inputs per job, is the largest documented here. Source on this site.

Every page on this site about Azure

• Full review. Azure Speech in Foundry Tools Review 2026: Pricing and Access Gates, Docs-Verified — the complete documentation-based assessment, with sources and dates against every claim.
• Alternatives. Azure Speech alternatives — organized by the reason you would leave, not as a ranked list.
• Head to head. Amazon Polly vs Azure Speech, Azure Speech vs Google Cloud.
• In our buying guides. Best AI Voice Generators in 2026: 10 Providers Compared, Best AI Dubbing Software 2026: Only Four Vendors Actually Sell It, Best Free AI Voice Generators 2026: What “Free” Actually Buys You, Most Realistic AI Voices: What Every Vendor Charges for the Claim, Best Multilingual AI Voice Generators: Every Count, Recounted, Real-Time AI Voice: Why the Latency Numbers Cannot Be Compared, Best AI Voice Cloning Software 2026: Price, Access and Consent Compared.
• In our comparisons of published data. AI Voice Generator Pricing Comparison: Every Rate Normalized, Best Text-to-Speech APIs: Limits, Transports and Auth Compared.
• In our explainers. Can You Use AI Voices Commercially? What Ten Vendors Actually Say, How AI Voice Cloning Works: What Ten Vendors Require of You, How AI Voice Generators Work, According to the Vendors Themselves.
• By what you are making. AI Voice for Audiobooks: Will Chapter 20 Match Chapter 1?, AI Voice for Business: What Survives a Procurement Review, AI Voice for eLearning: Courses Change, and the Audio Must Follow, AI Voice for Podcasts: Only Three Providers Do Two Speakers, AI Voice for YouTube: What You Must Disclose Before You Publish.

What we could not verify

Honesty about gaps beats a page that looks complete. Standing limits on everything above, as of September 3, 2026:

• How Azure sounds. Our listening benchmark is designed but not running. Nothing on this page or on any page it links to ranks Azure on audio quality, naturalness or speed.
• Anything Azure does not publish. Every page indexed here is built from the vendor’s own documentation. Where that documentation is silent, we record the silence rather than estimating around it.
• Whether a documented feature works as documented. We read documentation; we did not create an account or make an API call.
• Currency. Each linked page carries its own check date. This index inherits them and does not re-verify on load.

Change log

September 3, 2026 — First publication. Assembled from pages already published on this site; no new vendor claim was introduced by this page.

Published September 3, 2026 · This page shows no audio and reports no listening results, because our benchmark has not run. Corrections are recorded with their date on the corrections page. Independence note: this page contains no affiliate links, and no vendor paid for placement or influenced the order — see how we make money and our editorial policy.