# What makes an AI voice expressive? Tags, prompts, and controls in 2026

By Samuel Oram-Jones

A voice agent can say "I understand" to an angry caller and to a confused one in exactly the same tone, and better wording does not fix that. Delivery is decided somewhere else: inside the model, in a plain-English prompt, in tags placed in the script, or in a number set on the request, and every text to speech provider publishes a different mix of those four.

That mix matters to a buyer for a practical reason, which is that a script written for one provider rarely moves to another unchanged. Bland Speech takes its direction from bracketed tags and two numeric controls.

## What makes an AI voice sound expressive?

An AI voice sounds expressive when its delivery follows the meaning of a line, which happens either because the model works out the emotion from the text or because a person writes the delivery into the script.

Models that work it out alone are the easier ones to use. Azure says its HD voices "automatically detect emotional cues, and adjust speaking style and tone in real time", and Amazon describes Amazon Polly's Generative engine as "emotionally engaged, assertive, and highly colloquial". ElevenLabs calls Eleven v4 "Our most emotionally rich, expressive speech synthesis model", and ElevenLabs still advises that "Explicit dialogue tags yield more predictable results than relying solely on context."

That last line explains why direction exists. A model that reads emotion from text will sometimes read it wrong, and a phone agent that greets every caller with the same line has no use for sometimes. Providers therefore add ways for a writer to say what a line should sound like, and those ways fall into four kinds.

## How do text to speech providers let a developer direct delivery?

As of November 2026, providers offer four ways to direct delivery, which are inline tags, a style prompt, markup, and numeric controls, and each provider publishes its own mix of them.

### Inline tags

Bland Speech takes bracketed tags that sit right before the words they shape, such as [laughs] or [say warmly], and Bland says the model performs them. Bland's published set covers emotions ([say warmly], [say excitedly], [say in a whisper]), intensity ([very loud], [very fast], [voice lowered]), and vocalizations ([laughs], [sighs], [clears throat], [gasps]). Bland says to use those strings verbatim because they are "the set the model was trained on", and tags work on Bland's current-generation voices only.

ElevenLabs documents audio tags such as [whispering], [shouting], and [laughing] on Eleven v4. Google's Gemini-TTS takes inline tags in four groups: non-speech sounds such as [sigh], style modifiers such as [whispering], vocalized adjectives such as [curious], and pacing such as [short pause]. Grok TTS offers inline tags such as [laugh] and [sigh] plus wrapping tags such as `<whisper>` and `<loud>` that surround the words they affect.

Fish Audio goes the other way. Its S2 models take natural-language bracket tags "not limited to a fixed set", so a writer can put [whispers sweetly] or [laughing nervously] in a script. A fixed set can be checked against a published list, while an open-ended one has to be tested phrase by phrase, and that trade is the one a team makes when it picks between them.

### A style prompt

OpenAI's gpt-4o-mini-tts takes an `instructions` prompt in plain English, such as "Speak in a cheerful and positive tone.", covering accent, emotional range, intonation, impressions, speed, tone, and whispering. OpenAI also says that gpt-4o-mini-tts, tts-1, and tts-1-hd shut down on 6 January 2027, so steering by prompt on that endpoint is tied to a model with a published end date.

Google's Gemini-TTS has a style prompt too, which Google calls "The primary driver of the overall emotional tone and delivery."

### Markup

SSML, short for Speech Synthesis Markup Language, is a W3C Recommendation for wrapping text in tags that control pauses, pace, volume, and emphasis. Amazon Polly supports a subset of SSML 1.1 covering pauses, pace, volume, emphasis, phonetic pronunciation, breathing sounds, and whispering, plus a Newscaster style that its Generative engine does not support. Azure uses an express-as element for styles, style degree, and role on a subset of its neural voices, and its HD voices accept styles such as whispering and ecstatic.

ElevenLabs' newest models take another route. ElevenLabs says "Eleven v4 and Eleven v3 do not support SSML break tags", so pauses come from punctuation and audio tags. Google's Chirp 3: HD takes a `speaking_rate` from 0.25 to 2.0 and pause tags for delivery, and Google says "SSML tags are not currently supported for streaming requests."

### Numeric controls and pause markers

Cartesia offers speed from 0.6 to 1.5 and volume from 0.5 to 2.0, and its emotion controls are in beta and work in English only. Bland Speech has two controls on a scale from 0.0 to 1.0: expressiveness, which sets how much intonation varies, and stability, which sets how much the same input varies between renders. Timed pauses are written as `<|0.5|>` and can run from 0.1 to 10.0 seconds.

Directing a voice can also change what a script costs. Cartesia counts each break tag as one credit, and Bland charges one standard rate, $0.04 per 1,000 characters, whether a script is directed or not.

## How do listeners judge how human an AI voice sounds?

Design Arena judges how human an AI voice sounds by playing listeners two recordings of the same transcript, blind and in random order, and asking which voice sounds more like a real person.

Each choice is one vote, and Design Arena turns the votes into an Elo rating with the Bradley-Terry model. Elo is a rating where a higher number means a voice is chosen as more human more often. Design Arena also enters a human reading of each prompt into the pool, so every model is measured against a real person.

The prompts are 500 or more private scripts, most up to roughly 40 words, split into three categories: Phone Agents, Explainers, and Conversations. Each model is represented by one female and one male American-English voice. Design Arena says acting-heavy speech such as game characters and dramatized fiction is out of scope for this version and "deferred to a future expressive track", which means the test measures how conversational a voice sounds, not how wide its emotional range is.

