A cloned voice has two clocks, how long until the clone exists and how long until it starts speaking. Smallest AI says a clone takes under 10 seconds to make, Bland says an instant clone is ready in a minute, and Gradium, Cartesia, Inworld, and Smallest AI all publish a first-audio figure below 100ms for speech. Those are claims about different things, and a voice agent depends on both.
Real time also means the text arrives as a language model writes it and the caller can interrupt.
What is the best real-time voice cloning API?
Bland Speech is the best real-time voice cloning API for a clone that will speak on a live phone call, as of November 2026, because an instant clone from one sample of about 10 seconds is ready in a minute, its WebSocket takes language model tokens as they arrive and lets a new turn cut off the one playing, and Bland Speech v3 ranked first of 13 AI voices at 1365 in Design Arena's Audio Realism Bench results of 4 August 2026.
Bland publishes 199ms to first audio, which is slower than Gradium's under 50ms, Inworld's as low as 50 ms, Smallest AI's sub-100ms, and Cartesia's under 90ms, so a team whose only constraint is milliseconds should weigh Gradium first. Every figure here is the company's own claim, and Design Arena scored each model's own stock voices, so the realism result says nothing about how closely a clone matches its speaker.
- Bland Speech: Best overall for a clone on a live phone call with interruption
- Gradium: Best for the lowest published first-audio figure on a voice agent API
- Cartesia: Best for a sub-90ms claim across 44 languages
- Inworld: Best for a clone from 3 seconds streamed over a sentence-aware WebSocket
- Smallest AI: Best for a clone made in one HTTP call
- Sarvam: Best for streaming a clone in Indian languages
What these rows mean
- Clone from: the sample length each company documents for an instant clone.
- Streaming text in: whether text can arrive in pieces, as a language model writes it, over a WebSocket.
- First-audio figure: each company's own claim, measured by its own method. They are not a race anyone ran.
- 8 kHz mu-law: the format US phone networks carry. Sarvam offers mulaw as a WebSocket codec option.
What makes a voice cloning API real time?
A voice cloning API is real time when a clone can be used the moment it exists, text streams in from a language model, audio streams out, and a caller can interrupt. Bland says a cloned voice is private to the organization and usable immediately, and Smallest AI makes a clone in one POST request and returns a voice ID for any speech call. Sarvam goes furthest on streaming a clone itself, with a WebSocket cloning endpoint in beta.
Streaming text in matters as much as the clone. Bland's WebSocket buffers language model tokens into speech segments and keeps the voice session warm across turns, and a new context ID preempts the turn that is playing, which makes barge-in a normal part of its protocol. Cartesia's WebSocket accepts partial text through contexts, Gradium's accepts text token by token and multiplexes several requests on one socket, and Inworld's sends sentences to the model as soon as it detects a boundary.
The number comes last, because a published figure depends on how it was measured. Inworld says first audio arrives in as little as 50 ms on Realtime TTS-2 Flash, while its model table gives 100 ms for TTS-2 and 20 ms for TTS-2 Flash at the 90th percentile, measured on the server and excluding the network. Smallest AI says sub-100ms and also publishes about 200 ms to first byte at 40 concurrent requests over WebSocket. Gradium says under 50ms and also states an expected time to first token below 300 ms when streaming. A team deciding on latency should time a cloned voice on its own network, under its own load.
How do the 6 best real-time voice cloning APIs compare?
These six differ most in how a clone reaches a streaming connection and what the speed number measures.
1. Bland Speech: Best overall for a clone on a live phone call with interruption
What does it do? Bland Speech makes an instant clone from one clean sample of about 10 seconds and a professional clone from 30 minutes or more of audio after a consent and rights review. A clone returns a voice ID that Bland says is usable immediately with its speech endpoint, and the WebSocket at /v2/tts/ws takes language model tokens as they arrive, and the voice is set when the connection opens. Bland publishes 199ms to first audio on a streamed request, and output arrives as PCM from 8 to 48 kHz or 8 kHz mu-law.
Who is it for? Bland Speech suits a team putting a cloned voice on a live phone agent, through Bland's own endpoints or through the first-party Pipecat service and the Python LiveKit Agents plugin.
