# SSML vs performance tags (2026): Two ways to direct a text to speech voice, and what Bland Speech uses

By Samuel Oram-Jones

A developer who sends Deepgram's Flux TTS a line of SSML gets spoken audio and a warning: Deepgram strips recognized SSML tags, returns an `INPUT_MARKUP_STRIPPED` warning, and carries on synthesizing. The markup was recognized, and the voice never read it.

SSML, short for Speech Synthesis Markup Language, is a standard for marking up text with angle-bracket tags such as `<break>`. Performance tags are the other approach, bracketed words such as [laughs] or [say warmly] typed into a script and read by the model as stage directions. A script written for one style can do nothing, or something unwanted, in an API built for the other.

## What is the difference between SSML and performance tags?

SSML is a W3C standard of angle-bracket tags for pauses, pace, volume, emphasis, and pronunciation, used by Amazon Polly, Azure, and Google, while performance tags such as [laughs] are bracketed words that Bland Speech, ElevenLabs, Google's Gemini-TTS, and Grok TTS read as delivery directions, as of November 2026.

Amazon Polly says it supports "a subset of the SSML markup tags" from SSML Version 1.1, a W3C Recommendation, and names pauses, pace, volume, emphasis, phonetic pronunciation, breathing sounds, whispering, and a Newscaster style among its controls. SSML decides how a sentence is mechanically spoken. A performance tag decides how it is meant: Bland says tags such as [laughs] or [say warmly] "sit right before the words they shape, and the model performs them."

Two other approaches sit beside those. Plain-language direction passes a sentence of instruction: OpenAI's `gpt-4o-mini-tts` takes an `instructions` prompt such as "Speak in a cheerful and positive tone", Google calls a style prompt the "primary driver of the overall emotional tone and delivery" for Gemini-TTS, and Inworld lets TTS-2 take a delivery described in plain words. Numeric dials set a range instead: Deepgram's Flux TTS has a delivery dial from -2 calm to 2 animated, marked beta, Cartesia sets speed from 0.6 to 1.5, and Bland has two controls from 0.0 to 1.0, expressiveness and stability.

## Which text to speech APIs take SSML, and which take performance tags?

Amazon Polly, Azure, and Cartesia take SSML, ElevenLabs v4, Gemini-TTS, Grok TTS, and Bland Speech take bracketed tags, and Deepgram strips SSML and offers dials, with several companies spanning more than one method.

| API | How a developer directs the voice | Limit or extra published |
|---|---|---|
| Bland Speech | Square-bracket tags such as [say warmly] and [laughs], timed pauses from 0.1 to 10.0 seconds, and expressiveness and stability from 0.0 to 1.0 | Tags work on current-generation voices only |
| Amazon Polly | SSML, a subset of W3C SSML 1.1 | `<audio>`, `<lexicon>`, `<lookup>`, and `<voice>` tags are not supported, and Newscaster style is not available on the Generative engine |
| Azure | SSML, with `mstts:express-as` styles and a style degree from 0.01 to 2 | Some voices don't support all SSML tags |
| Google | Gemini-TTS inline tags and a style prompt, and a subset of SSML in Cloud TTS | SSML tags are not supported for Chirp 3: HD streaming requests |
| ElevenLabs | Audio tags such as [whispering], [shouting], and [laughing] on Eleven v4 and v3 | Eleven v4 and v3 do not support SSML break tags |
| Cartesia | Speed, volume, and emotion through `generation_config` or SSML | Emotion controls are beta and English only |
| Deepgram | A delivery dial, a speed control from 0.5 to 1.5, and inline IPA pronunciation | Recognized SSML is stripped with a warning |
| Inworld | Plain-word steering and bracket tags on TTS-2, 50 non-verbal sounds such as [laugh], and SSML break tags for pauses | Steering works on `inworld-tts-2` only |
| OpenAI | Plain-language `instructions` on `gpt-4o-mini-tts` | `gpt-4o-mini-tts` shuts down on 6 January 2027 |
| Grok TTS | Inline tags such as [laugh] and [pause], and wrapping tags such as `<whisper>` and `<loud>` | Pronunciation replacements, including IPA |

