# Azure text to speech vs Google text to speech (2026): Which cloud voice suits a phone call?

By Samuel Oram-Jones

The voices from Microsoft and Google that listeners rated highest in blind tests are the two hardest to put a price on. Microsoft's MAI-Voice-2 ranked 2nd of 13 AI voices at 1224 in Design Arena's Audio Realism Bench results of 4 August 2026, and Microsoft says its MAI-Voice models are in public preview, "not recommended for production workloads", with no service-level agreement. Google's Gemini voices ranked 5th, 8th, and 10th in the same results, and Google bills them by the token with no free usage.

A team choosing between the two clouds for a voice agent on phone calls should also weigh Bland Speech, which bills $0.04 per 1,000 characters with no plan, and whose v3 model ranked 1st in those results at 1365. The [text to speech comparison hub](/compare/text-to-speech) and the [rival comparisons](/compare/text-to-speech/rival-vs-rival) cover the other vendors and pairs.

## Which is better, Azure text to speech or Google text to speech?

Azure is the pick for a team that wants a production voice at a published price and latency, as of November 2026: Azure's neural voices cost $15 per million characters, and Microsoft states less than 300 ms for its HD and standard voices. Google is the pick for a team that wants to direct delivery inside the text itself, because Gemini-TTS takes bracket tags such as [whispering] and [short pause], and Chirp 3: HD gives the first 1 million characters free each month.

The catch sits in each cloud's newest voices. Microsoft prints N/A where MAI-Voice pricing would go, and Google prices Gemini-TTS per million tokens, so a monthly bill for either is a calculation, not a rate.

| Feature | Bland Speech | Azure text to speech | Google text to speech |
|----|----|----|----|
| Audio Realism Bench, 4 August 2026 | Ranked 1st of 13, 1365 | MAI-Voice-2 ranked 2nd, 1224 | Gemini 2.5 Pro TTS ranked 5th, 1109; Gemini 3.1 Flash TTS 8th, 1037; Gemini 2.5 Flash TTS 10th, 985 |
| Price per million characters | $40, one standard rate | $15 neural, $22 neural HD in East US, N/A for MAI-Voice | $30 Chirp 3: HD, $16 Neural2, $4 WaveNet and Standard, Gemini by the token |
| Free usage | 50,000 characters, once | 0.5 million neural characters a month | First 1 million Chirp 3: HD characters a month, none for Gemini-TTS |
| First audio, as each company publishes it | 199ms | Less than 300 ms for HD and standard voices | "Low-latency streaming", in Google's words |
| Phone audio | Mu-law at 8 kHz | Mu-law and A-law at 8 kHz | MULAW and ALAW |
| Voice cloning | Instant clone from about 10 seconds | Personal and professional voices through a Limited Access review | Chirp 3 instant custom voice closed to new requests, Gemini 3.8 replication in Preview |
| HIPAA | Bland is HIPAA compliant with a signed BAA | HIPAA BAA through Microsoft's Product Terms | Text-to-Speech covered by Google Cloud's BAA |

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

- **Audio Realism Bench, 4 August 2026:** Design Arena's blind listening test, where native speakers hear two recordings of the same line and pick the one that sounds more like a real person. Azure's standard neural voices and Google's Chirp 3: HD weren't in those results, so only the models named in the cells have a score.
- **Price per million characters:** Each rate is as the company publishes it. Microsoft's Retail Prices API shows neural HD at $22 in East US while its public pricing page prints N/A. Google's Gemini voices bill $1.00 per million text tokens in and $20.00 per million audio tokens out on 2.5 Pro and 3.1 Flash, so a per-character cost depends on how much audio a text produces.
- **First audio:** The time from sending text to hearing the first sound. Each figure is the company's own claim, and Google's wording is a quotation.
- **Phone audio:** Mu-law and A-law are the two 8-bit G.711 telephony formats.
- **Voice cloning:** A clone copies one person's voice from recordings. Microsoft restricts custom voices to approved registrations, and Google points new requests to Gemini 3.8 voice replication.
- **HIPAA:** A BAA is the business associate agreement a US healthcare customer signs with a vendor that handles patient data.

</details>

## What does Azure text to speech do well, and where does it fall short?

Azure sells four kinds of voice: standard neural voices, HD voices (DragonHD and DragonHDOmni), Azure OpenAI voices, and the MAI-Voice models. Standard voices cover 100+ languages and locales, and Microsoft publishes less than 300 ms for both HD and standard voices.

The best-scoring voice is the weakest-supported one. MAI-Voice is in public preview, and Microsoft offers MAI-Voice-2.1 and MAI-Voice-2.1-Flash through Azure Speech where Design Arena scored MAI-Voice-2, so a buyer would need to confirm which version a deployment serves before treating 1224 as Azure's.

