# How much audio does it take to clone a voice in 2026?

By Samuel Oram-Jones

Inworld says it can clone a voice from as little as 3 seconds of audio, and Cartesia sets 30 minutes as the minimum for a professional clone. Both numbers are true, because the two companies are selling different things, and a buyer who reads only one of them plans the wrong recording session.

Bland Speech sits in the middle of that range: one clean sample of about 10 seconds for an instant clone, and 30 minutes or more for a professional one. Providers also disagree about what counts as enough, and most of them pair a length with a demand about how clean the recording has to be.

## How many seconds of audio does an instant voice clone need?

An instant clone needs as little as 3 seconds at Inworld and NVIDIA Riva, 5 at Qwen and Smallest AI, and about 10 at Bland Speech, ElevenLabs, Cartesia, and MiniMax, while Grok TTS recommends 90 to 120 seconds.

| Company | Shortest clip | Clip length the company recommends |
|---|---|---|
| Bland Speech | About 10 seconds | One clean sample of about 10 seconds |
| Inworld | 3 seconds | Longer is better up to a 30-second limit |
| ElevenLabs | 10 seconds on Eleven v4 | Says 10 seconds captures a voice with high fidelity |
| Cartesia | 10 seconds | Up to 60 seconds to retain an accent |
| NVIDIA Riva | 3 seconds | About 5 seconds, 3 to 10 allowed |
| Smallest AI | 5 seconds | 5 to 15 seconds |
| Speechify | 10 seconds | 10 to 30 seconds |

<details><summary>What these rows mean</summary>

- **Shortest clip:** the least audio the company says a clone can be made from, as of November 2026. Bland states about 10 seconds, and Inworld states "as little as 3 seconds".
- **Clip length the company recommends:** what each company's own guidance asks for when quality matters. Cartesia says 10 seconds is enough to get started and up to 60 seconds better retains the speaker's accent. NVIDIA allows 3 to 10 seconds and aims for about 5. Grok says clips under 30 seconds may lack detail and recommends 90 to 120 seconds.

</details>

A short recording is also not the same as a free choice of recording. OpenAI's custom voices, which are limited to eligible customers, need at least 5 seconds of actual speech and recommend 10 to 60 seconds, with a 60-second ceiling, and Azure's personal voice takes a 5 to 90 second sample plus a spoken consent statement. Typecast accepts a single file of 5 to 150 seconds for an instant clone, Sarvam recommends a reference clip of 10 to 15 seconds, and Grok TTS takes clips up to 120 seconds.

Bland Speech takes exactly one audio file of up to 10 MB for an instant clone, and the clone is ready in about a minute according to Bland.

## How much audio does a professional voice clone need?

A professional clone needs 10 minutes at the low end, at Inworld and Fish Audio, and 30 minutes at Bland Speech, ElevenLabs, Cartesia, and Mistral, with the ceilings running from 90 minutes to 3 hours.

Bland says a professional clone is fine-tuned on 30 minutes or more of audio and passes Bland's consent and rights review. ElevenLabs recommends approximately 30 minutes of high-quality audio, and Cartesia says 30 minutes is the minimum and more is better. Mistral recommends 30 minutes to 3 hours of clean, high-quality audio from one speaker.

Fish Audio asks for at least 10 minutes of clean audio, up to 180 minutes, and Inworld asks for at least 10 minutes up to 1 GB across as many as 50 samples. Typecast takes one file of 5 minutes to 3 hours. Murf's Enterprise professional cloning takes up to 90 minutes of studio-grade recordings and runs 1 to 4 weeks, managed by Murf.

The ceiling is a ceiling and not a target. Rime says 30 to 60 minutes of clean, consistent recordings makes a basic clone and two to five hours or more makes the strongest one, and IBM sells a custom branded voice from as little as one hour of recordings as a Premium feature built with IBM.

## Does the quality of the audio matter more than the length?

Quality matters as much as length once a clip passes a provider's minimum, because Mistral says a clone copies the emotion and pace of the recording as well as its background noise and the microphone.

