Emotion control in text to speech comes in three kinds: a named emotion written into the text, a dial that sets how strongly a style applies, and a model that reads the emotion off the sentence. Voxtral TTS 3 documents 34 named emotions, Azure sets a style degree anywhere from 0.01 to 2, and Typecast's Smart Emotion infers the emotion from the sentences around a line, while Bland Speech documents 12 emotion tags.
A caller never hears that count. Bland Speech comes first here because its 12 emotion tags run inside a sentence on the voice that ranked first in Design Arena's listening test of 4 August 2026, with mu-law audio for a phone line, and the facts for each rival are as of November 2026.
What are the best text to speech APIs with emotion control?
The best text to speech API with emotion control for a phone call is Bland Speech, and the widest menu of emotions sits with Voxtral TTS 3, which lists 34 emotion tags against Bland Speech's 12.
- Bland Speech: Best overall for emotion on a phone call
- Azure: Best for a strength dial on each style
- Voxtral TTS 3: Best for 34 named emotions, in public preview
- Typecast: Best for emotion presets and a setting that reads emotion from context
- MiniMax: Best for an emotion parameter set per request, with 40 languages
- Cartesia: Best for speed, volume, and emotion controls on a low-latency model
What these rows mean
- Kind of emotion control: whether a team names the emotion, sets how strong a style is, or leaves the model to infer it, as each company documents it.
- Published price: each company's published rate for its main text to speech model. Typecast bills 1 credit per character, and Cartesia bills 1 credit per character from a monthly plan.
- 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 a text to speech API with emotion control?
A buyer should start with which kind of control the API gives, because each kind fails in its own way. Named tags let a team say exactly where a call turns, and they take a script that someone has marked up. A strength dial helps when one style is right and only its intensity is off. Inferred emotion needs no markup, which suits a team with thousands of scripts, and it still needs listening to confirm that the model read a line the way its writer meant.
The second check is whether the emotion works on the model and language in production. Cartesia says emotion tags are supported only for English and calls emotion controls Beta. MiniMax's Speech 2.8 models don't take the whisper value that its 2.6 models do. Mistral says free-form style instructions aren't available at launch for Voxtral TTS 3. Bland Speech's emotion tags work on its current-generation voices, and older voice providers strip bracketed text before it reaches the model.
The third is the phone. Phone audio is 8 kHz mu-law, the format US phone networks carry, and a buyer should hear the emotion in that format. Bland Speech, Azure, MiniMax, and Cartesia all return it, while Voxtral TTS 3 documents mp3, opus, wav, flac, and pcm at 24 kHz.
Fit comes last, and no tag count measures it. An agent that says "I understand your frustration" in a cheerful voice has an emotion problem that a longer list won't fix.
How do the 6 best text to speech APIs with emotion control compare?
These six APIs differ in who chooses the emotion, the writer, the dial, or the model, and Bland Speech puts the choice in tags a script writer controls, on a voice built for calls.
1. Bland Speech: Best overall for emotion on a phone call
What does it do? Bland Speech is the text to speech model inside Bland, a voice AI provider that builds and runs its own speech model. Emotion is written as a tag right before the words it shapes, from a documented list of 12: [say angrily], [say angrily with force], [say warmly], [say excitedly], [say quietly], [say in a whisper], [say nervously], [say sadly], [say wearily], [say playfully], [say confidently], and [say loudly]. Three intensity tags, [very loud], [very fast], and [voice lowered], push the delivery further. Two numeric controls from 0.0 to 1.0 set how much intonation varies, called expressiveness, and how much the same input varies between renders, called stability.
Who is it for? Bland Speech suits teams putting a voice on phone calls, where an emotion should change at the right clause and the audio should be phone-ready.
Pros
- Bland Speech v3 ranked first of 13 AI voices in Audio Realism Bench's results of 4 August 2026
- 12 emotion tags and 3 intensity tags, documented as the strings the model was trained on
- A stability control that sets how much the same input varies between renders
- Mu-law at 8 kHz, with text streamed in from a language model over a WebSocket
Cons
- More than a hobby project needs, because Bland Speech is built for production calls
Pricing: $0.04 per 1,000 characters, which is $40 per million, with 50,000 characters free for a new account, the same rate whether a line is directed or not. The Bland Speech page has a studio that autocompletes tags as a writer types a bracket.
