A clone made from a careful, flat read of a script stays a careful, flat voice for as long as it exists. Mistral says a clone "copies more than the timbre of the speaker" and picks up the emotion and pace of its recording, and Murf says styles aren't offered on cloned voices because a clone already carries the delivery of its reference sample.
That makes expressive cloning a decision at the microphone before it is a decision about an API. Six APIs are ranked on what each company says a recording passes on, which controls remain for steering a voice afterward, and what a clone costs to run, as of November 2026. Bland Speech costs $40 per million characters, and Murf, Mistral, and Grok TTS charge less than half of that.
What is the best expressive voice cloning API?
Bland Speech is the best expressive voice cloning API for a voice that has to perform on phone calls, because a clone made on its default engine is a BTTS V3 voice, the generation Bland documents performance tags and expressiveness and stability controls for.
That ranking rests on how each company's published guidance fits together, not on a test of clones. Bland Speech v3 ranked first of 13 AI voices on Design Arena's Audio Realism Bench, a blind listening test, on 4 August 2026, but Design Arena's results scored one female and one male voice per model, each its most naturally conversational American-English voice, so they say nothing about how a particular person's clone sounds.
- Bland Speech: Best overall for a clone that stays expressive on a phone call
- Mistral Voxtral TTS 3: Best for the plainest guidance on recording delivery, with 34 emotion tags
- Grok TTS: Best for free console clones and speech tags on the same endpoint
- Inworld: Best for cloning from 3 seconds and steering in plain words
- Murf: Best for a low rate on clones that keep a reference sample's delivery
- Cartesia: Best for a professional clone that learns pacing and loudness
What these rows mean
- What the recording passes on: what each company says about how a recording shapes a clone.
- Controls documented: what each company documents for steering its voices.
- Price per million characters: each company's list price for speech. Murf's $10 is $0.01 per 1,000 characters, and Cartesia's range divides each plan's price by its credits.
Where does a clone's expression come from?
A clone's expression comes from its recording first, and Bland Speech, Mistral, Grok TTS, Inworld, and Murf all say so in their own guidance.
Mistral's advice is to "record the clip in the style you want to hear". Grok TTS says to speak expressively because the voice will match the recording, and warns that a speaker who sounds like they are reading a script gets a voice that does the same. Bland asks for emotional variety, and Inworld asks for "a variety of emotions to capture the full range of the voice". Grok TTS adds that a support voice should be recorded on realistic support exchanges, matched to the content the voice will speak.
The room gets cloned as well. Grok TTS says background noise is cloned along with the voice, so a rep recorded beside an air conditioner brings the air conditioner onto every call.
Once a clone exists, the controls move it only so far. Mistral says higher expressiveness values trade voice similarity and stability for stronger delivery, and Bland says its expressiveness and stability dials pull against each other, so high expressiveness with low stability gives a different animated read each time. A better recording is the safer fix for a flat clone than a higher dial, and Murf says a clone can't be updated, so a different result means a new clone.
How do the 6 best expressive voice cloning APIs compare?
These six APIs differ most in what a clone inherits from its recording and what a team can still steer, and Bland Speech documents its sample guidance, V3 controls, and performance tags together.
1. Bland Speech: Best overall for a clone that stays expressive on a phone call
What does it do? Bland Speech makes an instant clone from one clean sample of about 10 seconds, and Bland's guidance for that sample is one speaker, no background noise, music, or reverb, and emotional variety in the delivery. A professional clone is fine-tuned on 30 minutes or more of audio after Bland's consent and rights review. Every clone is private to the customer's organization, and the customer confirms the voice is theirs to use.
A clone on the default engine is a BTTS V3 voice. Bland documents expressiveness and stability, each from 0.0 to 1.0, as calibrated for BTTS V3, and documents 30 performance tags for current-generation voices, such as [say warmly], [laughs], and [sighs], which Bland calls "the set the model was trained on". A clone's voice ID goes into POST /v2/tts like any other voice, and output comes as PCM from 8 to 48 kHz or 8 kHz mu-law. Bland Speech is a closed-source model Bland built itself, and it suits a team cloning one speaker's voice for callers.
Pros
- Sample guidance that asks for emotional variety, and clones that are BTTS V3 voices
- 30 documented performance tags plus expressiveness and stability controls
- A professional clone fine-tuned on 30 minutes or more, after a consent and rights review
- 8 kHz mu-law and PCM up to 48 kHz
- Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026
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, the same in the studio or the API, directed or not, pay as you go with no monthly plan. A new account gets 50,000 characters free once. Bland Speech has a studio for trying a clone.
2. Mistral Voxtral TTS 3: Best for the plainest guidance on recording delivery, with 34 emotion tags
What does it do? Voxtral TTS 3 has been in Public Preview since 20 October 2026. It clones a voice from a recording of a few seconds to a few minutes with no training, and Mistral recommends 30 minutes to 3 hours of clean audio for a professional clone. Mistral says a clone picks up the emotion and pace of the recording and tells a speaker to record in the style they want to hear. It documents 34 emotion tags and 2 pace tags, infers emotion from the text when no tag is present, and offers an expressiveness setting from 1.0 to 5.0. Voxtral TTS 3 suits a team that wants a larger emotion vocabulary and will accept a preview model.
Pros
- Direct advice to record a clip in the style wanted
- 34 emotion tags, 2 pace tags, and an expressiveness setting
- EU-hosted API or self-deployment
Cons
- Voxtral TTS 3 is in Public Preview
- Style instructions aren't available at launch
Pricing: $13 per million characters.
