Fish Audio instant voice clone best practices (and ElevenLabs IVC)

Instant clones are not the same job as a production audiobook sample. A 90-second phone take can sound like you on a demo line; a thin, noisy take will also sound like you: thin and noisy, forever. This is what we actually do when authors clone on Fish Audio through Lex, with a short ElevenLabs instant-clone (IVC) comparison when that is the tool.

Key numbers (Fish, Lex default path): hard floor 10 seconds of usable speech, target ~30 seconds, better at ~90 seconds, cap around 2 minutes for this lane. Room noise and clipping kill more clones than "not enough minutes." For a full-book production sample (hours of clean narration), use our longer brief: how to record a voice sample under 2 hours.

What is an "instant" voice clone?

An instant clone builds a usable voice from a short reference instead of a long training set. On Fish you can either upload samples into a saved voice, or pass reference audio (plus transcript) on a single generation. On ElevenLabs, Instant Voice Cloning is a short-sample product separate from Professional Voice Cloning (tens of minutes to hours). At Lex, author self-narrate intake for quick demos lands on Fish; longer partner / author production samples use a different, stricter pipeline.

Instant is for "hear yourself on a page of your book now." Long samples are for "ship a full title without the clone collapsing at hour six." Do not mix those quality bars.

How long should the Fish sample be?

Fish Audio's voice cloning docs set the floor at about 10 seconds per clip, and Fish says a minute or two of clear, single-speaker speech improves fidelity. Its voice cloning best practices suggest two or three clips of 15–20 seconds that form a complete paragraph, and 30–60 seconds if the result sounds robotic. Below ~10s the clone often lacks prosody; past a few clean minutes of instant material you usually stop buying fidelity and start buying inconsistency.

What we gate in product: 10s minimum, 30s target, 90s better, about 120s max for the self-narrate flow. If you only have one good take, one continuous read of the guided script beats stitching three bad phone clips.

Twenty seconds is often enough. We published four twenty-second author clones, each built from one clip of 20 to 25 seconds with one voice and no music. Levels on those clips ranged from −29 to −18 dB RMS, and every one cloned. Noise and music broke clones. Level mattered least.

What makes Fish clones sound like you?

Signal quality comes first. Transcript and performance consistency come after it, and they only help if the signal is clean.

• One speaker, dry room. No music, HVAC roar, café wash, or bathroom slap. The model treats background as part of the voice.
• Distance and level hold still. Two fists from the mic, same seat, same energy. Riding the chair in and out becomes volume and timbre drift in every later take.
• Peaks under about −3 dB. Do not slam the sample into 0 dBFS. Clipped source reads as metallic or crunchy later. Loud enough to be clearly above noise is enough.
• Exact transcript when you can send one. Fish accepts a transcript alongside the reference audio and says it sharpens pronunciation, so send matching text, punctuation included. Lex's guided scripts exist so the words in the file match the words we send. Hand-edited freestyle is fine if you paste a true transcript; wrong words hurt prosody more than people expect.
• Natural pace, not auctioneer, not ASMR. Read the way you want the book to sound. A forced "announcer" voice clones as an announcer.
• Variety inside one register. A question and a slightly warmer line, still the same person. Do not half-whisper and half-shout inside one clip.

Format is secondary. Clean mono speech at a normal speech sample rate beats a noisy 48 kHz stereo export. WAV or high-bitrate M4A/MP3 are both fine for this lane; compression artifacts only matter when the room already sounds cheap.

Fish can enhance noisy samples before training. Useful for a slightly hissy bedroom take. If you already have a dry, studio-clean capture, skip heavy "enhance / denoise" so the model does not learn the processor.

What should you actually read?

On Lex, use the on-screen guided script unless you have a reason not to. Those lines are written for continuous speech at a calm narration pace and we attach them as the Fish texts so alignment stays honest. Reading a page of your own book works if your mic take matches a transcript you can provide word-for-word.

Skip poems unless the book is poetry. Skip songs and co-reads. Skip "I am testing, one two three" into the only file you upload. The clone will sound like a mic check.

