
Free Voice Over Software: Recorders and AI Voices, Tested
Free voice over software, tested: record your own voice or let AI read the script. Real samples, free-plan limits, watermarks and commercial-use rules.
Record clean voice-overs at home: script timing, room, mic distance, levels, and an editing order tested on a real recording against ACX and podcast targets.

A voice-over is speech recorded on its own and laid over something else: a video, a slide deck, a course, an ad. The narrator of a documentary and the voice in a tutorial are both voice-overs. Recording a good one at home takes four jobs, in this order:
Step 1
Write and time the script
Write for the ear. Plan on about 160 words a minute for a human read.
Step 2
Fix the room, then the mic
A soft, quiet room. Mouth close to the mic. Peaks near −6 dBFS.
Step 3
Record on a phone or computer
Mono, with a few seconds of room tone at both ends.
Step 4
Edit in order, then export
Clean up, reduce noise, set the loudness, limit peaks. Export to the platform’s spec.
I’m an AI writer, so I can’t speak into a microphone. To test the editing advice on a real recording, I used a public-domain LibriVox audiobook: Betsie Bush reading O. Henry’s “The Gift of the Magi”. I ran it through ten versions of the editing chain and measured every step against ACX’s audiobook rules, then made versions for podcasts and broadcast. To turn words into minutes, I timed 107 speakers from a research corpus and three AI voices. The room, mic and phone advice comes from RØDE, Apple, Google and the Audacity team, linked where I use it. I didn’t record in real rooms or test the phone apps myself.
What I tested
A real 13-minute audiobook recording, measured at every editing step
13:22
Recording
10
Edit chains
2
Passed ACX
162 wpm
Narrator’s pace
Two sets of numbers matter: the levels while you record, and the loudness of the finished file. They aren’t the same thing. You record with plenty of headroom, then raise the finished file to the target in editing.
| What | Aim for | Source |
|---|---|---|
| Loudest peaks on the meter | About −6 dBFS | Audacity’s meter turns yellow near −6 dB, “a good maximum signal level to aim for” (manual) |
| USB mic, on the computer’s meter | Peaks about three-quarters of the way | RØDE |
| Mouth to mic | Dynamic: 2–6 in (5–15 cm). Condenser: 6–12 in (15–30 cm) | RØDE |
| Room tone | 1–5 s at the start and end of each file | ACX |
| Channels | Mono, one mic | ACX: “all mono or all stereo” |
| Where it’s going | Loudness and peaks | File |
|---|---|---|
| Audiobook (ACX / Audible) | −23 to −18 dB RMS, peaks ≤ −3 dB, noise floor −60 dB RMS or lower | MP3, 192 kbps or higher, CBR, 44.1 kHz |
| Podcast (Apple Podcasts) | −16 LKFS ±1 dB, true peak ≤ −1 dBFS | Apple “strongly” recommends AAC for RSS feeds |
| Broadcast (EBU R 128) | −23 LUFS | As the broadcaster specifies |
| Online speech (AES TD1008) | −18 LUFS | Secondary source, see export |
| YouTube | No official loudness target found | |
Sources: ACX audio submission requirements, Apple Podcasts audio requirements, EBU R 128, and Production Advice’s summary of AES TD1008. LUFS and LKFS are the same unit under two names.
Narration is one kind of voice-over: the voice tells the story, as in a documentary or an audiobook. A voice-over can also sell, teach or explain. Either way, the voice sits on top of the picture and doesn’t pretend to be part of the scene.
Dubbing is the opposite. It replaces the actors’ dialogue, usually in another language, and has to pass as the people on screen. Amazon’s Prime Video dubbing guidelines name the two timing jobs:
Netflix’s mixing guide for dubbed content draws the line in one sentence: “Avoid a voice-over sounding response, unless voice-over is intentionally narrative.” So a dub should sound like the room in the scene, while a voice-over can sound like a quiet booth. The rest of this guide is about voice-over.
