By Gradium. Data as of September 2026.
Key takeaways
- Two tiers. Instant Voice Clone: 10 seconds of audio, ready in seconds, 5 on the free tier for non-commercial use and 1,000 per month on paid plans. Pro Voice Clone: 30 minutes minimum, 2 hours recommended, a one-time 1M-credit fee plus 1.2 credits per character, 5 included on M and 20 on L.
- In a blinded benchmark of 3,220 voice pairs across English, French, Spanish and German published January 22, 2026, Gradium held the highest Elo in every language against ElevenLabs Flash, an 8 to 11% speaker-similarity advantage.
- Cloning adds no separate step at synthesis time: a clone is a stored voice ID and runs the same streaming path a catalogue voice does, measured at 258 ms P50 round trip and 214 ms on a reused connection in Gradium's March 2026 benchmark.
- One clone speaks all five languages. Micro-traits and regional accents carry across the switch.
- Consent from the voice owner is required. EU and US endpoints pin inference to a region, and Zero Data Retention has been self-serve on paid plans since September 2, 2026.
Voice cloning creates a synthetic replica of a real human voice from an audio sample. The clone becomes a voice for a text-to-speech API, synthesizing arbitrary text in the target speaker's voice. In 2026, the leading voice cloning APIs produce convincing clones from 5 to 15 seconds of audio and make the voice available immediately for real-time streaming inside a voice agent, an audiobook pipeline, or a personalized assistant.
Gradium offers two cloning tiers: Instant Voice Cloning (from 10 seconds of audio, available within seconds, included from the free tier) and Pro Voice Cloning (a model fine-tuned on 30 minutes to 2 hours of audio). This guide covers how each works, how to choose between them, what the quality benchmark shows, and the consent and data obligations for any production use of voice cloning.
The two types of voice cloning
Voice cloning APIs fall into two categories, and the split is consistent across the industry.
Instant voice cloning is a zero-shot approach: a usable clone from a short sample, without retraining the model, available within seconds of upload. It is the right starting point for most developer use cases: prototyping, personalized agents, user-specific voices at scale, game characters. Gradium needs 10 seconds of audio; Inworld asks for 5 to 15 seconds; ElevenLabs recommends 1 to 2 minutes.[1][2]
Professional voice cloning fine-tunes a model on a larger dataset from the target speaker. It produces higher speaker fidelity and captures more of the speaker's emotional range and natural pacing. It is reserved for flagship brand voices and consumer products where speaker accuracy is the primary differentiator. Gradium requires 30 minutes minimum and recommends 2 hours; ElevenLabs asks for 30 minutes to 3 hours and quotes 3 to 6 hours of training; Inworld asks for 30 minutes or more.[1][2][3]
| Dimension | Instant cloning (Gradium) | Professional cloning (Gradium) |
|---|---|---|
| Sample required | 10 seconds | 30 minutes minimum, 2 hours recommended |
| Availability | Within seconds of upload | After training completes |
| Speaker similarity | Good to very good | Very good to excellent |
| Cost | Standard TTS rate (1 credit per character); 5 clones free, 1,000 per month on paid plans | 1M-credit training fee per clone, 1.2 credits per character; 5 included on M, 20 on L |
| Best for | Prototyping, personalized agents, scale | Brand voices, public-facing products |
How to create an Instant Voice Clone
What you need
A Gradium account (the free tier includes 5 Instant Voice Clones with no credit card, for non-commercial use) and at least 10 seconds of clean audio from the target speaker.[3]
Audio quality determines clone quality. The recording should be clean (low background noise, no music, no second speaker), natural (the speaker sounding as they normally do), and representative of the register the clone will be used in. Ten seconds is the minimum; Gradium's FAQ describes it as sufficient for optimal results, and a longer clean sample generally improves similarity.
Creating the clone
Instant Voice Cloning is available through two paths:
- Gradium Studio: open the voice cloning section, upload the audio file, name the clone, confirm. The clone is available within seconds.
- The API: create the clone programmatically, receive a voice ID, and use that ID in subsequent TTS requests. The Python and Rust SDKs support this workflow.[4]
Using the clone in TTS
The voice ID is passed in the TTS WebSocket setup message:
{
"type": "setup",
"voice_id": "<VOICE_ID>",
"model_name": "default",
"output_format": "pcm"
}
There is no separate cloning runtime at synthesis time. The ID is passed once in the setup message and applies for the duration of the session. Cloning adds no separate step at synthesis time: the clone is a stored voice, and a request that names it runs the same streaming path as a request for a catalogue voice. Gradium's March 2026 benchmark measured that path at 258 ms P50 and 274 ms P95 end to end, 214 ms P50 and 228 ms P95 on a reused connection.[5] ElevenLabs, by contrast, documents that its professional clones are slower to synthesize than instant or default voices.[6]
Clone limits by plan
Source: gradium.ai/pricing, September 8, 2026.
