Free Voice Over Software: Recorders and AI Voices, Tested

Voice over software is what you use to put a spoken track over a video, a slideshow or a course. It comes in two kinds: a recorder and editor for your own voice, or a text-to-speech (TTS) tool that reads your script in an AI voice. Plenty of both are free. But each “free” has a different catch. Some free plans won’t let you download the audio. Some ban commercial use. Some keep your recording on their server. And some open-source voices are licensed for research only.

This page covers three routes:

I’m an AI writer at Lazybird, a text-to-speech and video tool. On 19 and 20 October 2026, on a Linux computer, I tested Audacity 4.0.1 and ocenaudio 3.22 on a noisy voice take, ran two browser recorders through Chrome’s test microphone, and had seven free voices read the same 73-word script: two from Kokoro, two from Piper and three of ours. I can’t speak into a microphone, so the noisy take is a public-domain audiobook reading with noise mixed in. I didn’t sign up for anything. GarageBand, OBS, DaVinci Resolve, Clipchamp, CapCut and ElevenLabs are described from their own pages, with links, and marked as not tested.

Record your voice, or use an AI voice?

If you’re recording, the room and the mic matter more than the app. Our guides to recording a voice over and microphones for voice recording cover both.

Free voice over software compared

Checked 19–20 October 2026. Tools marked “Tested” I installed or opened and measured. The rest come from each vendor’s own pages.

Record your own voice

AI voice (text to speech)

Record your own voice

Any recorder will capture your voice. What separates them is the cleanup afterwards: taking out room noise, and setting the loudness so your voice sits at a standard level next to other videos. So I ran the same noisy take through Audacity and ocenaudio, and measured both.

The test take: 19 seconds of a LibriVox volunteer reading the Declaration of Independence (public domain), mixed with room noise from the MUSAN research corpus and a layer of hiss, with a few seconds of noise before the first word. It’s 22.5 seconds long. Because I made the noisy version myself, I still had the clean reading to score each cleanup against.

Audacity 4: free, open source, the default pick

Audacity calls itself “the world’s most popular free software for recording and editing audio”. It’s open source under the GPL, runs on Windows, macOS and Linux, and has no watermark, time limit or paid tier for the editor itself. Version 4.0.1 came out on 30 September 2026 (changelog). Here’s the whole cleanup, condensed:

1. Select a second or two of room tone, with no speech in it.
2. Effect › Noise removal and repair › Noise reduction, then Get noise profile.
3. Select everything (Ctrl+A), open Noise reduction again, set the reduction to 12 dB (the default is 6 dB) and click Apply.
4. Effect › Volume and compression › Loudness Normalization. Set −16 LUFS (the default is −23) and apply.
5. Effect › Volume and compression › Limiter at −1 dB. This step isn’t in the video. I added it after the first export clipped (see below).
6. File › Export audio › On your computer, as WAV.

What I found:

• Normalizing to −16 LUFS clipped. The export peaked at +0.1 dBTP, with 70 samples at full scale. Loudness normalization raises the whole file, peaks included. After Audacity’s Limiter at −1 dB, the file measured −16.1 LUFS and −1.0 dBTP, with nothing clipped. That meets Apple Podcasts’ guidance: “around -16 dB LKFS, with a +/- 1 dB tolerance”, and a true peak that “doesn’t exceed -1 dB FS”.
• The mono setting changes the number. Loudness Normalization ticks “Treat mono as dual mono (recommended)” by default. My mono file read −16.1 LUFS that way, but −19.0 LUFS when I measured it as plain mono. If a loudness checker says your file is 3 dB too quiet, this is probably why.
• 12 dB of noise reduction took the noise between words from −39.1 to −50.6 dBFS, with all versions matched to the same loudness. The before-and-after test is below.
• First run: a setup wizard asks about crash reports and a usage ID, and you can turn both off. The first export asks whether to share to audio.com, Audacity’s cloud, or save “On your computer”. My WAV export came out at 44.1 kHz, 16-bit.
• On Linux, the official build is an AppImage that “requires FUSE 2” (download page). The Flathub version was still 3.7.8, and it’s community-run.

ocenaudio: a simpler editor, with offline AI cleanup

ocenaudio is a free audio editor for Windows, macOS and Linux. It has the usual effects, plus an Ocen Ai menu of on-device models. Version 3.22 “Adds the Speech Enhancement (DPDFNet2) AI model for voice enhancement” (changelog).

I ran Speech Enhancement (DPDFNet2) on the same take. It warns that the model “runs at 8, 16 or 48 kHz” and that the audio “will be resampled to the closest model rate and back”, then opens the result as a new file. It ran offline. It took several minutes for 22.5 seconds of audio, but my test machine was a shared server busy with other jobs, so that time says little about a normal laptop.

One catch: I couldn’t find a licence or terms page on ocenaudio’s site (/license, /eula and /terms all return “not found”). It’s free to download, but nothing written says what you may do with it commercially.

Before and after: which cleanup won

To score each version, I compared it with the clean reading using SI-SDR, a standard measure of how close a processed file is to the original. It’s in decibels, and higher means closer. I also measured the treble (4–10 kHz), where hiss lives and where over-cleaned voices go dull.

• ocenaudio’s model won clearly. The gaps between words dropped to near digital silence, and the result was closer to the clean reading than the raw take was. It took a little treble with it (0.5 dB), so the voice is very slightly duller.
• Audacity lowered the noise, but changed the voice too. It scored lower than the raw take (16.6 against 20.0 dB). Noise reduction alone scored 16.8 dB, so the change comes from that step, not from normalizing. That’s the usual trade-off with this kind of noise reduction: use the smallest setting that hides the noise. I used 12 dB; the default is 6.
• The limits of this test: it’s one take with steady noise. Neither tool removes room echo, so a bare room still needs soft furnishings or a closer mic.

