Design Arena left acting out of its Audio Realism Bench on purpose. Design Arena says "acting-heavy speech (game characters, dramatized fiction, persona-driven companion voices) is out of scope for this version", so a high score there says a voice sounds like a person talking, and says nothing about how far it can be pushed into a character.
That split decides this list, and the facts here are as of November 2026. Bland Speech comes first because it pairs 30 documented delivery tags with the voice that ranked first in the listening test on 4 August 2026 and with audio built for a phone line. Fish Audio, Inworld, and Voxtral TTS 3 each offer more tags, which matters more for narration and characters than for a call.
What are the most expressive text to speech APIs?
The most expressive text to speech API for a phone call is Bland Speech, while the widest control over a performance sits with Fish Audio, Inworld, and Voxtral TTS 3.
- Bland Speech: Best overall for an expressive voice on a phone call
- ElevenLabs: Best for audio tags on a model it calls its most emotionally rich
- Fish Audio: Best for directing a performance in plain language
- Grok TTS: Best for sound tags and style tags at $15 per million characters
- Inworld: Best for 50 non-verbal sounds and steering in plain words
- Voxtral TTS 3: Best for a catalog of 34 named emotions, in public preview
- Azure: Best for teams on Azure that want styles with a strength dial
What these rows mean
- Price per million characters: each company's list price for its main model. Fish Audio bills per million UTF-8 bytes, close to one byte per character in English.
- Audio Realism Bench: Design Arena's blind listening results of 4 August 2026, by place among 13 AI voices and score, for the model named in the cell.
What should a buyer check in an expressive text to speech API?
A buyer should check three things before counting tags: whether the vocabulary is open or closed, where the control sits, and whether it survives a live call.
An open vocabulary accepts any phrase in brackets, as Fish Audio's does, while a closed one accepts a fixed list the model was trained on, as Bland Speech's does. Open is more flexible, and a closed list is easier to test on a phone script.
Control sits in tags inside the text for most of these, while Inworld's TTS-2 also takes a plain-words instruction and Azure uses markup with a style and a strength. A tag that changes mid-sentence suits a phone agent that has to sound relieved in one clause and careful in the next.
Live calls add limits that a demo hides. Bland Speech's tags work on its current-generation voices, and older voice providers strip bracketed text before it reaches the model, so a team should confirm that its streaming path and phone format use the same model as the demo.
How do the 7 most expressive text to speech APIs compare?
These seven APIs split on whether they were built for a performance or a conversation, and Bland built its voice for calls.
1. Bland Speech: Best overall for an expressive voice on a phone call
What does it do? Bland Speech runs on Bland's own speech model, and delivery is shaped with performance tags in square brackets that sit right before the words they shape: 12 emotions such as [say warmly] and [say nervously], three intensity tags such as [voice lowered], and 15 vocal sounds such as [laughs], [sighs], and [clears throat]. Expressiveness and stability run from 0.0 to 1.0, a timed pause of 0.1 to 10.0 seconds goes in as <|N|>, and the studio's Norm director writes and performs a script from a scene description. Audio comes back as mu-law or as PCM from 8 to 48 kHz.
Who is it for? Bland Speech suits teams putting a voice on phone calls and voice agents, where delivery should move with the conversation without sounding performed.
Pros
- Bland Speech v3 ranked first of 13 AI voices in Audio Realism Bench's results of 4 August 2026
- 30 documented tags in three groups: emotion, intensity, and vocal sounds
- Mu-law at 8 kHz and a WebSocket that takes text from a language model
- One standard rate, $0.04 per 1,000 characters, whether the delivery is directed or not
Cons
- A fixed list of 30 tags rather than an open vocabulary, so a team directing a character's performance will want more range
Pricing: $0.04 per 1,000 characters, which is $40 per million, with 50,000 characters free for a new account and credits that never expire. Tags run in the Bland Speech studio and through the text to speech API.
2. ElevenLabs: Best for audio tags on a model it calls its most emotionally rich
What does it do? ElevenLabs calls Eleven v4 its "most emotionally rich, expressive speech synthesis model", launched on 28 September 2026. Audio tags such as [whispering], [shouting], and [laughing] work on Eleven v4, and the same techniques apply to Eleven v3. Neither model takes SSML break tags, so pauses come from punctuation and audio tags. The v4 family speaks 90+ languages.
Who is it for? ElevenLabs suits teams that want tag control on a large multilingual platform.