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

- **How a developer directs the voice:** the controls each company documents for its text to speech API, in that company's own terms.
- **Limit or extra published:** a limit or add-on the same company states. Bland's note is that tags work on current-generation voices, and that older voice providers strip bracketed text before it reaches the model.
- **Inworld and ElevenLabs:** Inworld names steering as available on TTS-2 only, and ElevenLabs says phoneme tags work only on `eleven_flash_v2`.

</details>

## What happens when markup reaches an engine that does not read it?

Markup is stripped, ignored, or spoken, depending on the API: Deepgram strips SSML with a warning, Bland says older voice providers strip bracketed text, and Google warns that some Gemini-TTS tags are read aloud.

Each company handles the mismatch differently, and the exact limits matter more than the style. ElevenLabs says Eleven v4 and v3 do not take SSML break tags, so pauses come from punctuation and audio tags, while older models take `<break time="x.xs" />` for up to 3 seconds. ElevenLabs also says SSML phoneme tags work only on `eleven_flash_v2`, and Eleven v4 takes IPA between forward slashes instead. Google says SSML tags are not supported for Chirp 3: HD streaming requests, and that pause tags work only in the `markup` field of that model, not in `text`. Azure says some voices don't support all SSML tags, and its HD voices infer emotion from text on their own.

Spoken tags are the oddest failure. Gemini-TTS has tags for vocalized adjectives such as [scared], [curious], and [bored], and Google says they are spoken aloud, "likely an undesired side effect for most use cases". A bracket that is a stage direction on one voice is a word on another.

The consequence is portability. A script tuned with Polly's SSML meets a stripped-markup warning on Deepgram, and an ElevenLabs script that relies on `<break>` tags loses its pauses on Eleven v4, which takes none. Moving a voice agent between APIs can mean re-directing lines.

## How does Bland Speech direct a voice?

Bland Speech takes performance tags in square brackets, timed pauses written `<|N|>`, and two numeric controls, expressiveness and stability, on its current-generation voices.

A tag sits right before the words it shapes. Bland publishes a fixed set and says "Use these strings verbatim. They are the set the model was trained on." Emotion tags include `[say warmly]`, `[say angrily]`, `[say excitedly]`, `[say quietly]`, `[say in a whisper]`, `[say nervously]`, and `[say playfully]`. Intensity tags include `[very loud]`, `[very fast]`, and `[voice lowered]`. Vocalizations include `[laughs]`, `[chuckles]`, `[sighs]`, `[clears throat]`, `[gasps]`, and `[yawns]`. In Bland's studio, typing `[` opens an autocomplete list of supported tags.

A pause is `<|N|>`, where N is 0.1 to 10.0 seconds. Amazon Polly's `<break>` element tops out at 10 seconds each, and ElevenLabs' break tag at 3 seconds on older models, so the longest pause a script can add is similar on Bland and Polly. Expressiveness sets how much intonation varies, and stability sets how much the same input varies between renders, both from 0.0 to 1.0. Bland says those two controls are calibrated for `BTTS_V3`, its current generation.

Bland's studio also has Norm, a single-speaker director that, in Bland's words, writes "the words, the stumbles and the timing, then performs them." It lives in the dashboard studio. A team moving from OpenAI's `gpt-4o-mini-tts` writes the direction from its `instructions` prompt as tags on Bland Speech. Bland's [text to speech API](/text-to-speech-api) page covers the endpoints.

## Which should a team use, SSML or performance tags?

A team with SSML scripts already running on Amazon Polly or Azure keeps SSML, and a team directing a voice agent's warmth, laughs, and hesitations uses performance tags, which Bland Speech, ElevenLabs v4, and Gemini-TTS take.