Text streams in over a WebSocket V2 endpoint, and Microsoft says text streaming is supported in C#, C++, and Python through the Speech SDK. A standard resource defaults to 30 transactions per second and can be raised to 1,000, and the free F0 tier allows 20 transactions per 60 seconds.

Custom voices are gated. Microsoft says custom neural voice is a Limited Access feature "available by registration only, and only for certain use cases", and that only customers working directly with Microsoft account teams are eligible. Personal voice needs a 5 to 90 second sample, a spoken consent statement, and approved access.

<div class="cmp-proscons" data-rival="azure"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Neural voices at $15 per million characters, falling to $7.50 at the 2,000 million tier</li><li>A published latency figure of less than 300 ms for HD and standard voices</li><li>Mu-law and A-law output at 8 kHz, plus 100+ languages and locales</li><li>Speech containers for teams that host the service on their own hardware</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>MAI-Voice is in public preview with no SLA, and its price shows as N/A</li><li>Custom and personal voices need Limited Access approval</li><li>Text streaming over WebSocket covers C#, C++, and Python only</li><li>Commitment tiers exclude HD, Azure OpenAI, custom, and personal voices</li></ul></div></div>

### Azure text to speech pricing

Azure charges $15 per million characters for neural voices, real-time or batch. Microsoft sells commitment tiers on neural voices: $960 for 80 million characters with overage at $12 per million, $3,900 for 400 million with overage at $9.75, and $15,000 for 2,000 million with overage at $7.50. Custom voice synthesis is $24 per million characters, $48 for neural HD, plus $52 per compute hour of training up to $936 and $4.04 per model per hour of hosting. Microsoft says prices are estimates and vary by region.

Ten million neural characters a month costs $150 at the pay as you go rate, against $400 for the same text on Bland Speech. That gap is real, and since Azure's standard neural voices weren't in the 4 August results, the sound side of the trade is left to a buyer's own listening.

Use Azure when:

- A team will provision a Speech resource inside an existing Azure subscription and manage regional quotas
- A team will apply for Limited Access approval to get a custom or personal voice
- A team will stream text through the Speech SDK in C#, C++, or Python

## What does Google text to speech do well, and where does it fall short?

Google sells Chirp 3: HD voices, the Gemini-TTS models (gemini-3.1-flash-tts-preview, gemini-2.5-flash-tts, gemini-2.5-flash-lite-preview-tts, and gemini-2.5-pro-tts), and older WaveNet, Neural2, Studio, Polyglot, and Standard voices. Google describes Chirp 3: HD as offering "low-latency streaming" and states 380+ voices across 75+ languages and variants. Text-to-Speech sits inside Google Cloud's HIPAA BAA and SOC 2 scope.

Gemini-TTS is where Google's expressiveness lives. Inline tags come in four groups: non-speech sounds such as [sigh] and [laughing], style modifiers such as [whispering], vocalized adjectives such as [curious], and pauses such as [short pause]. Google warns that vocalized-adjective tags can be spoken aloud, which it calls "likely an undesired side effect for most use cases". Chirp 3: HD has no emotion control and takes only a speaking rate from 0.25 to 2.0 and pause tags in its markup field, and SSML, the markup language for pauses and pronunciation, isn't supported on its streaming requests.

Chirp 3: HD allows 200 requests per minute and 100 concurrent streaming sessions per project, and Gemini 2.5 Flash TTS allows 150 queries per minute by default. Google's mu-law output carries a WAV header, which a raw phone stream has to account for. On cloning, Google says "You can no longer request access to Chirp 3: Instant Custom Voice" and points to Gemini 3.8 voice replication, which is in Preview.

<div class="cmp-proscons" data-rival="google"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Gemini-TTS inline tags for sounds, style, and pacing</li><li>First 1 million Chirp 3: HD characters free each month</li><li>MULAW and ALAW output for phone lines</li><li>Text-to-Speech covered by Google Cloud's HIPAA BAA</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Gemini-TTS bills by the token with no free usage, and 3.1 Flash is in Preview</li><li>Vocalized-adjective tags can be spoken aloud, in Google's words "likely an undesired side effect"</li><li>Chirp 3: HD has no emotion control and no SSML on streaming requests</li><li>Instant custom voice closed to new requests</li></ul></div></div>

### Google text to speech pricing

Chirp 3: HD is $30 per million characters, with the first 1 million characters each month free. Instant custom voice is $60 per million with no free usage. Legacy WaveNet and Standard voices are $4 per million, Neural2 is $16, and Studio is $160. Gemini 2.5 Pro TTS and Gemini 3.1 Flash TTS bill $1.00 per million text tokens in and $20.00 per million audio tokens out, and Google counts 25 audio tokens per second of audio, so output alone is $0.03 per minute of speech before input tokens.