Pros
- Audio tags on Eleven v4 and v3
- 90+ languages on the Eleven v4 family
- Eleven v4 Turbo at $0.04 per 1,000 characters
Cons
- Eleven v3 ranked 12th of 13 AI voices on Audio Realism Bench on 4 August 2026
- Eleven v4, its most expressive model, costs $0.08 per 1,000 characters
- The standard text to speech WebSocket doesn't take Eleven v3 or v4
Pricing: $0.08 per 1,000 characters for Eleven v4 and v3, and $0.04 for v4 Turbo, billed in dollars on plans from Free to Business at $990 a month. Eleven v4 came out after Audio Realism Bench's results of 4 August 2026 and isn't in them.
3. Fish Audio: Best for directing a performance in plain language
What does it do? Fish Audio's S2.1-Pro, its recommended production model, takes natural-language tags in square brackets that aren't limited to a fixed set, such as [whispers sweetly] and [laughing nervously]. Drama 3, launched on 13 October 2026, is a model Fish titles "The Most Emotive Text-to-Speech Model We've Built", and it takes direction in plain language, a sound or a pause placed where it happens, and a delivery limited to specific words.
Who is it for? Fish Audio suits teams producing dialogue, characters, and narration, where a director's note in plain words is the natural way to work.
Pros
- Tags aren't limited to a fixed set
- Drama 3 limits a delivery to specific words
- S2.1-Pro at $15 per million UTF-8 bytes, with no subscription for API access
Cons
- Drama 3 costs $50 per million UTF-8 bytes at list price, with a launch price of $10 through 30 November 2026
- Its voices weren't in Audio Realism Bench's results of 4 August 2026
Pricing: $15 per million UTF-8 bytes for S2.1-Pro and $50 for Drama 3, pay as you go.
4. Grok TTS: Best for sound tags and style tags at $15 per million characters
What does it do? Grok TTS is the text to speech API in the Grok API from xAI, now named SpaceXSI, generally available since March 2026. It takes 14 inline tags such as [laugh], [sigh], [pause], and [breath], and 12 wrapping tags such as <whisper>, <soft>, <loud>, and <singing>. It streams over HTTP and a WebSocket, returns mu-law and A-law at 8 kHz, and speaks 20 languages.
Who is it for? Grok TTS suits teams that want a fixed, documented set of sound and style tags on a pay as you go rate, including on phone audio.
Pros
- 14 inline tags and 12 style tags
- Grok TTS ranked 3rd of 13 AI voices on Audio Realism Bench on 4 August 2026, at 1156
- Mu-law and A-law at 8 kHz, with $15 per million characters
Cons
- Control is a fixed list of tags
Pricing: $15 per million characters, pay as you go, with no plan tiers.
5. Inworld: Best for 50 non-verbal sounds and steering in plain words
What does it do? Inworld's Realtime TTS-2, launched on 5 May 2026, takes steering in plain words, either as a per-request instruction or in brackets where the delivery changes partway through a line. Both TTS-2 and TTS-2 Flash render 50 non-verbal sounds such as [laugh] and [sigh], and pauses use SSML break tags. Inworld returns mu-law and A-law at 8 kHz.
Who is it for? Inworld suits teams building consumer apps, companions, and games.
Pros
- 50 non-verbal sounds on both models
- Steering in plain words, per request or inline
- Inworld says first audio arrives in as little as 90ms at P90 for TTS-2, measured on Coval
Cons
- Steering works on TTS-2 only, and TTS-2 Flash ignores it
- TTS-1.5 Max ranked 13th of 13 AI voices on Audio Realism Bench on 4 August 2026, at 763, and TTS-2 wasn't in those results
Pricing: On-Demand is $0 a month, with TTS-2 at $25 per million characters and TTS-2 Flash at $15.
6. Voxtral TTS 3: Best for a catalog of 34 named emotions, in public preview
What does it do? Voxtral TTS 3 is Mistral's text to speech model, released on 20 October 2026 and in public preview. It takes 34 emotion tags and two pace tags written as [emotion=<value>] or [pace=<value>], infers emotion from the text when a tag is missing, and has an expressiveness parameter from 1.0 to 5.0. The model speaks 35 languages.
Who is it for? Voxtral TTS 3 suits teams that want to name an emotion from a long list on Mistral's EU-hosted API or on their own infrastructure.