3. Grok TTS: Best for free console clones and speech tags on the same endpoint
What does it do? Grok TTS, from xAI (now SpaceXSI), clones a custom voice from a reference clip of up to 120 seconds, and recommends 90 to 120 seconds because clips under 30 seconds may lack detail. Grok TTS says the model picks up the delivery patterns of the clip. A custom voice is interchangeable with a built-in voice on POST /v1/tts, which takes 14 inline tags such as [laugh] and [sigh] and 12 wrapping tags such as <whisper>. Audio from its voice APIs carries an invisible watermark by default, custom voices included. Grok TTS suits a US team that will clone in the xAI console and wants tags on one endpoint.
Pros
- Custom voices free to create in the console
- Recording guidance by use case
- 14 inline tags and 12 wrapping tags
Cons
- Custom voices are available only in the United States, except Illinois
- Cloning through the API is gated to teams on an Enterprise plan
Pricing: $15 per million characters, pay as you go, with no plan tiers or monthly fee.
4. Inworld: Best for cloning from 3 seconds and steering in plain words
What does it do? Inworld clones from as little as 3 seconds of audio, with a 30-second cap, on every plan, and asks a speaker to use a variety of emotions. Professional cloning, in beta, needs at least 10 minutes and works on TTS-2 only. TTS-2 takes a description of delivery in plain words, either as an instruction or in square brackets partway through a line, and both TTS-2 and TTS-2 Flash render 50 non-verbal sounds such as [laugh] and [sigh]. Inworld suits a team that wants the shortest recording.
Pros
- Cloning from 3 seconds on every plan
- Delivery described in plain words on TTS-2
- Cloned voices billed at the stock-voice rate
Cons
- Steering in words works on TTS-2 only, and TTS-2 Flash ignores it
- Zero data retention doesn't cover clone samples
Pricing: On-Demand is $0 a month and pay as you go at $25 per million characters for TTS-2 and $15 for TTS-2 Flash. Creator is $25 a month with TTS-2 at $20 per million.
5. Murf: Best for a low rate on clones that keep a reference sample's delivery
What does it do? Murf clones on Falcon 2.1, and its instant cloning is Enterprise only, through the API, switched on by Murf's sales team. One reference sample of up to 30 seconds becomes a clone in a few minutes, and a professional clone uses up to 90 minutes of studio recordings and takes one to four weeks because Murf's team builds it. Murf says styles aren't available on clones, since a clone already carries the delivery of its reference sample, while clones take the same rate and pitch controls as library voices. Murf suits an enterprise team that will record one sample carefully.
Pros
- $0.01 per 1,000 characters, the lowest rate here
- No extra charge for creating or keeping a clone
Cons
- Instant and professional cloning are Enterprise only
- Styles aren't available on cloned voices, and a clone can only be replaced, not updated
Pricing: $0.01 per 1,000 characters for Falcon 2.1 on pay as you go, $10 per million, and a clone adds nothing. Enterprise API pricing is custom.
6. Cartesia: Best for a professional clone that learns pacing and loudness
What does it do? Cartesia sells instant clones from 10 seconds of audio, with up to 60 seconds on Sonic 3.6 to keep a speaker's accent, and professional clones on Startup and above, with training that costs nothing. Cartesia says a professional clone learns pacing and loudness from its dataset, so speed and volume controls don't apply to it. Sonic's emotion controls are in beta and English only. Cartesia suits a team that will record a long, consistent dataset.
Pros
- Free training for a professional clone
- Instant clones from 10 seconds
- Mu-law and A-law at 8 kHz
- Cartesia says first audio arrives in under 90ms
Cons
- Speed and volume controls don't apply to professional clones
- Emotion controls are in beta and English only
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.
How were these expressive voice cloning APIs compared?
Samuel Oram-Jones compared six APIs on what each company says a recording passes on to a clone, the controls each documents for steering a voice, the audio behind an instant and a professional clone, and price per million characters. Rival details are what each company publishes, and Bland Speech sits first on how its guidance, controls, and tags fit together, not on price or audio length.
Bland's voice cloning APIs ranked on audio and price and the most expressive text to speech APIs cover other angles, its text to speech API page has the endpoints, and its voice cloning comparisons sit inside the text to speech comparison hub.
Frequently asked questions
What is the best expressive voice cloning API?
Bland Speech is the best expressive voice cloning API for clones that will speak on phone calls. Bland asks for emotional variety in the sample, a clone on its default engine is a BTTS V3 voice with tags and dials documented for it, and output comes as 8 kHz mu-law.
Does a cloned voice keep the emotion of its recording?
Yes, at the companies that say so. Mistral says a clone picks up the emotion and pace of the recording, and Grok TTS says the voice matches the expressiveness of the recording. Bland says emotional variety in the sample yields a more expressive clone.
How should a recording be made for an expressive clone?
In the style the voice will use, in a quiet room, with one speaker and the kind of content the voice will speak. Grok TTS recommends 90 to 120 seconds, Mistral says to record in the style wanted, Inworld asks for a variety of emotions, and Bland asks for no background noise, music, or reverb.
Can a cloned voice be steered with tags after cloning?
Partly, and the companies differ. A Bland Speech clone is a BTTS V3 voice, and Bland documents tags and dials for current-generation voices. Grok TTS clones work on the endpoint that takes its tags. Murf says rate and pitch work on clones and styles don't, and Cartesia says speed and volume don't apply to professional clones.
What is the cheapest expressive voice cloning API?
Murf at $0.01 per 1,000 characters, which is $10 per million, though its cloning is Enterprise only. Mistral's Voxtral TTS 3 costs $13 per million, Grok TTS $15, and Inworld $15 on TTS-2 Flash and $25 on TTS-2. Bland Speech costs $40 per million, with no monthly plan.