Bland Speech v3 ranked first of 13 AI voices in Design Arena's [Audio Realism Bench results of 4 August 2026](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/), at 1365, and the human recording scored 1500. Bland Speech fits a phone agent that needs a voice to sound like a person and a script written in plain bracketed words from one fixed list. A team that wants SSML markup or free-form instructions has Amazon Polly, Azure, or Inworld to look at first.

<div class="cmp-usewhen"><div class="cmp-usewhen-own"><p><strong>Use Bland Speech when</strong></p><ul><li>A team will write scripts with bracketed performance tags and learn the fixed list Bland publishes</li><li>A developer will tune expressiveness and stability per voice on a current-generation voice</li><li>A buyer will drive a phone voice agent and add timed pauses of 0.1 to 10.0 seconds</li></ul></div><div><p><strong>Use an SSML-based API when</strong></p><ul><li>A team has SSML scripts in production on Amazon Polly or Azure and will keep them</li><li>An engineer will maintain pronunciation and emphasis markup across a large catalog of lines</li><li>A project will port scripts between SSML engines and test each voice for unsupported tags</li></ul></div></div>

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Plain bracketed tags with an autocomplete list in the studio</li><li>Timed pauses from 0.1 to 10.0 seconds and two numeric controls on current-generation voices</li><li>A model that ranked first of 13 AI voices in Design Arena's results of 4 August 2026</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>

The [performance tags list](/blog/best-tts-performance-tags) compares which APIs take tags, the [expressive text to speech list](/blog/most-expressive-tts-apis) covers delivery, and the [explainer on what makes a voice expressive](/blog/what-makes-ai-voice-expressive) goes further into how. [Bland Speech](/speech) has a studio for trying tags, the [text to speech hub](/compare/text-to-speech) holds every comparison, and more explainers sit under [text to speech guides](/compare/text-to-speech/guides).

## Frequently asked questions

### What performance tags does Bland Speech support?

Bland publishes a fixed list: emotion tags such as `[say warmly]` and `[say angrily]`, intensity tags such as `[very loud]` and `[voice lowered]`, and vocalizations such as `[laughs]`, `[sighs]`, and `[clears throat]`. Bland says they are the strings the model was trained on and should be used verbatim.

### Do tags work on every voice?

No. Bland says performance tags are supported on current-generation voices, and older voice providers strip bracketed text before it reaches the model. Azure says some voices don't support all SSML tags, and ElevenLabs says its phoneme tags work only on `eleven_flash_v2`.

### How long a pause can a script add?

Bland's `<|N|>` pause runs from 0.1 to 10.0 seconds, Amazon Polly's `<break>` element takes at most 10 seconds each, and ElevenLabs' break tag takes up to 3 seconds on older models. ElevenLabs says Eleven v4 and v3 do not take SSML break tags.

### Can a text to speech model laugh?

Bland Speech performs `[laughs]`, `[chuckles]`, and `[sighs]`, ElevenLabs v4 takes audio tags such as [laughing], Gemini-TTS takes [laughing] and [sigh], and Grok TTS takes [laugh] and [chuckle]. The tag has to match the API, because a tag can be stripped on one API and spoken aloud on another.

### Can a performance tag be read aloud by mistake?

Yes, on Gemini-TTS. Google says its vocalized-adjective tags such as [scared] and [curious] are spoken aloud, "likely an undesired side effect for most use cases". Bland says older voices strip bracketed text before it reaches the model.

### Is there a way to direct a voice besides SSML and tags?

Yes. OpenAI's `gpt-4o-mini-tts` takes plain-language `instructions`, Inworld's TTS-2 takes steering in plain words, and Google's Gemini-TTS takes a style prompt. OpenAI says `gpt-4o-mini-tts` shuts down on 6 January 2027.

### Which is better for a phone voice agent, SSML or performance tags?

Performance tags suit a voice agent whose job is sounding like a person, since tags such as `[laughs]` and `[say warmly]` direct delivery in plain words, while SSML suits scripts that need mechanical control of pauses and pronunciation. Bland Speech takes tags and timed pauses.