Pros
- An instant clone from about 10 seconds, ready in a minute
- A WebSocket that takes language model tokens and supports barge-in
- A first-party Pipecat service and a Python LiveKit Agents plugin
- Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026
- Mu-law and PCM output ready for phone calls
Cons
- More than a one-off narration project needs, because Bland Speech is built for production calls
Pricing: $0.04 per 1,000 characters, which is $40 per million, pay as you go with no monthly plan, and 50,000 characters free once for a new account. The Bland Speech page has a studio to try a clone, and the text to speech API page lists the endpoints.
2. Gradium: Best for the lowest published first-audio figure on a voice agent API
What does it do? Gradium sells real-time speech for voice agents, with a WebSocket that takes text token by token and multiplexes several requests over one socket. It clones a voice from 10 seconds of clear audio, enhances new clones by default since 23 October 2026, and says text to speech starts in under 50ms. Output includes 8 kHz mu-law and A-law, and Gradium supports English, French, German, Spanish, and Portuguese.
Who is it for? Gradium suits a team building a voice agent in those five languages that wants a clone and a first-audio claim on one API.
Pros
- Under 50ms to first audio, by Gradium's claim
- A clone from 10 seconds, with 6 custom voices on the $0 Free plan
- Mu-law and A-law output at 8 kHz
Cons
- Five languages
- A professional clone starts on the $312 M plan, with a one-time training fee of 1 million credits
- Commercial use starts on the $12 XS plan
Pricing: 1 credit per character, with Free at $0, XS at $12 a month, S at $39, M at $312, and L at $1,480, and rates from $58 per million credits on XS to $36 on L.
3. Cartesia: Best for a sub-90ms claim across 44 languages
What does it do? Cartesia sells Sonic 3.6, which streams over HTTP, server-sent events, and a WebSocket whose contexts accept partial text. Cartesia says first audio arrives in under 90ms. An instant clone takes 10 seconds of audio, up to 60 to retain accent, and Cartesia offers professional cloning on Startup and above. Output includes mu-law and A-law at 8 kHz.
Who is it for? Cartesia suits a team that wants one model across 44 languages and on-premise or air-gapped deployment under an enterprise contract.
Pros
- Under 90ms to first audio, by Cartesia's claim
- Pro Voice Clone training is free
- 44 languages from one model
Cons
- Credits come from a monthly plan, with overages at $65, $45, or $38 per million
- Sonic 3.5 ranked 6th of 13 AI voices on Audio Realism Bench on 4 August 2026, and Sonic 3.6, released on 27 August 2026, isn't in 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.
4. Inworld: Best for a clone from 3 seconds streamed over a sentence-aware WebSocket
What does it do? Inworld clones a voice from as little as 3 seconds of audio on every plan and streams over a persistent WebSocket whose autoMode detects sentence boundaries as a language model produces text. Inworld says first audio arrives in as little as 50 ms on Realtime TTS-2 Flash, and its output includes 8 kHz mu-law and A-law. Cloned voices bill at the same rate as stock voices.
Who is it for? Inworld suits a developer building a realtime voice app who wants the least audio for a clone and the stock-voice price.
Pros
- A clone from 3 seconds on every plan
- Sentence-boundary streaming for language model text
- Cloned voices bill at the stock-voice rate
Cons
- Inworld states first audio as low as 50 ms in one place and 20 ms or 100 ms server-side in another
- Zero data retention doesn't cover voice cloning samples
- Professional cloning is in beta and runs on TTS-2 only
Pricing: $25 per million characters for TTS-2 and $15 for TTS-2 Flash on pay as you go, lower on monthly plans from Creator at $25 a month to Growth at $1,500.
5. Smallest AI: Best for a clone made in one HTTP call
What does it do? Smallest AI makes an instant clone from a 5 to 15 second sample with one POST request that returns a voice ID, usable with Lightning v3.1 Pro's streaming and WebSocket endpoints. Smallest AI says speech starts at sub-100ms and publishes about 200 ms to first byte at 40 concurrent requests over WebSocket.
Who is it for? Smallest AI suits a developer who wants a one-call clone at about $19.50 per million characters.