In Design Arena's [Audio Realism Bench results of 4 August 2026](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/), Bland Speech v3 ranked first of 13 AI voices at 1365, against 1500 for the human recording. Design Arena's article names Microsoft's MAI-Voice-2 second at 1224 and xAI's Grok TTS third at 1156. A voice can sit high on a realism test and still need tags to play a line sad or amused, so the two questions stay separate.

## What goes wrong when a team directs a voice?

Direction goes wrong in four documented ways: a tag is read aloud, a script is tied to one provider's syntax, a control stops applying to some voices, and the model carrying the instructions is retired.

### A tag is read aloud

Google warns that its vocalized-adjective tags, such as [scared], [curious], and [bored], are spoken aloud, and calls that "likely an undesired side effect for most use cases". Bland says older voice providers strip bracketed text before it reaches the model, which is why Bland's tags work on current-generation voices only.

### A script is tied to one syntax

[whispering] is how ElevenLabs and Google write a whisper, Bland writes [say in a whisper], and Grok TTS wraps the words in `<whisper>`. A script full of one syntax needs rewriting for another, which is a reason to keep delivery directions in a layer a team controls, and the trade between [SSML and bracketed tags](/blog/ssml-vs-performance-tags) is a question of its own.

### A control stops at some voice types

Cartesia says speed and volume are fixed on Pro Voice Clones, which learn pacing and loudness from a dataset. A team that tuned speed on a stock voice finds the knob gone on a clone.

### The instruction model is retired

The OpenAI case above is the clearest one: instructions work on gpt-4o-mini-tts, and gpt-4o-mini-tts has a shutdown date. A team that chose it for steering has to rebuild the steering on a new model.

## Where does Bland Speech fit among expressive text to speech options?

Bland Speech fits phone agents whose delivery is written into the script, because direction comes from a published tag set and two numeric controls, and it costs the same standard rate, $0.04 per 1,000 characters, as a plain read.

Bland's studio, where a line renders immediately on the same models the API uses, autocompletes tags when a writer types `[`, and it makes 1, 2, or 3 takes of one script in a single run so a team can hear how much a render varies. Bland's director, Norm, works on one selected voice: given a scene, it writes the words, the stumbles, and the timing, then performs them. Norm lives in Bland's dashboard studio.

The catch is the one the fixed list implies. Fish Audio accepts open-ended directions that Bland's set does not, and Design Arena's realism test leaves out acting-heavy speech, so a team building dramatized characters should audition those voices on their own scripts. A new Bland account gets 50,000 free characters once, enough for that audition.

<div class="cmp-usewhen"><div class="cmp-usewhen-own"><p><strong>Use Bland Speech when</strong></p><ul><li>A team will write delivery into scripts with bracketed tags from a published list and tune two controls from 0.0 to 1.0</li><li>A phone agent will repeat the same lines and the team needs repeat renders to stay close, which the stability control sets</li></ul></div><div><p><strong>Use another provider when</strong></p><ul><li>A team will maintain SSML documents against the subset each voice supports, as Amazon Polly and Azure publish it</li><li>A team will write open-ended directions in plain language and test each phrase, as Fish Audio's bracket tags allow</li></ul></div></div>

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>A published tag set that includes [laughs], [sighs], and [clears throat]</li><li>Expressiveness and stability controls from 0.0 to 1.0, plus pause markers from 0.1 to 10.0 seconds</li><li>One standard rate, $0.04 per 1,000 characters, pay as you go, directed or not</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>

[Bland Speech](/speech) carries these controls, Bland documents the endpoints that take them with its [text to speech API](/text-to-speech-api), and Bland's [text to speech comparison](/compare/text-to-speech) and [developer guides](/compare/text-to-speech/guides) set the providers side by side.

## Frequently asked questions

### What is an expressive text to speech voice?

An expressive text to speech voice is one whose delivery changes with the meaning of a line, through tone, pace, and sounds such as laughs or sighs. Providers get there by inferring emotion from the text, by tags in the script, by a style prompt, or by numeric controls.

### Do emotion tags work on every text to speech API?

No, tags are published per provider and often per model. Bland's tags work on its current-generation voices, ElevenLabs documents audio tags on Eleven v4 and v3, Google documents them for Gemini-TTS, and Google's Chirp 3: HD takes pause tags and a speaking rate.

### Can an AI voice laugh or sigh?

Yes, on providers that publish the tag. Bland Speech takes [laughs] and [sighs], ElevenLabs takes [laughing], Google's Gemini-TTS takes [sigh] and [laughing], and Grok TTS takes [laugh] and [sigh].

### What is SSML, and do voice agents still use it?

SSML is a W3C Recommendation for marking up text with pauses, pace, volume, and emphasis. Amazon Polly and Azure publish it, ElevenLabs says "Eleven v4 and Eleven v3 do not support SSML break tags", and Bland Speech takes bracketed tags and pause markers instead.

### Does directing a voice cost more?

On Bland Speech it does not: one standard rate of $0.04 per 1,000 characters applies directed or not, studio or API, and a new account gets 50,000 free characters once. Cartesia counts each break tag as one credit.

### How can a team check a direction before putting it in code?

A team can render the line in Bland's studio, which uses the same models as the API, and compare takes. The studio makes 1, 2, or 3 takes from one script in a run.