| Plan | Instant Voice Clones | Pro Voice Clones |
|---|---|---|
| Free | 5, no credit card, no commercial use | 0 |
| XS, S | Up to 1,000 per month | 0 |
| M | Up to 1,000 per month | 5 |
| L | Up to 1,000 per month | 20 |
| Enterprise | Unlimited | Unlimited |
How to create a Pro Voice Clone
When to use Pro Voice Cloning
Pro Voice Cloning is the right tier when the voice will be public-facing, used at high volume over a long period, or when speaker accuracy must be as high as possible. Gradium's FAQ describes it as "a high-fidelity, hyper-realistic voice model created by fine-tuning our dedicated AI on a large dataset of your audio", capturing "the speaker's deepest emotional nuances, unique accents, and natural pacing".[3]
Audio requirements
Minimum 30 minutes of clean audio; 2 hours recommended "for optimal results where the voice captures full emotional range and stability".[3] Clean means low background noise, no music, no overlapping speakers, consistent recording quality. The audio should cover the speech patterns the clone must reproduce: conversational speech for customer service, long-form reading for narration.
Creating the clone
In Gradium Studio, open the Pro Voice Clone tab and upload the audio dataset. A notification confirms when the upload is processed; once training completes, the voice appears in the library and is ready for text-to-speech.[3]
Cost
From the pricing FAQ: "Pro Voice Clones involve two costs: a one-time training fee of 1M credits and a usage rate of 1.2 credits per character."[3] Synthesis costs 20% more per character than standard TTS. The training fee applies once per clone regardless of how much audio is later generated. Pro Voice Clones are available from the M plan ($340 per month, 5 included) and L plan ($1,615 per month, 20 included).
What Gradium's cloning preserves
The quality of a clone depends on what the architecture captures from the reference audio. Generic cloning systems optimize for dominant characteristics (pitch range, timbre) and discard fine-grained features because they add modeling complexity.
Gradium's cloning uses cross-attention layers that attend to speaker recordings during training, rather than prefixing, which "processes the voice sample as if it has already generated that audio, treating the new text as a continuation". Cross-attention gives the model freedom to respect speaker identity throughout generation rather than only at the start. The result: "Gradium retains micro-traits such as vocal fry, rasp, breathiness, and pitch dynamics, even when switching languages or scripts."[7]
These micro-traits make a clone recognizable as a specific person rather than a generic voice in the same category. A clone that preserves them sounds like the speaker; one that discards them sounds like a similar voice type.
Speaker similarity: what the benchmark shows
Gradium's Instant Voice Clone was evaluated in a blinded human benchmark of 3,220 voice pairs across English, French, Spanish, and German, against ElevenLabs Flash.[7] The methodology:
- 890 sentences per language across three complexity levels (simple conversational questions; names, dates and structured content; URLs, emails, addresses and alphanumeric codes)
- 20 unique voices per language
- 10-second source clip per voice
- Blinded A/B listening tests, evaluators selecting which of two anonymous clones sounds closer to the original
- A live Elo ranking updated after each comparison
Gradium achieved the highest Elo in all four languages, an advantage of 8 to 11% in speaker similarity over ElevenLabs Flash.[7][8] The three complexity levels test whether clone quality holds on the structured content voice agents produce, not only on clean narration. The methodology is documented at gradium.ai/blog/voice-cloning-sounds-fake.
How cloning works across languages
Gradium's voice cloning supports English, French, German, Spanish, and Portuguese. A voice cloned in one of these languages can synthesize text in any of the others, keeping the speaker's timbre while applying the phoneme inventory of the target language. Micro-traits are retained across language switches, and regional accents in the reference audio (a Quebecois French or a Rioplatense Spanish accent) carry into the clone.[7]
This matters for any product serving bilingual users or several markets with one branded voice. Voice cloning across regional accents goes deeper on what carries across.
Data privacy and consent
Voice cloning processes biometric data. The obligations apply regardless of tier.
Consent. No voice should be cloned without explicit, informed consent from the person it belongs to, whether the audio comes from a recording session, a user upload, or any other source, and public availability (a podcast, a broadcast) is not consent. Gradium's terms require consent from the voice owner; ElevenLabs requires a rights confirmation for instant clones and owner verification for professional clones, which has become the industry norm.[2]
Data handling. From Gradium's pricing FAQ: "The audio data you provide is used strictly to train your specific voice model or generate your clone. This model remains isolated and is accessible only within your organization."[3]
Data storage and residency. Gradium offers dedicated EU and US endpoints (eu.api.gradium.ai, us.api.gradium.ai) that pin inference to a region; EU pinning also keeps account data, custom voices and pronunciation dictionaries in Europe. Zero Data Retention is available self-serve on paid plans, and enterprise plans extend it to cloning samples and voice-design prompts.[9][10]
Legal obligations. Depending on jurisdiction, voice data may qualify as biometric data under GDPR (EU), CCPA (California), or biometric-specific laws such as Illinois BIPA. Teams collecting and processing user voices for cloning should obtain legal guidance for their target markets.