Providers publish the same short checklist in different words. NVIDIA asks for clear speech with minimal background noise, trimmed silence at the start and end, a quiet and echo-free room, consistent volume, and no music, sound effects, or overlapping speakers. Inworld says to record in a quiet place, avoid microphone noise, and speak with a variety of emotions so the clone captures the full range of the voice. Mistral says to record the clip in the style a business wants to hear, since a clone inherits the delivery.

Some providers filter for it. Neuphonic rejects a sample it cannot align to a transcript, with an error explaining why, and Qwen says its Plus model copes better with noise and reverberation in reference audio.

## How much audio should a business record for a phone agent?

A business should record about 10 seconds to start a test clone with Bland Speech, and 30 minutes or more of clean, single-speaker audio recorded in the delivery the calls need if a professional voice is the goal.

The 10-second test is cheap and fast: one clean sample, a voice ready in about a minute, and a way to hear whether the speaker's voice survives the trip before anyone books a recording day. The 30-minute set is where a business chooses the microphone, the room, and the delivery once, for every call the voice will take.

Neither route skips consent. Bland has a customer confirm that every voice is theirs to use, and a professional clone passes a consent and rights review before it is trained. Clones are private to the customer's organization. The providers and their checks are compared in [Is voice cloning legal for business?](/blog/is-voice-cloning-legal), and [instant vs professional voice cloning](/blog/instant-vs-professional-voice-cloning) covers which kind a business needs.

Bland Speech is built for phone calls, and its [text to speech API](/text-to-speech-api) returns audio at 8 to 48 kHz and in mu-law, a compressed encoding used on phone lines. More providers are ranked in the [best voice cloning from 10 seconds of audio](/blog/best-voice-cloning-10-seconds) list, and the full set is on the [voice cloning comparison hub](/compare/text-to-speech/voice-cloning) and the wider [text to speech comparison](/compare/text-to-speech). A new Bland Speech account gets 50,000 free characters once, and speech is billed at $0.04 per 1,000 characters, pay as you go, so [Bland Speech](/speech) can be tried before any recording day is booked.

## Frequently asked questions

### Can a voice be cloned from 3 seconds of audio?

Inworld says an instant clone can be made from as little as 3 seconds, and NVIDIA's zero-shot model accepts a 3-second prompt, though NVIDIA aims for about 5 seconds and says prompts shorter than 3 or longer than 10 seconds can produce lower quality. Bland Speech asks for about 10 seconds, and Inworld notes that similarity improves as the sample grows toward its 30-second limit.

### Is 10 seconds enough to clone a voice?

Ten seconds is enough to start at Bland Speech, ElevenLabs, Cartesia, and MiniMax. Cartesia says up to 60 seconds better retains the speaker's accent, and Grok says clips under 30 seconds may lack detail, so 10 seconds is a test and not a final recording.

### What is the minimum audio for a professional voice clone?

The minimum is 10 minutes at Inworld and Fish Audio, 5 minutes at Typecast, and 30 minutes at Bland Speech, ElevenLabs, Cartesia, and Mistral. Rime says 30 to 60 minutes makes a basic enterprise clone.

### Does the voice sample need to be in the language the clone will speak?

Not always. Sarvam says the reference clip's language does not have to match the output, and Speechify says a voice keeps the accent of the recording it was made from. Descript says its clones are generated from a model based on US English pronunciation, so accents may not be preserved accurately.

### What file types can a voice sample be?

Providers accept different files. Bland Speech takes one audio file of up to 10 MB, Smallest AI accepts mp3, wav, mp4, and webm under 5 MB, Inworld accepts wav, mp3, and webm up to 12 MB, and NVIDIA wants a 16-bit mono WAV at 22.05 kHz or higher. A recording's conditions still carry into the clone, so a clean source beats a convenient one.

### Does more audio always make a better clone?

More audio helps up to a point and then stops mattering. Inworld says similarity improves up to its 30-second instant limit, and it trims longer samples. Professional clones are built for far more audio, and Rime says its strongest results take two to five hours or more.