Pros
- A clone from one POST request with a 5 to 15 second sample
- Lightning v3.1 Pro at about $19.50 per million characters
- Mu-law output at 8 kHz
Cons
- Clones made on Lightning v3.1 are affected by its 30 November 2026 retirement
- Lightning v3.1 Pro ranked 11th of 13 AI voices on Audio Realism Bench on 4 August 2026
Pricing: About $0.175 per 10,000 characters for Lightning V3.1 and $0.195 for V3.1 Pro, which is $17.50 and $19.50 per million, with Enterprise custom.
6. Sarvam: Best for streaming a clone in Indian languages
What does it do? Sarvam's Instant Voice Cloning API, in beta, streams over HTTP at POST /voices/clone/stream and over a WebSocket at GET /voices/clone/ws, from a reference clip of 10 to 15 seconds recommended. A clip in one language can speak any of 12 Indian languages. Sarvam says speech starts in under 250 ms in production, and its WebSocket offers mulaw as a codec. Instant Voice Cloning carries a Sarvam watermark, with a detector added on 9 November 2026.
Who is it for? Sarvam suits a team building voice agents for Indian-language callers that wants cloning and streaming on one API.
Pros
- HTTP and WebSocket streaming for cloned voices
- A reference clip in one language can speak another supported language
- A watermark and a detector for cloned speech
Cons
- Instant Voice Cloning is in beta
- Cloned speech covers 12 Indian languages
- Prices are in rupees, excluding GST
Pricing: ₹2.00 per 1,000 characters for real-time voice cloning on every tier, with ₹100 of free credits for every verified user.
What does real-time cloning cost?
Real-time cloning costs the speech rate, plus a plan at most companies, and list prices run from $15 to $58 per million characters, with Bland Speech's $40 in the middle. Smallest AI charges $17.50 to $19.50, Inworld $15 to $25 at list, and Gradium $36 to $58 per million credits by plan. Cartesia charges 1 credit per character from plans of $5, $49, and $299 a month, and Sarvam charges ₹2.00 per 1,000 characters. Bland Speech bills each character as used, with no plan to climb. Bland's voice cloning API pricing guide compares the full bills.
How were these real-time voice cloning APIs compared?
Samuel Oram-Jones compared six APIs on the audio a clone needs, how text streams in, whether a new turn can interrupt, the first-audio figure each company publishes, and phone output. Every rival detail comes from that company's own pricing, documentation, and product pages.
Bland's text to speech comparison hub covers the same companies by job, and the voice cloning comparisons go one provider at a time. A team building the agent itself can read best voice cloning for voice agents, and a team deciding between clone types can read instant vs professional voice cloning.
Frequently asked questions
What is the best real-time voice cloning API?
Bland Speech is the best for a clone on a live phone call. An instant clone takes about 10 seconds of audio, the WebSocket takes language model tokens and supports barge-in, and Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026. Gradium publishes a lower first-audio figure, under 50ms against Bland's 199ms.
Can a cloned voice speak in real time?
Yes. Sarvam streams cloned speech over HTTP and a WebSocket, Smallest AI documents clones on its streaming and WebSocket endpoints, and Bland's clone returns a voice ID that works with its speech calls.
Which voice cloning API has the lowest latency?
By published claims, Gradium says under 50ms and Inworld as low as 50 ms, Cartesia under 90ms, Smallest AI sub-100ms, Sarvam under 250 ms, and Bland 199ms, each measured by the company's own method.
How fast can a voice be cloned?
Smallest AI says a clone takes under 10 seconds to make from a 5 to 15 second sample, and Bland says an instant clone is ready in a minute. Inworld needs as little as 3 seconds of audio, and Gradium and Cartesia need 10.
Do real-time voice cloning APIs work on phone calls?
Yes, with 8 kHz mu-law output. Bland Speech, Gradium, Cartesia, Inworld, and Smallest AI return mu-law, and Sarvam offers mulaw as a WebSocket codec.
Does a cloned voice need a plan?
At Gradium, commercial use starts on the $12 XS plan, and Cartesia bills credits from monthly plans. Bland Speech bills pay as you go with no monthly plan, and Inworld includes instant cloning on every plan.