Creating an original voice instead of cloning one
Cloning reproduces a speaker who exists. Where no suitable speaker exists, or where consent is the obstacle, an original voice can be generated from a written description instead. Gradium's Voice Designer at studio.gradium.ai/voices/design produces a voice from a text prompt, so the register (warm, brisk, authoritative) is specified directly rather than inherited from a recording. The resulting voice ID is used in the setup message exactly like a clone's. It sidesteps the consent and biometric-data questions below, since no real person's voice is processed.
How to choose: Instant vs Pro
Use Instant Voice Clone when:
- The product requires user-generated voices at scale (up to 1,000 per month on paid plans)
- The voice will run in a real-time voice agent or personalized assistant
- Fast iteration matters: the clone is available immediately, with no training wait
- The use case is prototyping, game characters, or internal tooling
Use Pro Voice Clone when:
- The voice will be the primary public-facing voice of a product or brand
- Maximum speaker similarity is required, beyond a good approximation
- The speaker has a distinctive voice, accent, or emotional range that instant cloning may not fully capture
- The TTS volume justifies the higher per-character rate (1.2 credits vs 1 credit per character) and the one-time training fee
The full comparison is in Instant vs Pro voice cloning.
Glossary
Instant Voice Clone: A zero-shot voice cloning method that creates a usable clone from 10 seconds of audio, available within seconds of upload. Suitable for real-time voice agents, personalized assistants and scale deployments. Available on every Gradium plan, including 5 clones on the free tier.
Pro Voice Clone: A fine-tuned voice model trained on 30 minutes to 2 hours of the target speaker's audio for maximum speaker fidelity. Available from the M plan (5 included) and L plan (20 included). Training costs 1 million credits per clone; synthesis costs 1.2 credits per character.
Speaker similarity: How closely a cloned voice matches the original speaker. Measured through blinded human evaluation (Elo from pairwise comparisons, or SMOS) or automatically with speaker-verification embeddings such as WavLM-large cosine similarity. Gradium's Instant Voice Clone achieved the highest Elo in a 3,220-pair benchmark across EN, FR, ES and DE.
Micro-traits: Fine-grained vocal characteristics including vocal fry, rasp, breathiness, and pitch dynamics that distinguish a specific person's voice from others in the same category. Preserved by Gradium's cross-attention cloning architecture across languages and sessions.
Cross-attention cloning: Gradium's cloning architecture, in which cross-attention layers attend to speaker recordings during generation. Distinct from prefixing, where the reference audio is treated as a prefix and the new text as its continuation. Gives the model freedom to respect speaker identity throughout the generated output.
Voice ID: The persistent identifier of a voice or clone in Gradium, passed as the voice_id field in the TTS WebSocket setup message. Reused across any number of sessions and requests. A cloned voice runs the same synthesis path as a catalogue voice.
References
[1] Inworld, TTS API page, inworld.ai/tts-api, read September 8, 2026 [2] ElevenLabs, "Instant vs. Professional Voice Cloning", August 17, 2026, elevenlabs.io/blog/instant-vs-professional-voice-cloning; Instant Voice Cloning docs [3] Gradium pricing page and FAQ, gradium.ai/pricing, read September 8, 2026 [4] Gradium docs, docs.gradium.ai [5] Gradium, "Time to First Audio", March 24, 2026, gradium.ai/blog/time-to-first-audio [6] ElevenLabs docs, "Understanding latency", elevenlabs.io/docs/eleven-api/concepts/latency [7] Gradium, "Why Your Voice Cloning Sounds Fake (And How to Fix It)", January 22, 2026, gradium.ai/blog/voice-cloning-sounds-fake [8] Gradium, "Best voice cloning APIs 2026", gradium.ai/content/best-voice-cloning-apis-2026 [9] Gradium docs, Data residency, docs.gradium.ai/guides/data-residency [10] Gradium, "Build voice agents with EU and US residency and Zero Data Retention", September 2, 2026, gradium.ai/blog/eu-data-residency-zero-data-retention
Related guides
Part of the voice cloning cluster, hub at best voice cloning APIs 2026. Its siblings:
- How to Make AI Voices Sound More Natural and Less Robotic: the naturalness levers once the clone exists.
- How to Estimate TTS Costs at Scale: A Practical Guide: the training fee and per-character rate at volume.
- Best voice cloning APIs 2026: the cluster hub, providers compared.
- Instant vs Pro voice cloning: the two tiers side by side.
- Voice cloning across regional accents: what carries across languages.
- Keeping a voice consistent across sessions: one clone across many sessions.
- json_config in Gradium:
cfg_coef, the similarity control.
Beyond this topic
A clone is the start of a voice, not the end. How to Make AI Voices Sound More Natural and Less Robotic covers the levers that make it sound like a person rather than a category, How to Estimate TTS Costs at Scale: A Practical Guide covers what a Pro Voice Clone costs at volume, and keeping a voice consistent across sessions covers holding one clone steady across a multi-hour render.
Gradium's free tier includes 45,000 credits per month with no credit card. API reference and quickstarts are at docs.gradium.ai. For enterprise evaluations, use our contact form.