Why does my conversion sound robotic?

Nearly always the source rather than the settings. The four usual causes are more than one voice in the recording, background noise, a speaker recorded across a room, and a target voice a long way from your own in pitch. What you will hear when one of them is at work is that the identity slips, consonants smear, the pitch wobbles, breaths go strange and background comes alive. None of them produces an error — the job finishes and the credits go — so the only protection is judging the recording before you upload it.

Should I clean up a noisy recording first?

Yes, and in that order. Steady noise is part of the signal as far as conversion is concerned. Hum and hiss do not get ignored; they get re-voiced along with the speech, and what comes back is a strange, unstable texture underneath the words rather than the clean hum you started with. Run the cleanup first, then convert the cleaned file. Doing it the other way round does not work: once the noise has been re-voiced it is inside the speech, and no cleanup separates it out again. Cleaning a recording well is its own subject — see /guides/denoise.

Can it convert two people in one recording?

Not usefully. Conversion has no concept of who is talking. It converts the audio, not a speaker — so a second voice, an interviewer's agreement, a laugh off-mic all get pushed through the same target voice, and the result is one person appearing to talk over themselves. Cut to a stretch where only your speaker is audible, or run speaker separation first and convert one speaker's track.

Should I turn Preserve Timing off?

Holds your source recording's rhythm — where the pauses fall, how long each word takes — instead of letting the target voice re-time the line. Leave it on. It is what keeps a converted take in sync with video, and turning it off is a choice you should make deliberately rather than while exploring.

It still isn't right. Should I just run it again?

A re-run at the same settings on the same file gets you the same class of result and costs the same again — 4,950 credits for a five-minute take. Change something first: a shorter, cleaner excerpt, a closer take, or a nearer target. If the read itself is the problem, record the source again; conversion cannot fix a performance.

Voice Changer

Getting a convincing result

Q: Does the Quality setting make it more convincing?

Less than you would hope, and less than the menu suggests. The menu offers three options and the request carries two. High is sent as high; Low and Medium are both sent as standard, so switching between those two changes nothing whatsoever. Treat it as a two-position switch — High, or not High — and ignore the third entry. It costs nothing to leave it on High, since the credit figure is the length of your source either way, so there is no reason to run anything else.

Q: Will Pitch Shift make it sound more like the target?

No, and reaching for it is the commonest mistake on this tab. Transposes the result up or down after conversion. It is a transposition, not a re-performance: every note moves by the same amount, so a large shift makes the voice sound processed rather than making it sound like a different person. This is the control most likely to be misused, because it is the one that looks like an effect. If a target is too far from your voice, a transposition on top of the conversion makes the result sound more processed rather than more like the person. Change target instead.

Almost every bad conversion is the source recording, not the settings. One speaker, a quiet room, close miking, and a target within reach of your own voice — plus how to hear a conversion that has not worked.

Lesson 2 · core · 8 min read

Open Voice Changer

Conversion is only as good as what you hand it. It keeps your timing and delivery and swaps the timbre — which means everything already printed onto your recording gets carried through the swap, and some of it survives the journey badly. Almost every disappointing result on this tool traces back to the source file, and almost none of them traces back to the settings pane.

Careful

None of what follows produces an error. A recording with three people in it and a fan running converts successfully, costs full price, and comes back unusable. The tool does not check, does not warn, and reports the job as complete — so knowing this beforehand is the only protection there is.

Five things about the source

In rough order of how much difference they make. The first two are the ones that ruin results outright; the last three are the difference between usable and convincing.

