ElevenLabs voice cloning: How it works, what it costs and how to use the API

Plans, prices and features checked on the official pages on October 21, 2026.

ElevenLabs can clone a voice from a short recording, but not on the free plan. According to its pricing page, Instant Voice Cloning starts on the $6 Starter plan, and Professional Voice Cloning starts on the $22 Creator plan.

Instant clones need about one to two minutes of clean audio and are ready almost at once. Professional clones are trained on a longer recording and take hours. Below: the plans, credits, licensing rules, the API and the other tools in the platform.

What ElevenLabs does and which products it covers

ElevenLabs describes itself as an AI research and product company. Its documentation lists three main product areas: ElevenCreative, a no-code web app for voiceovers, music, dubbing and studio projects; ElevenAgents, for building voice agents; and ElevenAPI, with REST endpoints and official Python and TypeScript SDKs. It also lists Reception.ai, an AI phone receptionist for small and medium businesses.

The company’s about page says it is backed by investors including Andreessen Horowitz, ICONIQ, Sequoia and NEA, and carries a September 30, 2026 headline reporting a $22 billion valuation. The page does not state a founding date, founders or headquarters. Wikipedia says the company was founded in 2022 by Piotr Dabkowski and Mateusz Staniszewski, though that is third-party context.

For a broader assessment of the product, see the aiholics ElevenLabs review. For the company’s newest agent technology, see ElevenLabs bets on one-model voice agents.

Instant versus professional cloning: What each needs and delivers

The documentation separates two kinds of clone.

• Instant Voice Cloning (IVC) does not train a custom model. It needs about 1 to 2 minutes of good audio. The docs say audio beyond 2 to 3 minutes adds little and can hurt stability, and that audio quality matters more than the number of files.
• Professional Voice Cloning (PVC) trains a dedicated model on your voice. It needs roughly 30 to 180 minutes of clean, consistent recordings without background noise, reverb or several speakers. Training usually takes 3 to 6 hours, though the docs say it can take longer depending on the queue, and one FAQ says up to 24 hours.

ElevenLabs recommends PVC for very unique voices or uncommon accents that an instant clone may not capture well.

Consent is built in. Each professional clone must pass a verification step confirming the voice is yours. The docs say you can create a PVC only of your own voice; even with someone’s consent you cannot clone theirs this way. To share a voice, the owner creates and verifies it in their own account and shares it by link.

Plans and prices: What each tier includes

Prices checked on October 21, 2026. Source: the ElevenLabs pricing page.

Scale includes 3 seats and Business includes 10. Enterprise adds custom terms, SLA and DPA assurances, HIPAA BAAs, SSO and volume discounts. The Creator and higher plans add 192 kbps audio, and Pro adds 44.1 kHz PCM output through the API.

How credits work, when they reset and how to get more

Credits were previously called characters, and the value is unchanged. For text to speech, the docs say one input character costs one credit. Other operations are billed differently: the voice isolator and voice changer cost 1,000 credits per minute of audio, and sound effects with a set duration cost 40 credits per second.

Self-serve plans reset monthly, on your billing date. On paid monthly and annual plans, up to two months of unused credits roll over, so the balance is capped at three times the monthly quota. Downgrading or cancelling forfeits them, and rollover does not apply to the Free plan.

The documentation reviewed for this guide describes no daily allowance; the free plan’s 10,000 credits are a monthly amount. To get more credits, upgrade or buy Pay As You Go top-ups. Top-up credits last 12 months and are not available on Enterprise or legacy subscriptions. Usage-based billing for overage is a legacy feature that new self-serve subscriptions do not get.

To cancel, use the billing settings. Cancellation takes effect at the end of the billing cycle, unused credits are lost, and the account moves to the free plan. You can resume before the cycle ends. Content is not deleted, but cloning becomes unavailable.

Licensing: When the audio is free to use commercially

The pricing page lists a commercial license for every paid plan and none for Free. ElevenLabs’ help center says audio generated during a paid subscription keeps its commercial rights after the subscription ends, with no expiration. It also says audio generated outside a subscription always requires attribution, and that there is no guarantee content stays accessible on the free tier.

Eleven Music is described in the docs as cleared for nearly all commercial uses, including film, TV, podcasts, social media, advertising and gaming, with details by plan in the music terms. Free-tier dubs are watermarked automatically; paid ones are not.

Getting started and using the API, step by step

1. Create an account. New accounts land on the free plan.
2. Choose a voice from the library, or upgrade to Starter or higher and add an instant clone from a clean recording of one to two minutes.
3. Pick a model. The docs list Eleven v3 (70+ languages, 5,000 characters per request), Flash v2.5 (32 languages, about 75 ms model latency, 40,000 characters, 50% lower API cost per character), Multilingual v2 (29 languages, 10,000 characters) and the newer v4 family (90+ languages).
4. Generate and download. MP3 is the default; PCM, Opus, μ-law and A-law are also offered, and higher-quality options require a paid tier.
5. For code, create an API key in your account settings under API keys, store it as ELEVENLABS_API_KEY and send it in the xi-api-key header.

The docs say you can restrict a key by endpoint scope, credit quota and IP allowlist, and warn against exposing it in client-side code. SDKs install with pip install elevenlabs or npm install @elevenlabs/elevenlabs-js. A text to speech call takes the text, a voice_id, a model_id and an output_format. Business plans list low-latency text to speech from about 5 cents per minute.

Controlling emotion and pauses in generated speech

According to the docs, the models read emotion from the text itself, such as dialogue tags or punctuation, and the Stability and Similarity settings affect consistency. Guidance written into the text may be spoken aloud, so you may need to trim it afterward.

For pauses, the documentation says to use <break time="1.5s" />, up to 3 seconds, but not with Eleven v3 or v4, where punctuation, ellipses and text structure are the substitutes. Too many break tags in one generation can cause instability. Eleven v3 and v4 support audio tags such as [whispers], [sighs] and [laughs]. For pronunciation, v4 accepts IPA between slashes, while <phoneme> tags work only with eleven_flash_v2; other models can use alias tags in pronunciation dictionaries.

Beyond text to speech: Music, sound effects, dubbing and more

• Music: yes, ElevenLabs can generate music. Tracks run from 3 seconds to 5 minutes, with Music v2.5 the most advanced model. The Music API is for paid subscribers. An earlier aiholics piece covers how the company approaches music and voice.
• Sound effects: text prompts produce effects up to 30 seconds, with looping options, as MP3 or 48 kHz WAV for non-looping effects.
• Voice isolator: removes background noise from audio or video, up to 500 MB and one hour per file. The docs say it is not specifically optimized for separating vocals from music.
• Voice changer: converts a recording into another voice while keeping whispers, laughs and delivery. Segments run up to 5 minutes.
• Dubbing: translates audio and video into 90+ languages with Dubbing v2, keeping each speaker’s voice, up to 32 speakers per file. Live dubbing is not available.
• Speech to text: the Scribe v2 and Scribe v2 Realtime models.

Alternatives and what to check before switching

Competitors in text to speech and voice cloning include the speech services from large cloud providers and OpenAI, as well as specialist voice startups. This guide did not read their pricing pages, so it gives no figures for them. When comparing, check the same four points used above: whether cloning is on the free tier, the commercial license terms, the per-character or per-minute rate, and the number of supported languages.

Common questions