Microsoft says of MAI-Voice, the newest voice family in Azure Speech, "This preview is provided without a service-level agreement, and is not recommended for production workloads." A phone line that has to answer every call has little use for a voice that ships with that sentence, and Azure's other premium paths come with gates of their own: custom voice is a Limited Access feature, and personal voice needs approval.
Moving from Azure to Bland Speech changes the shape of each request more than it changes the audio. An SSML document (the XML markup Azure Speech reads for voice, style, and pauses) becomes a JSON body, a voice name becomes a voice ID, and a custom voice is cloned again from audio. Bland keeps more text to speech migration guides beside its text to speech comparison hub.
Why would a team move from Azure text to speech to Bland Speech?
A team moves from Azure text to speech to Bland Speech when its newest Azure voice is preview-only, its custom voice needs Microsoft's approval, or its cost for HD voices can't be read off a public price page, and it accepts a higher per-character rate than Azure's Neural voices, as of November 2026.
On the voice, Microsoft says MAI-Voice is offered to customers as MAI-Voice-2.1 and MAI-Voice-2.1-Flash, reached through the standard Azure Speech API and SDK with SSML, in public preview. On cost, Microsoft's public price for MAI-Voice reads "N/A" per million characters, as does Neural HD's in the default region, while Microsoft's Retail Prices API prices Neural HD at $22 per million characters in East US, effective 1 March 2026, and carries no MAI-Voice meter at all.
Cloning is gated more tightly than the price. Microsoft says custom neural voice "is a Limited Access feature available by registration only, and only for certain use cases", and that "Only customers managed by Microsoft, meaning those who are working directly with Microsoft account teams, are eligible for access." Personal voice, which needs a 5 to 90 second sample, is restricted to eligible customers and approved use cases, and MAI-Voice cloning, which Microsoft recommends doing with 5 to 60 seconds of audio, goes through the same Limited Access Review. Bland's instant clone needs one clean sample of about 10 seconds, and the customer confirms every voice is theirs to use. A professional clone is fine-tuned on 30 minutes or more of audio and goes through Bland's consent and rights review.
Bland Speech costs $0.04 per 1,000 characters ($40 per million), pay as you go with no monthly plan, and a new account gets 50,000 free characters once. Azure is cheaper for plain neural speech. Azure Neural and Neural HD Flash voices are $15 per million characters, Azure's F0 free tier gives 0.5 million neural characters a month, and its commitment tiers fall to $7.50 per million at 2,000 million characters for $15,000. Azure also offers more than 100 languages and locales for standard voices, and Microsoft's HIPAA Business Associate Agreement is available through the Product Terms by default.
Published latency numbers can't settle the choice. Microsoft says Azure Speech HD voices and its standard voices run in less than 300 ms, and Bland publishes 199ms to first audio on a streamed request. Each company defines its own clock, so a team compares them on its own network.
What maps to what when moving from Azure Speech to Bland Speech?
An Azure SSML request maps to a JSON body for Bland's POST /v2/tts, an Azure text streaming session maps to Bland's WebSocket at /v2/tts/ws, and an Azure output format such as raw-8khz-8bit-mono-mulaw maps to an audio object with mulaw at 8,000 Hz.
What these rows mean
- Request shape: Bland's native endpoints take a plain
textfield, and styling comes from bracket tags and two numeric controls rather than markup. - Style: Azure's style degree runs from 0.01 to 2 and Bland's
expressivenessruns from 0.0 to 1.0, and they aren't the same dial, so values are re-tuned by ear instead of converted. - Price: each side's published rate, in US dollars. Azure's regional prices vary, and Microsoft calls its page prices estimates.
Azure's expressive features have close Bland counterparts, with different spelling. The express-as styles Microsoft documents, such as whispering and ecstatic, map to Bland tags such as [say in a whisper] and [say excitedly], and Bland publishes the full trained set and says invented variants are "ignored or garbled", so each style is matched to a tag from that list. Azure's HD voices read emotional cues from the text on their own, and Bland's tags sit before the words they shape. A pause is written <|0.8|>, with a length from 0.1 to 10.0 seconds, and each marker adds about a second of breath on top of that length. Bland's tags work on its current-generation voices only, the BTTS_V3 family.
Frameworks are the easy case. The Pipecat team maintains services for both Azure and Bland, and Bland's installs with uv add "pipecat-ai[bland]", with BlandTTSService for the WebSocket and BlandHttpTTSService for complete text. LiveKit Agents has Python plugins for Azure AI Speech and for Bland, and Bland's installs with uv add "livekit-agents[bland]~=1.8".
What are the steps to switch from Azure text to speech to Bland Speech?
A switch takes six steps: build a voice table, rewrite SSML as plain text, change the output format, port streaming, clone voices again, and run both providers in parallel before cutover.
- A team builds a voice table with each Azure voice name beside a Bland voice UUID.