The recording: Why it matters, and the fix
One person, speaking alone: Conversion has no concept of who is talking. It converts the audio, not a speaker — so a second voice, an interviewer's agreement, a laugh off-mic all get pushed through the same target voice, and the result is one person appearing to talk over themselves. Cut to a stretch where only your speaker is audible, or run speaker separation first and convert one speaker's track.
A quiet background: Steady noise is part of the signal as far as conversion is concerned. Hum and hiss do not get ignored; they get re-voiced along with the speech, and what comes back is a strange, unstable texture underneath the words rather than the clean hum you started with. Clean the source first, then convert. The noise-reduction track covers doing that well, and running the cleanup before the conversion is the right order — after is too late, because the artefacts are now inside the voice.
Close, present speech: A voice recorded across a room arrives mixed with the room. Reverb moves the way speech moves, so it cannot be told apart from the voice, and a distant source converts into a distant, smeared result no matter which target you pick. Use the closest-miked take you have. If you are recording for conversion, record close and dry — this is worth more than any setting on the tab.
A target in reach of the source: The further the target voice sits from yours in pitch and build, the more work the conversion has to do and the more of it you hear. A very low voice converted to a very high one is the case that most often comes back sounding synthetic, and no amount of retrying changes that. Pick a target closer to your own range first and hear what a good result sounds like. Then judge the far-off target against that rather than against your expectations.
A delivery worth keeping: Timing, emphasis and energy all come from your recording, not from the target. Conversion cannot make a flat read sound engaged — it makes a flat read sound like somebody else being flat. Perform the source take. If the read is the problem, the fix is another take, not another target voice.

What decides a conversion, and what to do about each.

Tip

The order matters and it is the one thing people get backwards: clean, then convert. Noise that has been through a conversion is no longer noise sitting under speech — it has become part of the voice, and nothing takes it out afterwards. /guides/denoise covers doing the cleanup well, including the recordings it cannot rescue.

How far the target can be from you

The further the target voice sits from yours in pitch and build, the more work the conversion has to do and the more of it you hear. A very low voice converted to a very high one is the case that most often comes back sounding synthetic, and no amount of retrying changes that. Pick a target closer to your own range first and hear what a good result sounds like. Then judge the far-off target against that rather than against your expectations.

Careful

Do not try to close that gap with Pitch Shift. Transposes the result up or down after conversion. It is a transposition, not a re-performance: every note moves by the same amount, so a large shift makes the voice sound processed rather than making it sound like a different person. This is the control most likely to be misused, because it is the one that looks like an effect.

What a failed conversion sounds like

A conversion rarely fails loudly. It comes back intelligible, on time, every word present — and not quite anybody. Five things to listen for, and any one of them means stop and change the input rather than running it again.

The identity slips. It sounds like the target voice on the loud words and like nobody in particular on the quiet ones. Consistency across a whole sentence is the hardest thing for a conversion to hold, and the ends of phrases are where it lets go.
Consonants smear. The s, t and f sounds arrive soft or doubled. This usually means the source was noisy or distant rather than that the target was wrong.
The pitch wobbles. Sustained vowels drift or warble slightly, the way a held note does when it cannot decide. Almost always a sign the target sits too far from the source.
Breaths go strange. Breaths and mouth noises are audio too, so they get converted as well, and a converted breath is one of the most recognisably artificial sounds this tool produces.
Background comes alive. A steady hum that was merely dull in the source becomes something that moves and shimmers. That is the noise being re-voiced, and it is the clearest possible sign the source needed cleaning first.

Tip

Listen to a whole sentence, not a word. Conversion holds identity best on loud, well-articulated syllables and worst at the quiet ends of phrases, so a single word from the middle of a strong line will sound fine in a take that falls apart everywhere else.