How is ElevenLabs IVC different?

ElevenLabs Instant Voice Cloning wants slightly more speech than Fish's shortest path, and it is even more allergic to inconsistency.

ElevenLabs' documentation draws the line clearly. An Instant Voice Clone works from under 2 minutes of audio and is ready right away. A Professional Voice Clone trains a dedicated model on 30–180 minutes of good audio, and fine-tuning usually takes 3–6 hours.

1. In the ElevenLabs dashboard, open Voices and click the plus icon.
2. Choose Instant Voice Clone.
3. Upload or record your sample.
4. Name the voice, confirm you have the right and consent to clone it, and click Save voice.
5. Find it under My Voices and click Use voice.

• About 1–2 minutes of clean audio is the happy band. Past roughly 3 minutes of IVC material, gains flatten and bad takes can make the clone worse.
• Consistency over mileage. One excellent minute beats five mixed sources. Same mic, same room, same delivery, same distance.
• Aim average speech around −23 to −18 dB RMS, true peak under about −3 dB. That matches what we already tell authors for longer samples.
• ElevenLabs' documentation advises MP3 at 128 kbps and above, and its file FAQ now recommends 192 kbps or higher. WAV rarely improves the clone and can cause upload problems. How you recorded matters more than the codec.

If you are logging hours for a professional / long-form ElevenLabs clone, that is a different brief (think tens of minutes minimum, often hours). That job belongs in the production sample guide.

What ruins an instant clone in practice?

• Noise as identity. Fridge, fan, traffic, keyboard. The clone will bring them into every sentence.
• Room reverb. Wet bathrooms and hard kitchens bake space into the speaker.
• Auto-gain and aggressive "AI denoise." Phones and consumer apps that pump or chirp train the model on the artifact.
• Multiple mics or days mashed together. Two different rooms read as two people.
• Acting extremes in a short sample. Instant clones have little average to fall back on; a cartoon villain line in a 20s clip can dominate the whole voice.
• Lying transcript (Fish). Mismatch between spoken audio and submitted text blurs rhythm and stress.

Quick checklist before you upload

1. Quiet room (closet full of clothes is fine).
2. One mic or phone, fixed distance, no headset processing if you can avoid it.
3. At least 10s of real speech; 30–90s if you want a fair self-narrate demo.
4. Natural narration energy; no music, no second speaker, no clipping.
5. Transcript matches the take (use Lex's script when you are in our flow).
6. Listen once before upload: if you hear noise or slap, re-record. Do not "fix it in software" unless you know what you are doing.

Want more than a demo line? Start with a free first-chapter sample on Lex Audio, or see Lex production pack pricing. For full-length author clones, record the longer clean sample. A single 30-second take stretched across a novel will show its seams.

FAQ

Can you clone someone else's voice with an instant clone?

Only with permission. Fish and ElevenLabs require it. Since 2 August 2026, the European Commission says EU AI Act Article 50 makes professional deployers disclose AI-generated deepfakes (not legal advice). On Lex, a cloned voice is a private model used only for its book, and Lex's voice cloning FAQ says it is not shared, sold, or reused for other clients.

Can an instant voice clone speak another language?

Yes, within limits. ElevenLabs' documentation says you can clone your voice in any language ElevenLabs supports for instant clones, and results can be unpredictable outside them. On Lex, Alan Berg's clone, built from twenty-one seconds of podcast audio, read his Spanish title, Ghosting, ¿por qué me ignoran?

Why does my clone's accent sound wrong?

Instant clones can miss uncommon accents, and ElevenLabs' documentation says you cannot change a clone's accent after creation. It adds that small changes to the sample can make a large difference. Re-record or re-trim it with one voice in your natural narration register. If the accent still drifts, move to a longer professional clone.

What file format should I upload for an instant clone?

For ElevenLabs, upload MP3 at 192 kbps or higher; its file FAQ says WAV usually does not improve quality and can cause upload problems. For Fish through Lex, WAV or high-bitrate M4A/MP3 are both fine. Clean mono speech beats a noisy stereo export, and recording quality matters more than format.