GET /v1/voiceslists every voice a team can use, and Speech Studio renders a line with the same models the API uses, so a team can listen to candidates before it writes the table. - A team rewrites each SSML document as plain text with Bland's tags and pause markers, and sets
expressivenessandstabilityper request where a flow needs a different read. - A team changes the output format. Azure's raw-8khz-8bit-mono-mulaw becomes
"audio": { "encoding": "mulaw", "sample_rate": 8000 }, and Azure's riff-8khz-8bit-mono-mulaw becomes the same object with"container": "wav". A team that doesn't need 8 kHz asks for 48,000 Hzpcm_s16le, whichBTTS_V3renders natively. - A team replaces Azure text streaming with one WebSocket per conversation, using
initonce,speakdeltas under onecontext_idper turn, andend_of_turn. Azure limits one request to 10 minutes of audio, while Bland's HTTP endpoint takes up to 5,000 characters and a WebSocket turn up to 4,000, so long scripts are split by characters instead of minutes. - A team clones each custom voice again from its original recordings, with the speaker's permission on file.
- A team routes a small share of calls to Bland, compares recordings of the same lines, and moves the rest once the results hold.
curl -X POST "https://api.bland.ai/v2/tts" \
-H "Authorization: Bearer $BLAND_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"text": "Great news. <|0.4|> Your results came back clear.",
"voice": "29158307-9893-4149-8a75-bc9ce313d64e",
"controls": { "expressiveness": 0.8, "stability": 0.4 },
"audio": { "container": "wav" }
}' \
--output news.wav
That request replaces an SSML document with one pause marker and two controls. The response carries x-model, x-cost, and x-latency headers, so a team can log what each call cost and which model a voice resolved to.
Do Azure voices and Bland Speech sound different to callers?
Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, at 1365, with the human recording at 1500, and Microsoft's MAI-Voice-2 ranked 2nd at 1224, 141 points behind.
- #1BlandBland Speech v31365
- 2ndMicrosoftMAI-Voice-21224
- 3rdxAIGrok TTS1156
- 4thOpenAIGPT Realtime 21119
- 5thGoogleGemini 2.5 Pro TTS1109
Ranks are among 13 AI voices, and Microsoft's own voice is on the list. Real human recordings scored 1500, at the right edge of this scale. Bars start at 600, not zero. Source: Design Arena's published results.
Version numbers matter here. Azure Speech offers MAI-Voice-2.1 and MAI-Voice-2.1-Flash, and neither was in those results, and neither were Azure's standard neural voices or its HD voices. A team on those voices has no score from this benchmark to compare. The benchmark has native speakers pick, blind, which of two clips of the same text sounds more like a real person, and it has a Phone Agents category. The practical test is a team's twenty most frequent lines rendered in Speech Studio next to the same lines from Azure.
What does the move cost, and what stays on Azure?
A move from Azure Neural to Bland Speech costs 2.7 times as much per character, and Azure custom voices, personal voices, and any self-hosted containers stay behind.
At 10 million characters a month, Azure Neural at $15 per million is $150 and Bland Speech at $40 per million is $400, both before free allowances. At 80 million, Azure's commitment tier is $960 and Bland is $3,200. Bland charges one standard rate whether a script is directed or not, in the studio or the API, and each request carries a $0.001 minimum, so very short prompts bill at the minimum. Azure's custom voice carries charges of its own: $52 per compute hour for training, up to $936 per training, and $4.04 per model per hour for endpoint hosting, which is about $2,909 for a 30-day month before any characters.
Voices don't transfer. Microsoft requires "explicit written permission from voice talent prior to creating a voice model", and Bland has each customer confirm every voice is theirs to use, so the permission on file is the first thing a team checks before cloning a talent's voice a second time. Microsoft also reserves the right to run speaker recognition on the recorded acknowledgement statement.
Azure Speech also runs in containers, and Microsoft says containers are "constrained only by the CPUs of the hardware that they're hosted on", so a team that hosts speech on its own hardware for data-location reasons weighs that before moving. Enterprise customers get dedicated orchestration and dedicated GPUs run by Bland. Enterprise pricing is custom, set with Bland's sales team.
The Bland Speech vs Azure text to speech head-to-head, the Azure text to speech alternatives list, and the Azure text to speech review cover the wider choice, and the custom voice cloning comparison covers the clone path. The text to speech API page lists each endpoint, and Bland Speech has the studio.
Frequently asked questions
Can an Azure custom voice or personal voice move to Bland Speech?
No, an Azure custom voice or personal voice can't move to Bland Speech as a voice, because Bland voices are made on Bland from audio. A team clones again from its original recordings, with an instant clone needing one clean sample of about 10 seconds and a professional clone needing 30 minutes or more plus Bland's consent and rights review.
How do Azure speaking styles translate to Bland Speech?
Azure styles map to Bland bracket tags such as [say warmly], [say excitedly], and [say in a whisper], and a team re-tunes expressiveness and stability by ear. Bland says tags must match its published set exactly, because invented variants are ignored or garbled.
Is Bland Speech cheaper than Azure text to speech?
No. Bland Speech is $0.04 per 1,000 characters ($40 per million), and Azure Neural is $15 per million characters, with commitment tiers down to $7.50 per million. Bland has no monthly plan, and a new account gets 50,000 free characters once.
Can a Pipecat or LiveKit agent switch from Azure to Bland Speech?
Yes. Pipecat ships a Bland service maintained by its own team, and LiveKit Agents has a Bland plugin for Python, so the swap is an install and a class change instead of a rewrite of the streaming code.
Does Azure's HIPAA agreement carry over?
Microsoft makes its HIPAA Business Associate Agreement available through the Product Terms by default. Bland is HIPAA compliant with a signed BAA, which it offers on Enterprise, so a team moving regulated audio settles that with Bland's sales team before any traffic moves.