Write this

Not this

Where you look first when it sounds wrong

The recording you uploaded

The settings pane

A noisy source

Clean it, then convert the cleaned file

Convert it and hope the noise gets lost

A target far from your voice

Pick a nearer target, or re-cast the part

Add pitch shift to meet it halfway

Two people on the recording

Separate them, convert one track

Convert the whole thing

A half-hour episode

Run thirty seconds first — 495 credits against 29,700

Run the whole thing and find out

A flat read

Record the source again — the delivery is yours, not the target's

Try another target voice

Where you look first when it sounds wrong

Write this:: The recording you uploaded
Not this:: The settings pane

A noisy source

Write this:: Clean it, then convert the cleaned file
Not this:: Convert it and hope the noise gets lost

A target far from your voice

Write this:: Pick a nearer target, or re-cast the part
Not this:: Add pitch shift to meet it halfway

Two people on the recording

Write this:: Separate them, convert one track
Not this:: Convert the whole thing

A half-hour episode

Write this:: Run thirty seconds first — 495 credits against 29,700
Not this:: Run the whole thing and find out

A flat read

Write this:: Record the source again — the delivery is yours, not the target's
Not this:: Try another target voice

What the settings are actually worth

Less than the source, by a long way. Two of the three are worth understanding mainly so that you stop reaching for them.

Note

The menu offers three options and the request carries two. High is sent as high; Low and Medium are both sent as standard, so switching between those two changes nothing whatsoever. Treat it as a two-position switch — High, or not High — and ignore the third entry.

Preserve Timing — On or off. On by default. Holds your source recording's rhythm — where the pauses fall, how long each word takes — instead of letting the target voice re-time the line. Leave it on. It is what keeps a converted take in sync with video, and turning it off is a choice you should make deliberately rather than while exploring.
Quality — Low, Medium, High. Starts on High. Trades processing time against result. It is not a strength dial and it does not change what you are billed — the credit figure is the length of your source and nothing else, so there is no cheaper setting to fall back to.
Pitch Shift — -12 to +12 semitones, in whole steps of 1. Default 0. Transposes the result up or down after conversion. It is a transposition, not a re-performance: every note moves by the same amount, so a large shift makes the voice sound processed rather than making it sound like a different person. This is the control most likely to be misused, because it is the one that looks like an effect.

Test on an excerpt

Cost scales with the length of the source and nothing else, so the price of being wrong is entirely a function of how much you uploaded. Thirty seconds from the middle of a long recording costs 495 credits and tells you everything the full run would: whether this source and this target produce a person.

Note

Take the excerpt from the middle, not the top. The first minute of any recording is the least representative part of it — people settle, rooms settle, and the noise you are actually fighting turns up once everyone has stopped being careful.

Get the source right and AudioPod's voice conversion is a good tool with a narrow, real job. Which leaves the question the tool itself never asks, and the one that decides whether you should be running it at all: whose voice is this?

Questions

What people ask about this

Free to start

Now go make one

Reading about a style description only gets you so far. The studio is free to use — write one sentence and hear what comes back.

Create a free account

1,000 credits every month. No card required.

All Voice Changer lessons

Voice Changer

Getting a convincing result

Lesson 2 · core · 8 min read

Open Voice Changer

Careful

Five things about the source

In rough order of how much difference they make. The first two are the ones that ruin results outright; the last three are the difference between usable and convincing.

The recording: Why it matters, and the fix
One person, speaking alone: Conversion has no concept of who is talking. It converts the audio, not a speaker — so a second voice, an interviewer's agreement, a laugh off-mic all get pushed through the same target voice, and the result is one person appearing to talk over themselves. Cut to a stretch where only your speaker is audible, or run speaker separation first and convert one speaker's track.
A quiet background: Steady noise is part of the signal as far as conversion is concerned. Hum and hiss do not get ignored; they get re-voiced along with the speech, and what comes back is a strange, unstable texture underneath the words rather than the clean hum you started with. Clean the source first, then convert. The noise-reduction track covers doing that well, and running the cleanup before the conversion is the right order — after is too late, because the artefacts are now inside the voice.
Close, present speech: A voice recorded across a room arrives mixed with the room. Reverb moves the way speech moves, so it cannot be told apart from the voice, and a distant source converts into a distant, smeared result no matter which target you pick. Use the closest-miked take you have. If you are recording for conversion, record close and dry — this is worth more than any setting on the tab.
A target in reach of the source: The further the target voice sits from yours in pitch and build, the more work the conversion has to do and the more of it you hear. A very low voice converted to a very high one is the case that most often comes back sounding synthetic, and no amount of retrying changes that. Pick a target closer to your own range first and hear what a good result sounds like. Then judge the far-off target against that rather than against your expectations.
A delivery worth keeping: Timing, emphasis and energy all come from your recording, not from the target. Conversion cannot make a flat read sound engaged — it makes a flat read sound like somebody else being flat. Perform the source take. If the read is the problem, the fix is another take, not another target voice.