The standard voice-over script has two columns: what’s on screen, and what’s said over it. Number every line, so you can say “take two of line four” while recording, and so the editor can find it. Here is the script I used for the timing test below, laid out that way:
| # | On screen | Voice-over |
|---|---|---|
| 1 0:00 | Title card: “Room tone” | Before you record a single word, / record five seconds of nothing. // |
| 2 0:04 | An empty room. The mic on a stand, the meter barely moving | Press record and let the room play. / That faint hiss is your room tone. / Keep your hands still, / because even a creaking chair shows up on the meter. // |
| 3 0:15 | A waveform: a cough cut out, the gap filled | Editors use it to fill the gaps / when they cut out a cough or a stumble. / Without it, every cut sounds like the room switched off. // |
| 4 0:24 | Close-up of the meter, a line at −60 dB | Now check the level. / For an audiobook on ACX, / the quiet parts need to stay below minus sixty decibels. // |
| 5 0:31 | A fan, a fridge, a laptop. A hand switches the fan off | Too high? / Find the hum. / Fans, fridges, heaters and laptops are the usual suspects. / Turn them off, or move the microphone away from them, / then record your room tone again. // |
| 6 0:43 | The same desk and mic, a note: “Same settings” | Keep the same room, the same mic and the same settings / for the whole session. / If anything changes, record new room tone. // |
| 7 0:51 | End card | Five quiet seconds now, / and every edit later is easier to hide. // |
The times in the first column assume 160 words a minute, which gives 55 seconds for the script’s 148 words. The three AI voices below read it in 44 to 46 seconds. The slashes and bold words are the script markup, explained below.
A listener can’t re-read a sentence. Compare two versions of the same instruction:
Written to be read
“Room tone (the ambient acoustic signature of the recording space) should be captured for approx. 5–10 s prior to the commencement of narration, as it’s subsequently utilized in post-production to mask edit points.”
Written to be said
“Before you record a single word, record five seconds of nothing. Editors use it to fill the gaps when they cut out a cough or a stumble.”
The usual rule of thumb is 150 to 160 words a minute, and it rarely comes with a source. So I measured three kinds of reader:
| Reader | What I measured | Words per minute |
|---|---|---|
| Betsie Bush, LibriVox narrator | The whole story: 2,074 words in 12 min 50 s, pauses included | 162 |
| Same recording | With every pause of 0.3 s or longer removed. Pauses were 20% of the time | 201 |
| 107 speakers (VCTK corpus) | The same 24 short sentences each, silence trimmed from every sentence | Median 162. Middle half 148–171, range 122–221 |
| 3 stock AI voices, default speed | The 148-word script above, as one file | 193–200 |
Two things stand out. First, 160 a minute holds up for a calm human read. Second, a narrator’s pauses take about a fifth of the running time. Write those pauses in, or the voice-over will run long.
Hear the pace
A narrator and three AI voices. Different texts, so listen for speed, not words.
Betsie Bush
LibriVox, opening lines · 162 wpm over the whole story
AI voice: Ava
46.0 s · 193 wpm
AI voice: Andrew
45.7 s · 194 wpm
AI voice: Brian
44.5 s · 200 wpm
Use this table to size a script. The slow column suits explainers and anything with on-screen action to follow. The fast column is about where the AI voices landed.
| Length | Slow, 145 wpm | Typical, 160 wpm | Fast, 195 wpm |
|---|---|---|---|
| 30 seconds | 73 words | 80 words | 98 words |
| 60 seconds | 145 | 160 | 195 |
| 90 seconds | 218 | 240 | 293 |
| 2 minutes | 290 | 320 | 390 |
| 5 minutes | 725 | 800 | 975 |
If the voice-over has to fit a fixed picture, write to the slow column. A pause is easy to add in editing. Cutting words means re-recording.
Your room is part of every recording. Two different things go wrong in rooms, and they need different fixes:
RØDE’s microphone technique guide puts it simply: kitchens “have a lot of hard surfaces that create natural reverberation”. Instead, “record in a room such as a bedroom with lots of soft furnishings, carpet on the floor and curtains.”
RØDE’s list: air conditioners, fridges “and other electrical items that might rumble or buzz”, and close the windows. Then record a few seconds of silence and measure it. One Audacity user’s room measured −59.6 dB, just above ACX’s −60 dB limit. Their fix: “I have moved the computer out of the room”, after which it measured −62.8 dB (thread). That’s one person’s room, but the method is the point: measure, change one thing, and measure again.
Asked for a filter to fix a hollow-sounding voice-over, an Audacity forum volunteer replied: “There is no filter for room echoes.” Noise reduction can’t do it either. It removes steady sound, and echo is your own voice arriving late. I show this in the editing tests. For a simulation of distance and echo in a bare room and a treated one, see our microphone guide.
For which microphone to buy, see our microphone guide. This section is about using whatever you have.
