Voice clone recording quality checklist
The quality of an AI voice clone is decided before you ever press “train” — it is set by the recording. Models faithfully reproduce whatever is in your samples, including the bad parts: hiss, room echo, inconsistent levels, and a monotone delivery all carry through to every future line. This checklist walks your setup against the things that actually move clone quality so you catch problems while they are still fixable.
How it works
Choose your target tool — ElevenLabs, Resemble, or PlayHT — and the checklist sets a sensible minimum duration target for that platform’s high-fidelity mode. Then tick off each requirement: a quiet noise floor, correct microphone distance, treated room acoustics, sample diversity across pace and emotion, and total recorded duration. The tool counts how many critical items pass and warns about any that will degrade the result, giving you a clear go / fix-first signal.
What each check actually tests
Noise floor
Record ten seconds of room silence and check the waveform is genuinely flat in your DAW or audio editor. Any visible hum, hiss, or HVAC rumble will be baked into the model’s “silence” and appear between every generated word. A noise floor below –60 dBFS is a safe target; below –70 dBFS is excellent.
Microphone distance and plosives
At around 15–20 cm with a pop filter, you get a full, consistent tone without proximity bass boost or plosive blasts. Too close and p/b/t sounds crack; too far and room reflections dominate. Stay fixed at one distance throughout — the model learns the combined distance-room signature.
Room acoustics
Hard surfaces (bare walls, glass, tile) add audible reverb that the model picks up as a consistent “tail” on every generation. Soft furnishings, carpets, heavy curtains, or a dedicated acoustic panel behind the speaker absorb those reflections. A walk-in wardrobe packed with clothes is a genuinely excellent quick studio.
Sample diversity
A clone trained only on slow, calm reading sounds flat when asked to be animated, or clips when asked to project. Record varied pace (slow-deliberate, normal, fast), varied pitch range (low emphasis, normal, high emphasis), and varied emotion (neutral, warm, urgent, questioning). Even 10 minutes of diverse speech trains more expressive range than 30 minutes of monotone delivery.
Duration targets by platform
| Platform | Instant clone | High-fidelity clone |
|---|---|---|
| ElevenLabs | 1 minute minimum | 30+ minutes recommended |
| Resemble AI | 5 minutes minimum | 45+ minutes for best results |
| PlayHT | 2 minutes minimum | 20+ minutes recommended |
Instant clones work from short samples but give narrower range and less speaker-identity accuracy. High-fidelity clones need more data to nail subtle prosody.
Tips for a clean capture
- Fix the room before the mic. Recording in a closet of clothes or under a blanket fort kills reflections more cheaply than any plugin.
- Set the noise floor first. Record ten seconds of silence and check it is inaudible — any audible hum will haunt every generation.
- Read with range. Capture calm, excited, questioning, and emphatic lines so the clone can perform, not just narrate.
- Keep levels consistent. Stay the same distance from the mic throughout; varying distance teaches the model an unstable voice.
- Avoid clipping. Peaks above –3 dBFS introduce digital distortion that cannot be removed cleanly; aim for peaks around –6 to –12 dBFS.
- Re-record, do not edit around bad takes. Cutting out noise events leaves short sections of different acoustic character that confuse the model.