How do I change the sound of my voice in a recording, but keep my own delivery?

This page may contain affiliate links from which Blur Studio earns commissions.

This is a narrower and more specific need than most voice-AI content assumes. Plenty of guides jump straight to "generate a voice from text," which throws away the one thing you actually like about your own recording: how you say it. The tool that fits here is ElevenLabs' Voice Changer, also called speech to speech, built specifically to swap voice identity while preserving performance.

Why this is different from generating a voice from text

Text-to-speech starts from written words and a voice model decides the pacing, emphasis and emotion for you, which you can steer with settings but never fully control the way you can by actually performing the line yourself. Speech to speech starts from your real recording and changes only the voice identity layered on top, keeping your actual timing, your pauses, your stress on particular words. If the problem is specifically "I don't like my voice, not my delivery," speech to speech is the tool built for exactly that distinction.

How to do it

1. Record the line the way you'd actually say it. Don't try to perform it differently in anticipation of the voice change; your natural delivery is the thing being preserved, so record it normally.
2. Pick a target voice. This can be one of ElevenLabs' 10,000+ Voice Library options, a voice you've designed from a text description, or a clone, as long as you have the rights to use it.
3. Upload it to Voice Changer and convert. The source file needs to stay under 50MB, and a single conversion caps at roughly 5 minutes on most plans, which covers most individual recordings without needing to split anything.
4. Listen for anything that slipped through. Breathing sounds, laughs, or filler words in your original take sometimes convert oddly, since the model is working from your actual audio, not a clean script. A second take of just the rough section is usually faster than trying to fix it after conversion.

A worked example: anonymous feedback that still sounds human

Say you're collecting internal feedback or a case-study quote and the person speaking wants to stay anonymous, but a flat, obviously synthetic voice would undercut how genuine the content feels. Text-to-speech from a script solves the anonymity problem but loses the real pauses, emphasis and emotion that make a testimonial land as sincere rather than scripted. Here's the actual sequence:

1. Record the person saying their real feedback, in their own words, with their own natural pacing. Don't have them read from a polished script; the more natural the original delivery, the better the final result carries that same authenticity.
2. Pick a target voice that's clearly not identifiable as them, either a designed voice or a different Voice Library option entirely.
3. Convert the recording and listen specifically for whether the emotional beats (a pause before an honest admission, emphasis on the part that actually mattered to them) survived the conversion. They should, since that's exactly what speech to speech preserves.
4. Use the converted clip, and disclose plainly that the voice has been changed to protect the speaker's identity, which in this specific case is the whole point rather than something to hide.

The result sounds like a real person speaking honestly, because it is, just without the voice that would identify them.

Choosing the right kind of target voice

Not every target voice choice serves the same goal, and it's worth being deliberate rather than picking whatever sounds pleasant. A Voice Library voice is the fastest option and works fine when you just want a different, generic-sounding voice with no particular identity attached. A voice you've designed from a text description gives you more control over specific qualities (warmth, age, accent) without borrowing anyone's actual identity, which matters if the context calls for a distinctive but clearly nobody's-real-voice option. A clone, of yourself in a different register, or of someone who's given explicit permission, is the right call when you specifically want the result to sound like a particular real person rather than just "a voice." Match the choice to why you're changing the voice in the first place, privacy, variety, or a specific desired identity, rather than defaulting to whichever option is first in the list. Playing several parts yourself is covered in playing several characters yourself.

A real tip from actually doing this

The quality of your source recording matters more than people expect. A clean, close-mic recording with minimal background noise converts noticeably better than a phone recording in a room with any echo or hum, because the model is working with your actual audio, not just extracting the words from it. If a conversion comes out sounding slightly off or artificial, check your source recording's cleanliness before assuming the voice choice is the problem. The fix is often a better recording environment, not a different target voice.