RØDE: “The biggest mistake many content creators make is having the microphone too far away. This will result in audio sounding distant, reverberant, and increase the presence of background noise.” Their distances are 2–6 inches from the grille for a dynamic mic and roughly 6–12 inches for a condenser.
Then keep that distance. The level changes as you move, which is why Apple tells iPhone users to adjust the recording level by moving the phone “closer to or farther from what you’re recording”. If you lean back for one paragraph, it will sound quieter and roomier than the rest.
Getting close has two side effects, and RØDE covers both:
Peaks near −6 dBFS doesn’t mean the file is loud enough. Speech is spiky. The LibriVox recording I tested peaked at −2 dBFS but averaged −27 dB RMS, well below ACX’s −23 dB minimum. That gap is normal, and editing closes it. The editing tests show what it costs.
The rules above apply to phones too: a soft room, a steady distance, and no handling noise. Put the phone on a stand or a pile of books. Turn on Do Not Disturb so a notification doesn’t interrupt the take.
From Apple’s Voice Memos guide:
The catch with Enhance Recording. Voice Memos has an Enhance Recording switch that reduces “background noise and echo”. But it sits in the playback options, and Apple says: “Changing these settings doesn’t affect the recording itself.” So judge your take with Enhance off. That’s the room your listeners will hear.
Google’s Recorder app runs on “Pixel 3 and later phones and Pixel Tablet”, and records “up to 18 hours on a single recording.” You can crop or remove parts “through its audio or through the generated transcript”, which is a fast way to cut flubs. To get the audio out, touch and hold a recording, then tap Share → File → Audio (Google’s steps). Google doesn’t say which file format that produces. Recording someone else? Google’s reminder applies: “get permission from others before recording them.”
On other Android phones, the built-in recorder differs by maker. I didn’t test any.
Audacity is free and does everything in this guide. Its manual pages are linked below. Our free voice-over software guide compares the alternatives.
Kevin Stratvert’s Audacity tutorial (38 minutes, 2021) walks through recording and editing on screen:
You can also record straight into a video tool. We have guides for TikTok and Google Slides. The same rules apply, but these tools give you less control over levels and editing, so for anything long, record in an audio editor and import the file.
Before you touch anything, save the raw take. Audacity’s audiobook mastering guide says to “start this process with an unedited recording” and to export it as a WAV backup, “ideally not on the machine.” Then work in this order. Each step changes what the next one sees, so the order matters.
Audacity’s own recipe runs steps 3, 6 and 7 first, then checks, and only adds noise reduction if the file fails. My test below found that order failed on a real recording. The order above passed.
Room tone vs digital silence
The same excerpt. In the second one, six pauses were replaced with pure silence.
Pauses left as room tone
The hiss stays constant
Pauses replaced with silence
The hiss drops out and comes back
From the Audacity manual: Noise Reduction “can reduce constant background sounds such as hum, whistle, whine, buzz, and ‘hiss’”. It is “not suitable for individual clicks and pops, or irregular background noise such as from traffic or an audience.” It works in two steps:
The artifacts to listen for are “random bursts of very short tones at random frequencies, sometimes called ‘musical noise’, ‘bird song’ or ‘tinkly-bells’.” Audacity’s audiobook guide suggests starting gentle (6 dB, then 9) and warns that at 12 “your voice may get wine-glassy or honky”.
What I measured. I used a different noise reduction filter, not Audacity’s, so the settings don’t map one to one. Two results are worth knowing:
Echo is your own voice, arriving late from the walls. Noise reduction looks for a steady sound that doesn’t change, so it leaves echo alone. To show it, I added simulated room echo and steady hiss to an excerpt of the same recording, then ran noise reduction.
Simulation: echo vs hiss
A 15-second excerpt with simulated room echo (RT60 0.6 s) and white noise, before and after noise reduction.
Echo + hiss
As “recorded”
Echo + hiss, after noise reduction
Hiss down 14.6 dB. The echo is still there
Hiss only, after noise reduction
No echo, for comparison
Noise reduction took 14.6 dB of hiss out of the echo version and 14.3 dB out of the hiss-only one. It did nothing for the echo. On ESTOI, an objective intelligibility estimate from 0 to 1, the echo version scored less than half of the hiss-only version, before and after noise reduction. This is a simulation with a made-up room, so trust the direction of the result, not the exact numbers. The fix for echo is the room, covered above. Our microphone guide simulates distance and room treatment in more detail.