2. Azure: Best for a strength dial on each style
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, Microsoft says, "automatically detect emotional cues, and adjust speaking style and tone in real time". MAI-Voice-2.1 and MAI-Voice-2.1-Flash are in public preview without a service-level agreement.
Who is it for? Azure suits teams on Microsoft's cloud that work in SSML and want to dial a style up or down rather than switch it on.
Pros
- A style degree from 0.01 to 2
- HD voices that infer emotion from the text
- Mu-law and A-law at 8 kHz, and neural voices at $15 per million characters
Cons
- Styles work on a subset of neural voices, and some voices don't support all SSML tags
- MAI-Voice is in public preview, which Microsoft says isn't recommended for production workloads
Pricing: $15 per million characters for Neural and Neural HD Flash voices, with 0.5 million characters free a month. Microsoft's MAI-Voice-2 ranked 2nd of 13 AI voices on Audio Realism Bench on 4 August 2026, at 1224, and MAI-Voice-2.1 wasn't in those results.
3. Voxtral TTS 3: Best for 34 named emotions, in public preview
What does it do? Voxtral TTS 3, released by Mistral on 20 October 2026, takes 34 emotion tags and two pace tags, written as [emotion=<value>] or [pace=<value>], and an expressiveness parameter from 1.0 to 5.0. It infers emotion from the text itself when a tag is missing, so a tag isn't always needed. The model speaks 35 languages.
Who is it for? Voxtral TTS 3 suits teams that want a long list of named emotions and run on Mistral's EU-hosted API or on their own infrastructure.
Pros
- 34 emotion tags and 2 pace tags
- Emotion inferred from the text when no tag is given
- $13 per million characters
Cons
- Public preview, with free-form style instructions not available at launch
- Output formats are mp3, opus, wav, flac, and pcm, at 24 kHz
- Its voices weren't in Audio Realism Bench's results of 4 August 2026
Pricing: $13 per million characters, with no monthly plan.
4. Typecast: Best for emotion presets and a setting that reads emotion from context
What does it do? Typecast's ssfm-v30 and ssfm-v31 models take seven emotion presets: normal, happy, sad, angry, whisper, toneup, and tonedown. Smart Emotion, Typecast says, "uses AI to automatically understand context then apply the most fitting emotion", and it reads the previous_text and next_text around a line. Pauses go in as <|1s|>.
Who is it for? Typecast suits creators and developers who want an emotion preset per line or an automatic choice, and who work in a studio as much as an API.
Pros
- Seven presets, plus Smart Emotion that reads surrounding text
- An API Lite plan at $15 a month for 200,000 credits
- Emotion controls on its Studio Pro plan at $29 a month
Cons
- Typecast says its current API does not support SSML
- Its free API plan can't be used commercially and requires attribution
- Its voices weren't in Audio Realism Bench's results of 4 August 2026
Pricing: 1 credit per character on its API plans. Free is $0 for 15,000 credits a month, Lite is $15 a month for 200,000 credits at $0.075 per 1,000 credits, and Plus is $280 a month for 4 million credits.
5. MiniMax: Best for an emotion parameter set per request, with 40 languages
What does it do? MiniMax's Speech 2.8 models, speech-2.8-hd and speech-2.8-turbo, released on 23 January 2026, take an emotion parameter per request with values happy, sad, angry, fearful, disgusted, surprised, and calm, and 19 inline interjection tags such as laughs, sighs, and gasps. Whisper and fluent values belong to its 2.6 models. The speech models cover 40 languages with mu-law output.
Who is it for? MiniMax suits teams that set one emotion per request in code and want broad language coverage from one API.