What you can actually adjust once you've converted

Voice Changer isn't a one-shot, take-it-or-leave-it conversion. Once you've picked a target voice, four sliders let you shape how the result lands: Speed (how quickly the converted voice speaks, independent of your original pacing), Tone (the overall character of the voice), Pacing (fine control over delivery rhythm beyond the raw speed setting), and Style Exaggeration (how much the target voice's distinctive traits get pushed, from subtle to pronounced). Start with all of them at their defaults and change your source recording before you reach for these, since they're meant for small adjustments to the output, not for fixing a problem that's really in your original take. A common mistake is cranking Style Exaggeration to cover for a flat performance; it makes the voice sound more "characterful," but it won't put back emotion that wasn't in your recording to begin with. Changing the accent while keeping the voice is covered in using the same voice with a different accent.

You'll also see a model choice (v3 or v4, with v4 billed as the more emotive option). For most voice-changing work, v4 is worth defaulting to, since the whole point of this tool is preserving your performance's emotional nuance, and that's exactly what the newer model is built to carry through more faithfully. Output comes back as WAV or PCM at 44.1kHz, which is the standard you want for anything going into further video or audio editing rather than a compressed format. Processing itself is fast, on the order of 400ms for a short clip, so iterating on a take (convert, listen, adjust the slider, convert again) is genuinely quick rather than something you queue up and wait on.

A second worked example: a recurring branded voice for a segment

Say you host a short weekly update, a "what's new" segment, a recurring internal briefing, and you want it to have a consistent, recognizable character voice rather than sounding like you reading a script every time. Text-to-speech from a script would give you consistency but none of your actual timing or inflection, episode to episode. Here's how to keep both:

1. Pick one target voice and stick with it across every episode. Consistency is the entire value of a recurring "character," so resist the urge to try a different voice each week even if you're tempted by the variety in the Voice Library.
2. Record each episode the way you'd actually deliver it live, keeping your natural rhythm rather than over-performing for the conversion. The tool carries your real delivery through, so an exaggerated, "podcast voice" read will sound exaggerated in the output too.
3. Keep your slider settings (Speed, Tone, Pacing, Style Exaggeration) identical from episode to episode, written down somewhere so whoever produces the audio doesn't have to guess or re-tune them each time. Small week-to-week drift in these settings is what makes a recurring voice start to sound subtly inconsistent over a season.
4. Convert, listen for the same emotional beats you intended live, and publish.

Done this way, listeners get a consistent character voice that still has the texture of a real person actually talking to them, not a script being read. Keeping one voice across a set of videos is covered in getting the same voice across demo videos.

When a conversion sounds off, and it's not the recording

If your source audio is clean and the result still sounds subtly wrong, the usual culprit is a mismatch between your natural voice and the target voice's underlying register, a deep, slow-speaking target voice applied to a fast, high-pitched original take can produce odd-sounding pacing artifacts, even though the system is technically doing its job correctly. The fix isn't more slider tweaking; it's picking a target voice whose natural range sits closer to your own, or dialing back Style Exaggeration so the target voice's distinctive traits don't fight against your original delivery's energy. This is also where testing with a short clip before converting a full recording pays off, since you'll hear this kind of mismatch in 10 seconds rather than after converting a 4-minute file. A similar energy problem with cloned voices is covered in fixing a cloned voice that sounds tired or slow.

When to disclose the change

If the context is one where people reasonably assume they're hearing your actual voice, a testimonial, a public statement, anything where trust in who's speaking matters, say plainly that the voice has been changed. There's no universal legal line here that applies to every situation, but there is an obvious reputational one: being caught using a changed voice somewhere people expected to hear the real you costs you more credibility than the content gains from sounding different.

How I'd approach this

For an occasional recording, this is a self-serve job, record, convert, check the result, done. Where it's worth something more built-out is a business doing this routinely, for example running the same presenter's recordings through a consistent voice change every time content gets produced. In that case I'd wire the conversion step directly into whatever system already handles the recording or editing, so it happens automatically rather than as a manual extra step someone has to remember.