Pros
- 34 emotion tags and 2 pace tags
- $13 per million characters
- An expressiveness parameter from 1.0 to 5.0
Cons
- Public preview, and free-form style instructions aren't available at launch
- Its voices weren't in Audio Realism Bench's results of 4 August 2026
Pricing: $13 per million characters for Voxtral TTS 3, with no monthly plan.
7. Azure: Best for teams on Azure that want styles with a strength dial
What does it do? Azure Speech sets a speaking style, a style degree from 0.01 to 2, and a role at the sentence level through mstts:express-as on a subset of neural voices. Its HD voices infer emotion from the text and take styles such as whispering and ecstatic. Microsoft's own MAI-Voice-2.1 and MAI-Voice-2.1-Flash are in public preview without a service-level agreement, and Azure returns mu-law at 8 kHz.
Who is it for? Azure suits teams already on Microsoft's cloud.
Pros
- A style degree from 0.01 to 2
- Microsoft's MAI-Voice-2 ranked 2nd of 13 AI voices on Audio Realism Bench on 4 August 2026, at 1224
- Neural voices at $15 per million characters, with 0.5 million characters free a month
Cons
- Styles work on a subset of neural voices, and some voices don't support all SSML tags
- MAI-Voice is in public preview, without a service-level agreement
Pricing: $15 per million characters for Neural and Neural HD Flash voices, with 0.5 million characters free a month. MAI-Voice-2.1, the version Azure offers now, wasn't in the results of 4 August 2026.
Which expressive text to speech API sounded most human in a blind test?
Among these APIs, Bland Speech v3 sounded most human in Design Arena's Audio Realism Bench results of 4 August 2026, at 1365, and only the human recordings scored higher, at 1500. Microsoft's MAI-Voice-2 followed in 2nd at 1224, Grok TTS was 3rd at 1156, ElevenLabs' Eleven v3 was 12th at 938, and Inworld's TTS-1.5 Max was 13th at 763.
- #1BlandBland Speech v31365
- 2ndMicrosoftMAI-Voice-21224
- 3rdxAIGrok TTS1156
- 12thElevenLabsEleven v3Near the bottom938
- 13thInworldTTS-1.5 MaxNear the bottom763
Ranks are among 13 AI voices, and each rival shows the highest listed voice available on it. Real human recordings scored 1500, at the right edge of this scale. Bars start at 600, not zero. Source: Design Arena's published results.
Fish Audio's and Mistral's voices weren't in those results, and neither were newer models such as Eleven v4. Design Arena's published results score how real a clip sounds to a listener, not how many tags a model accepts.
How were these text to speech APIs compared?
Samuel Oram-Jones compared seven APIs on the size and openness of each tag vocabulary, numeric delivery controls, where the control sits in a request, streaming, phone audio, price per character, and Design Arena's blind listening score of 4 August 2026.
The same vendors appear by job in Bland's text to speech comparison, and the best lists sort them by buying question. A team weighing tags against SSML can read SSML vs performance tags, and one that wants to know what makes a voice expressive can read what makes an AI voice expressive.
Frequently asked questions
What is the most expressive text to speech API?
The answer depends on the work. Bland Speech tops this list for phone calls and voice agents, because Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026 and ships 30 documented tags. Fish Audio's Drama 3 and ElevenLabs' Eleven v4, which each company describes as its most emotive or emotionally rich model, are built more for characters and narration.
Which text to speech API has the most tags?
Fish Audio takes any phrase in brackets, Inworld documents 50 non-verbal sounds, and Voxtral TTS 3 documents 34 emotion tags and 2 pace tags. Bland Speech documents 30 tags, and Grok TTS documents 14 inline tags and 12 style tags.
Do expressive tags work on a phone call?
They work when the model that accepts the tags also streams and returns phone audio. Bland Speech returns mu-law at 8 kHz and streams text from a language model over a WebSocket, and its tags work on current-generation voices.
Is a high Design Arena score the same as being expressive?
No. Listeners pick the clip that sounds more like a real person, and acting-heavy speech is out of scope, so a score shows how human a voice sounds in conversation rather than how many characters it can play.
Does directing delivery cost extra on Bland Speech?
No. Bland Speech charges one standard rate of $0.04 per 1,000 characters, directed or not, in the studio or the API. ElevenLabs charges $0.08 per 1,000 characters for Eleven v4 and Fish Audio prices Drama 3 at $50 per million UTF-8 bytes.