Use Google when:

- A team will write delivery directions into the text with bracket tags and a style prompt
- A team will work inside Google Cloud projects and manage per-project quotas
- A team will bill Gemini voices by the token and estimate audio length from 25 tokens per second

## Which sounds more human, Azure text to speech or Google text to speech?

Microsoft's MAI-Voice-2 scored higher than any Google voice in Design Arena's Audio Realism Bench results of 4 August 2026, ranked 2nd at 1224 against 5th at 1109 for Google's best, Gemini 2.5 Pro TTS. Elo is a rating that rises each time listeners pick a voice over another, so 115 points is a wide gap. Design Arena's [Audio Realism Bench article](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/) describes the test and names Bland Speech v3 at 1365, MAI-Voice-2 at 1224, and Grok TTS at 1156.

Two limits apply to that result. Google's other two scored voices, Gemini 3.1 Flash TTS and Gemini 2.5 Flash TTS, landed 8th at 1037 and 10th at 985, and neither cloud's older workhorse voices were in the results at all, so a buyer reading the chart learns about the newest models and nothing about Azure's neural voices or Google's Chirp 3: HD. Bland Speech v3 ranked 1st at 1365, which is 141 points above MAI-Voice-2, and the human recordings in the same pool scored 1500.

## Why weigh Bland Speech against Azure and Google?

Bland Speech is worth weighing when the call is the product, because Bland Speech v3 ranked 1st of 13 AI voices in the 4 August 2026 results and bills one standard rate, $0.04 per 1,000 characters ($40 per million), with no plan, no tokens, and no preview label. A new account gets 50,000 characters free once.

The trade is cost per character. Azure's neural voices at $15 and Google's Chirp 3: HD at $30 both undercut $40, and a team reading a rate sheet would pick one of them. Bland publishes 199ms to first audio and Microsoft states less than 300 ms, though each company defines its own measure, so a buyer can't read one figure against the other.

Bland Speech streams over a WebSocket at WSS /v2/tts/ws, where text from a language model goes in as it arrives and a new context interrupts the turn that is playing, a behavior called barge-in. Output is PCM from 8 to 48 kHz or 8 kHz mu-law for phone lines. The [text to speech API](/text-to-speech-api) has the details, and [Bland Speech](/speech) has the studio.

Bland is SOC 2 Type II audited, HIPAA compliant with a signed BAA, and FedRAMP 20x Class C certified at Moderate, granted 17 November 2026, and Bland says speech synthesis runs inside that FedRAMP boundary. The [Azure review](/blog/azure-text-to-speech-review) and the [Google review](/blog/google-text-to-speech-review) cover each cloud on its own, and [Bland Speech vs Azure](/compare/bland-speech-vs-azure-tts) sets Bland against one of them.

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Ranked 1st of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>One standard rate of $0.04 per 1,000 characters with no plan or token math</li><li>WebSocket streaming with barge-in and mu-law output at 8 kHz</li><li>Instant voice clone from about 10 seconds of audio</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>

## Frequently asked questions

### Is Azure or Google text to speech cheaper?

Azure is cheaper on its standard neural voices, at $15 per million characters against $30 for Google's Chirp 3: HD. Google is cheaper on legacy WaveNet and Standard voices at $4 per million, and Google's Gemini voices bill by the token, so no per-character comparison holds for them. Bland Speech is $40 per million characters.

### Does Azure or Google have better voice cloning?

Both gate it. Microsoft limits custom neural voice and personal voice to approved registrations, and Google says new requests for Chirp 3 instant custom voice are closed in favor of Gemini 3.8 voice replication, which is in Preview. Bland Speech makes an instant clone from one clean sample of about 10 seconds, and every clone is private to the customer's organization.

### Which is better for real-time streaming on a phone call?

Both stream, with different catches. Azure streams text in over a WebSocket V2 endpoint in C#, C++, and Python and states less than 300 ms for HD and standard voices. Google offers bidirectional streaming with text in and audio out, and its Chirp 3: HD streaming requests don't accept SSML.

### Do Azure and Google output phone-ready audio?

Yes, both do. Azure returns mu-law and A-law at 8 kHz, and Google's AudioEncoding offers MULAW and ALAW, with a WAV header on MULAW output. Bland Speech returns mu-law at 8 kHz.

### Are Azure and Google text to speech HIPAA compliant?

Microsoft makes a HIPAA business associate agreement available through its Product Terms to customers that are covered entities or business associates, and Google includes Text-to-Speech in the products covered by its BAA. Bland is HIPAA compliant with a signed BAA.

### Is MAI-Voice ready for production?

Microsoft says MAI-Voice is in public preview, provided without a service-level agreement, and "not recommended for production workloads".