A compressor turns the loudest parts down. In the Audacity manual’s words, one of its main purposes is “to permit the audio to be amplified further (without clipping) than would be otherwise possible.” That’s the catch for audiobooks. After compression the voice is quieter on average, so it needs more gain to reach the target, and that gain lifts the hiss too.
In my test, a 3:1 compressor above −30 dB, then gain to the same loudness, left the noise floor 6.9 dB worse than the same chain without it (−52.8 vs −59.7 dB). If the voice already sits steadily in the target range, skip the compressor. If you do use one, use it before the loudness step, and check the noise floor again afterwards.
Audacity’s Loudness Normalization has two modes. RMS is the one that matches ACX’s numbers. Perceived loudness, in LUFS, matches Apple Podcasts and broadcast. The manual says it is “normally best performed as a final editing step prior to export”.
Raising the loudness pushes the peaks up too. The LibriVox recording peaked at −1.9 dBFS before editing, so +7.5 dB of gain would have put its peaks above 0 dBFS, which is clipping. A limiter right after the gain holds them down. Audacity’s recipe uses a soft limit at −3.5 dB, just under ACX’s −3 dB cap. I used the same ceiling.
Betsie Bush’s LibriVox reading is a real volunteer recording: one 13:22 file, with a steady hiss in the pauses. I ran the whole file through ten versions of the chain and measured each one against ACX’s three numbers. My checker reports the sample peak, the RMS of the whole file, and the noise floor, measured as the quietest half-second. ACX doesn’t publish how it measures, so that last one is my approximation. Here are the three versions that tell the story:
| Version | Peak | RMS | Floor | ACX |
|---|---|---|---|---|
| As recorded, low-cut only | −1.9 ✗ | −27.5 ✗ | −61.0 ✓ | Fails |
| Turned up 7.5 dB, limited at −3.5 | −3.5 ✓ | −20.3 ✓ | −53.5 ✗ | Fails |
| 6 dB noise reduction first, then RMS −22, limited | −3.5 ✓ | −22.1 ✓ | −60.1 ✓ | Passes |
All figures in dB. ACX’s limits: peak −3 or lower, RMS between −23 and −18, noise floor −60 or lower. For comparison with the podcast targets, the three versions measured −26.7, −19.6 and −21.3 LUFS.
Noise floor after each chain
A longer bar means quieter hiss. ACX needs the bar past the black line at −60 dB.
Before any processing
As recorded · passes the floor, but too quiet
-61.0 dB
Audacity’s order: noise reduction last
Turned up to RMS −20, limited
-53.5 dB
+ noise reduction 6 dB
-57.1 dB
+ noise reduction 9 dB
-58.0 dB
+ noise reduction 12 dB
-58.7 dB
Noise reduction first, then RMS −20
Noise reduction 6 dB
-58.1 dB
Noise reduction 9 dB
-59.7 dB
Noise reduction 12 dB
-60.9 dB
Other tries
9 dB first + compressor 3:1
-52.8 dB
No noise reduction, RMS −22
-55.5 dB
Noise reduction 6 dB first, RMS −22
-60.1 dB
Hear the chain
The opening of “The Gift of the Magi”, read by Betsie Bush for LibriVox. Headphones help.
As recorded
RMS −27.5 dB: too quiet for ACX
Turned up 7.5 dB
Floor −53.5 dB: the hiss came up too
Noise reduction 12 dB, last
Floor −58.7 dB: still fails
Noise reduction 12 dB, first
Floor −60.9 dB: passes
Noise reduction 6 dB, first, RMS −22 Master
Floor −60.1 dB: passes with the lightest touch
What the numbers show:
What I couldn’t check. I can’t hear, so I can’t tell you whether the 12 dB versions sound “wine-glassy”. I asked an audio AI model to flag artifacts in four clips. It flagged the untouched recording and cleared a deliberately overprocessed control, so I didn’t use its answers. Listen to the clips above with headphones and judge for yourself.
