How to Clean Up a Voice Recording for Voice Cloning (ElevenLabs & More)
Why generative speech enhancers flatten voice clones, and how to strip background hiss, room echo, and dead air while keeping the speaker's natural timbre.
Updated October 26, 2026Michael NguyenTested October 2026
Quick answer
Voice cloning models like ElevenLabs copy everything in your training audio—including mic hiss, fan hum, and room echo—yet ElevenLabs' own documentation warns against heavy generative speech enhancers because rebuilding the voice flattens the speaker's natural timbre and prosody. To prepare a clean voice sample without changing how the person actually sounds, run Audo AI's non-generative Remove noise (/noise-removal, first minute free with no sign-up, then $1.80 an hour) to filter out background noise while leaving the real voice untouched, use Remove echo (/remove-echo) on a subtle setting only if the room rings, and trim dead air with Remove silence (/silence-remover, $0.60 an hour).
Open Remove noise| Voice clone requirement | Why it matters for the clone | What to run on Audo AI |
|---|---|---|
| Zero background hiss, hum, or music | The cloning model treats background noise as part of the voice and bakes hiss into generated speech | Remove noise (/noise-removal) or Voice isolator (/voice-isolator)—filters noise without regenerating the voice |
| Natural vocal timbre (no generative resynthesis) | Generative speech enhancers smooth out breathiness, pitch micro-shifts, and rasp, making clones sound generic | Use Remove noise (/noise-removal) instead of full-strength Audio enhancer so the real voice stays intact |
| Dry acoustics with no hollow room echo | Room reflections make the cloned voice sound boxy or distant on every line it speaks | If the room echoes, run Remove echo (/remove-echo) set toward Keep some of the room (Subtle/Medium) |
| One speaker only, tight pauses, no clipping | Long dead air wastes sample length, and a second voice or clipped peaks corrupts the voice profile | Cut other speakers with Cut audio (/audio-cutter), tighten pauses with Remove silence (/silence-remover), and verify 0 clipped samples on the free LUFS meter (/lufs-meter) |
Hear the difference
Play Original, then Cleaned.
Cleaned removes the chatter and dishes.
Made with Remove noise
Why generative speech enhancers ruin voice clones—and why non-generative noise removal works
Anyone preparing audio for ElevenLabs Instant Voice Cloning or Professional Voice Cloning runs into an apparent contradiction: the documentation says your sample must be free of background noise and room reverb, yet also warns against heavy noise-reduction or studio-enhancement filters.
The reason comes down to how two very different kinds of AI audio clean-up work:
- Generative speech enhancers (avoid at full strength for cloning). Tools like Adobe Podcast Enhance Speech, Descript Studio Sound, and Audo's Audio enhancer at 100% rebuild the voice from scratch. While that rescues a terrible phone call for a podcast listener, a voice cloner trained on resynthesized speech clones the enhancer's synthetic voice model rather than your unique vocal cords.
- Non-generative noise removal (safe for voice cloning). Audo AI's Remove noise (/noise-removal) does not resynthesize or invent speech. It separates and turns down steady hiss, air-conditioner hum, computer fans, and background music while keeping the speaker's real formants, breathiness, and delivery exactly as recorded.
Step-by-step workflow to prepare a voice sample on Audo AI
Whether you are preparing 1 to 2 minutes for an instant clone or 30 minutes to 3 hours for a professional voice clone, run only the steps your recording actually needs:
- Step 1: Strip background noise or music with Remove noise (/noise-removal). Drop your WAV, MP3, M4A, or video file into Remove noise. If the sample was pulled from a YouTube video or podcast intro with a music bed underneath, Remove noise (or /voice-isolator) strips the music and leaves the speaking voice.
- Step 2: Tame room ring gently with Remove echo (/remove-echo) only if needed. If your recording already sounds close and dry, skip this step. If you recorded in an untreated bedroom or office with noticeable slapback echo, run Remove echo with the slider pulled back toward Keep some of the room so the reverb tail disappears without making the voice sound artificial.
- Step 3: Cut any other voices and shorten long pauses. A training sample must contain only one voice. Use Cut audio (/audio-cutter) or the browser audio editor (/audio-editor) to trim out hosts or interviewers, then run Remove silence (/silence-remover) to shorten pauses over 0.7 seconds while keeping a natural 0.25-second breath between sentences.
- Step 4: Check noise floor and clipping on the free LUFS meter (/lufs-meter). Drop the finished file into Audo's free LUFS meter (/lufs-meter). Confirm that Clipped samples reads 0, true peak sits at or below −1 dBTP, and the noise floor is below −60 dB.
How much audio to export, and in what format
ElevenLabs recommends 1 to 2 minutes of clean speech for Instant Voice Cloning (adding more than 3 minutes to an instant clone can actually reduce quality), and at least 30 minutes—up to 2 to 3 hours—of consistent recordings for Professional Voice Cloning.
Export as uncompressed WAV or high-bitrate MP3 (192 kbps or higher). If your source is an iPhone video or MKV recording, Audo's Extract audio from video (/extract-audio-from-video) or Audio converter (/audio-converter) converts it to WAV or MP3 in seconds.
When to upload your raw recording without any processing
The best voice clone sample is one recorded cleanly enough that it needs no noise removal at all:
- Upload the raw file untouched. If you recorded in a quiet room or clothes-filled closet, spoke a hand's width from a dedicated microphone with a pop filter, and Audo's free LUFS meter (/lufs-meter) already shows a noise floor below −60 dB with 0 clipped samples, upload the original WAV straight to your voice cloning platform.
- Clean with Audo AI first. If you are cloning from existing podcast episodes, interviews, videos, or a home-office mic with audible room hiss, fan hum, or long pauses, run non-generative Remove noise and Remove silence first.
Prep your voice sample without changing your voice
Drop your recording into Audo's non-generative Remove noise to strip hiss and hum while keeping your real vocal character, or tighten pauses with Remove silence. Free to try. No sign-up.
Questions
Should I use an AI audio enhancer before uploading to ElevenLabs?
Avoid full-strength generative enhancers (like Adobe Enhance Speech or Studio Sound) when preparing a voice clone sample, because they resynthesize the voice and strip away the subtle vocal traits the cloner needs to learn. Instead, use Audo AI's non-generative Remove noise (/noise-removal), which filters out background hiss, hum, and music while keeping the original voice intact.
How do I remove background music from a clip I want to use for voice cloning?
Drop the audio or video file into Audo AI's Voice isolator (/voice-isolator) or Remove music from video (/remove-music-from-video). Both run Audo's non-generative speech filter to strip background music and ambient noise while preserving the speaker's natural voice.
How can I check if my voice sample has clipping or a high noise floor?
Upload the file to Audo AI's free LUFS meter (/lufs-meter) or File info (/media-info). It measures integrated loudness in LUFS, true peak in dBTP, RMS noise floor in dB, and exact clipped sample count for the entire file—always free with no sign-up.
What does it cost to clean up a 30-minute Professional Voice Clone dataset on Audo AI?
The first minute of Remove noise and the first hour of Remove silence are free to try with no sign-up. Cleaning a full 30-minute recording with Remove noise costs $0.90 ($0.03 a minute, or $1.80 an hour), paid once from $5 with no subscription.