What decides a conversion, and what to do about each.

Tip

How far the target can be from you

Careful

What a failed conversion sounds like

The identity slips. It sounds like the target voice on the loud words and like nobody in particular on the quiet ones. Consistency across a whole sentence is the hardest thing for a conversion to hold, and the ends of phrases are where it lets go.
Consonants smear. The s, t and f sounds arrive soft or doubled. This usually means the source was noisy or distant rather than that the target was wrong.
The pitch wobbles. Sustained vowels drift or warble slightly, the way a held note does when it cannot decide. Almost always a sign the target sits too far from the source.
Breaths go strange. Breaths and mouth noises are audio too, so they get converted as well, and a converted breath is one of the most recognisably artificial sounds this tool produces.
Background comes alive. A steady hum that was merely dull in the source becomes something that moves and shimmers. That is the noise being re-voiced, and it is the clearest possible sign the source needed cleaning first.

Tip

Write this

Not this

Where you look first when it sounds wrong

The recording you uploaded

The settings pane

A noisy source

Clean it, then convert the cleaned file

Convert it and hope the noise gets lost

A target far from your voice

Pick a nearer target, or re-cast the part

Add pitch shift to meet it halfway

Two people on the recording

Separate them, convert one track

Convert the whole thing

A half-hour episode

Run thirty seconds first — 495 credits against 29,700

Run the whole thing and find out

A flat read

Record the source again — the delivery is yours, not the target's

Try another target voice

Where you look first when it sounds wrong

Write this:: The recording you uploaded
Not this:: The settings pane

A noisy source

Write this:: Clean it, then convert the cleaned file
Not this:: Convert it and hope the noise gets lost

A target far from your voice

Write this:: Pick a nearer target, or re-cast the part
Not this:: Add pitch shift to meet it halfway

Two people on the recording

Write this:: Separate them, convert one track
Not this:: Convert the whole thing

A half-hour episode

Write this:: Run thirty seconds first — 495 credits against 29,700
Not this:: Run the whole thing and find out

A flat read

Write this:: Record the source again — the delivery is yours, not the target's
Not this:: Try another target voice

What the settings are actually worth

Less than the source, by a long way. Two of the three are worth understanding mainly so that you stop reaching for them.

Note

Preserve Timing — On or off. On by default. Holds your source recording's rhythm — where the pauses fall, how long each word takes — instead of letting the target voice re-time the line. Leave it on. It is what keeps a converted take in sync with video, and turning it off is a choice you should make deliberately rather than while exploring.
Quality — Low, Medium, High. Starts on High. Trades processing time against result. It is not a strength dial and it does not change what you are billed — the credit figure is the length of your source and nothing else, so there is no cheaper setting to fall back to.
Pitch Shift — -12 to +12 semitones, in whole steps of 1. Default 0. Transposes the result up or down after conversion. It is a transposition, not a re-performance: every note moves by the same amount, so a large shift makes the voice sound processed rather than making it sound like a different person. This is the control most likely to be misused, because it is the one that looks like an effect.

Test on an excerpt

Note

Questions

What people ask about this

Free to start

Now go make one

Reading about a style description only gets you so far. The studio is free to use — write one sentence and hear what comes back.

Create a free account

1,000 credits every month. No card required.

All Voice Changer lessons