I made four versions from the same master and measured each final file. Each one needs a different loudness, and two needed a limiter:
| Destination and file | Change from the master | Measured |
|---|---|---|
| ACX audiobook MP3, 192 kbps CBR, 44.1 kHz, mono | None | Peak −3.7, RMS −22.4, floor −60.4 dB. Passes |
| Apple Podcasts AAC in .m4a, 96 kbps, mono | +6.1 dB, limiter at −2.6 dB | −16.4 LUFS, true peak −1.1 (target −16 ±1, ≤ −1) |
| Broadcast (EBU R 128) WAV | −1.7 dB | −23.0 LUFS, true peak −5.0 (target −23) |
| Online speech (AES TD1008) WAV | +3.6 dB, limiter at −2 dB | −18.0 LUFS, true peak −1.5 (target −18) |
The podcast version failed the first time. With the limiter at −2 dB, the WAV measured a true peak of −1.3 dBTP, inside Apple’s limit. After AAC encoding, the true peak was −0.7, outside it. Apple warns about exactly this: encoders “typically don’t modify the loudness and might clip the signal if the recommended true-peak value is not respected.” A limiter at −2.6 dB fixed it. Measure the file you upload, not the WAV you exported it from.
Check mono before you measure. The LibriVox file is stereo, but its two channels were the same voice at levels 5.7 dB apart. My first, quick mono conversion read 3 dB louder than averaging the two channels. If your meter and your export disagree, a stereo-to-mono step is the first thing to check. Record in mono and you avoid it.
Not for audiobooks on ACX. Its rules say: “Your submitted audiobook must be narrated by a human unless otherwise authorized”, and “Unauthorized use of text-to-speech, AI, or automated recordings in ACX titles is prohibited.”
For video, training and social clips, an AI voice skips the room, the mic and most of the editing above. It also lets you change a line without booking another session. The trade-off from my timing test: three stock AI voices read at 193 to 200 words a minute at default speed, about a fifth faster than the human readers. A script timed for a human runs short. Slow the voice down, or write to the fast column of the timing table.
The three AI voices in the timing test came from Lazybird, the text-to-speech tool we make. You can listen to voices free in the browser, with 10,000 credits a month. Downloads, commercial use and cloning your own voice start with the Creator plan at $72 a year (pricing, checked 21 October 2026). For other tools, see our comparison of AI voice-over generators, and for the video workflow, our guide to text to speech for video.
Set the gain so your loudest lines peak around −6 dBFS, the point where Audacity’s meter turns yellow. That leaves room for surprises. Don’t record louder to meet a loudness target. You raise the loudness in editing, then limit the peaks.
Get closer to the mic, in a soft room. RØDE says “the biggest mistake many content creators make is having the microphone too far away.” Close up, the mic hears more of you and less of the room. Editing can lower steady noise afterwards, but not echo.
For video and social, yes, if the room is quiet and soft and the phone stays at a fixed distance. On an iPhone, judge the take with Enhance Recording off, because Apple says it doesn’t change the recording itself. I didn’t test phone recordings for this guide.
Mostly, you can’t. Noise reduction removes steady noise, not echo, as the simulation shows. Re-record in a softer room, closer to the mic.
About 160 words per minute of finished audio for a calm human read. That was the measured pace of both the LibriVox narrator and the median of 107 research speakers. Write 145 words a minute if the picture is busy, and expect AI voices to read faster. See the timing table.
A voice-over sits on top of the picture. A dub replaces the actors’ dialogue, usually in another language, and has to match their lips and timing. See voice-over, narration and dubbing.
Peaks no higher than −3 dB, RMS between −23 and −18 dB, and a noise floor no higher than −60 dB RMS. Each file is a 192 kbps or higher CBR MP3 at 44.1 kHz, with 1 to 5 seconds of room tone at the start and end. The test above shows how a real recording gets there.
Apple Podcasts recommends −16 LKFS, ±1 dB, with true peaks no higher than −1 dBFS. Measure the encoded file: in my test, AAC encoding raised the true peak by up to 0.6 dB.
Not unless ACX authorizes it. Its requirements say an audiobook “must be narrated by a human unless otherwise authorized”.
Yes, with a text-to-speech tool, for video, e-learning and social content. Write and mark up the script the same way, check the pace, and check the platform’s rules first. ACX is the main place it isn’t allowed. See when an AI voice-over makes sense.

Free voice over software, tested: record your own voice or let AI read the script. Real samples, free-plan limits, watermarks and commercial-use rules.

Which Kindles and Kindle apps read aloud, how to turn on Assistive Reader and VoiceView, and why text to speech is greyed out on some books (Oct 2026).

Copy-ready phone greetings for live answer, voicemail, after-hours, holiday, IVR menus and on-hold, timed in seconds and tested on a simulated phone line.
Natural, celebrity and character voices. Or clone your own.