Pros
- An emotion parameter per request and 19 interjection tags
- 40 languages on the 2.6 and 2.8 models
- Audio packages from $5 a month
Cons
- speech-2.8-hd costs $100 per million characters and the turbo model $60
- Speech-02 HD ranked 7th of 13 AI voices on Audio Realism Bench on 4 August 2026, and Speech 2.8 wasn't in those results
Pricing: $100 per million characters for speech-2.8-hd and $60 for speech-2.8-turbo, pay as you go, or monthly audio packages from Starter at $5 for 100,000 audio points to Business at $999.
6. Cartesia: Best for speed, volume, and emotion controls on a low-latency model
What does it do? Cartesia's Sonic 3.6, released on 27 August 2026, speaks 44 languages and takes speed from 0.6 to 1.5, volume from 0.5 to 2.0, and emotion through generation_config or SSML. Cartesia calls the emotion controls Beta and says they are for English only, and says first audio arrives in under 90ms. Sonic returns mu-law at 8 kHz.
Who is it for? Cartesia suits teams that want one fast model with numeric speed and volume controls and are happy to treat emotion as an English-language extra.
Pros
- Speed, volume, and emotion in one request
- Cartesia says first audio arrives in under 90ms
- A Free plan of 20,000 credits a month for evaluation
Cons
- Emotion controls are Beta and English only
- Speed and volume are fixed on Pro Voice Clones
- Sonic 3.5 ranked 6th of 13 AI voices on Audio Realism Bench on 4 August 2026, and Sonic 3.6 came out after those results
Pricing: 1 credit per character from monthly plans: Pro at $5 for 100,000 credits, Startup at $49 for 1.25 million, and Scale at $299 for 8 million, with overages at $65, $45, or $38 per million credits.
Which emotion-control text to speech API sounded most human in a blind test?
Among these six, 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 was 2nd at 1224, Cartesia's Sonic 3.5 was 6th at 1096, and MiniMax's Speech-02 HD was 7th at 1093.
- #1BlandBland Speech v31365
- 2ndMicrosoftMAI-Voice-21224
- 6thCartesiaSonic 3.51096
- 7thMiniMaxSpeech-02 HD1093
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.
Voxtral TTS 3 and Typecast weren't in those results, and neither were newer models such as Sonic 3.6 and MiniMax's Speech 2.8. Design Arena's published results score how real a clip sounds, not how many emotions a model can name, and Design Arena left acting-heavy speech out of scope, so a long list and a high score are separate facts.
How were these text to speech APIs compared?
Samuel Oram-Jones compared six APIs on what kind of emotion control each offers, how many emotions it documents, where those controls stop working, phone audio, price per character, and Design Arena's blind listening score. Rival details come from each company's own pages.
Bland's text to speech comparison sorts the same vendors by job, and the best lists sort them by buying question. A team deciding between markup styles and tags can read SSML vs performance tags, and one directing whole scenes can read text to speech for directing a voice performance.
Frequently asked questions
What is emotion control in text to speech?
Emotion control is any way of telling a text to speech model how a line should feel, whether through a tag like [say sadly], a style with a strength setting, or a model that infers the feeling from the sentence.
Which text to speech API has the most emotions?
Voxtral TTS 3 documents 34 emotion tags, the longest named list here. Bland Speech documents 12 emotion tags, Typecast documents seven presets, and MiniMax's Speech 2.8 takes seven emotion values.
How does a team make an emotion stronger on Bland Speech?
A team adds an intensity tag such as [very loud] or [voice lowered], picks a stronger emotion such as [say angrily with force], or raises the expressiveness control, which sets how much intonation varies from 0.0 to 1.0.
Does emotion control work on a phone call?
Emotion control works on a call when the same model returns phone audio, and Bland Speech, Azure, MiniMax, and Cartesia all return mu-law at 8 kHz.
Does emotion control cost extra?
On Bland Speech it doesn't, because the standard rate of $0.04 per 1,000 characters applies whether a line is directed or not. Typecast includes emotion controls in its Studio Pro plan at $29 a month.
Does emotion control work in every language?
Not always. Cartesia says its emotion tags are supported only for English and lists emotion controls as Beta.