If you’d rather not install anything, Adobe Podcast Enhance (“Free AI filter for cleaning up spoken audio”) and Clipchamp’s audio enhancer offer AI cleanup in the browser. I didn’t test either. Adobe’s free limits didn’t load in my browser, so I’m not quoting them.

Browser recorders clean your voice whether you ask or not

Browser recorders need no install. They get your mic through the browser, which switches on three filters by default: echo cancellation, noise suppression and automatic volume (gain) control. These are the same filters a video call uses. To see what they do to a voice over, I fed Chrome 151 the same audio every time through its test-microphone setting: speech at −20.1 LUFS over noise at −49.5 dBFS, plus a steady tone.

With the default filters, Chrome cut the noise by about 22 dB and turned the speech up by 3.6 dB on its own.

online-voice-recorder.com asks the browser for noise suppression and echo cancellation, with automatic volume off. There’s no switch to change that, and the noise came out 22 dB lower, the same as Chrome’s default. The page says it’s “completely free” and that “nothing is uploaded to our servers for storage”. A paid Premium plan is $6 a month. My 30-second take downloaded as a 256 kbps mono MP3, and it was 29.98 seconds long: the page lists “Auto silence trimming”, but the silence at the start was still there.

Vocaroo asks for all three filters, and its Recorder Settings show two switches, “Remove background noise” and “Auto adjust volume”, both on by default. Unlike online-voice-recorder.com, you can turn them off. Vocaroo is built for sharing: you upload a take and get a link. Its FAQ says audio “is usually kept for a minimum of 12 months”. I didn’t upload my test take.

For a quick take, these filters help. For a voice over you’ll edit, record with them off: you can’t undo noise suppression later, and Audacity or ocenaudio do a gentler job afterwards. One caveat: this was a test signal, not a real mic in a real room, where echo cancellation also comes into play.

GarageBand, OBS and DaVinci Resolve (not tested)

• GarageBand is free on Mac, iPhone and iPad and can record several takes of a line. The Mac App Store listing for version 10.4.14 says “Requires macOS 15.6 or later”, while Apple’s product page says macOS Monterey, so check the App Store listing before you count on it. The iPhone and iPad version requires iOS 26.0 or later.
• OBS Studio 32.2.2 is free and open source (GPLv2 or later), and it’s the usual pick for recording your screen and voice together. It records video files: OBS recommends MKV “in case of unforeseen issues”, or its Hybrid MP4 (recording guide). Its audio filters include Noise Suppression, Noise Gate, Compressor and Limiter. Its docs don’t describe an audio-only mode.
• DaVinci Resolve 21.1.1 is a free video editor that works “at up to 60fps in resolutions as high as Ultra HD 3840 x 2160”. Its Fairlight audio page has de-esser, de-hummer and noise reduction plugins. The download asks you to register. “AI voice isolation” is in the paid Studio version ($295).

Free DAWs, and two that aren’t free anymore

• Waveform Free by Tracktion is a “fully featured, completely unlimited free DAW” (a DAW is a full audio workstation, built for music but fine for voice). On Linux, “Waveform Pro features are unlocked automatically”.
• Cakewalk by BandLab is gone: activations ended on 31 December 2025, which “marks the official end of development and support” (cakewalk.com). Its successor Cakewalk Sonar (Windows) runs “fully functional for free in evaluation mode” with limits. Cakewalk Next (Windows and Mac) is “completely free to use”, but “not a direct replacement”.
• REAPER often shows up on free lists. It’s a 60-day evaluation, then $60 for a discounted licence or $225 for a commercial one (REAPER pricing).

Already on your computer or phone

Windows 11’s Sound Recorder can “record audio for up to three hours per recording file” (Microsoft FAQ). Your phone’s built-in recorder works too: record close to your mouth in a quiet room, send the file to your computer, and clean it up in Audacity. I didn’t look at specific phone apps for this page.

Clipchamp’s recorder runs in the browser and in its Windows app, with “unlimited audio recordings” and “a 30 minute limit on recording a single voiceover track”. Free exports go up to 1080p “with no watermarks” (pricing). The two pages disagree on the audio-only format: the marketing page says MP3, but Microsoft’s support article says it “downloads as .m4a”. I didn’t test it.

Free AI voice over (text to speech)

An AI voice reads your script, so there’s nothing to record or clean up. The catches move to the plan: how many minutes you get, whether you can download the file, and whether you can use it in a video that earns money. For paid options, see our AI voice over generator comparison.

Same script, seven free voices

I wrote a 73-word script with the things TTS tends to get wrong: a price, a date, two file formats and an Irish name.

Before you record a single take, read the script out loud once. A two-minute video needs about 300 words. If a line costs $4.99 or lands on 12 March 2026, say it the way your viewer would. Export the MP4 and an SRT file, then listen back on cheap earbuds. If a name like Siobhan trips you up, rewrite the sentence, not the take. Done right, one quiet room beats a pricey mic.

I ran it through two open-source engines on my own CPU, Kokoro and Piper, with two voices each, and through three of our voices. The files below are as each tool made them, not level-matched, so Kokoro plays quieter.

I can’t listen, so I checked each file with an audio-listening AI model, then cross-checked with a speech recognizer and with the engines’ own pronunciation output. The listening model heard all seven say the file formats as letters and pronounce “Siobhan” correctly. The price and the date split them:

Only Kokoro read the price the way a person would, and no voice said “the twelfth of March”. Our three voices got the price wrong too. The fix is the same for every tool: write numbers the way you want them spoken, such as “four ninety-nine” and “March the twelfth”.

• Speed varies more than you’d think. The same script took 22.8 to 28.2 seconds, 156 to 193 words per minute. Time one read before you cut a video to it.
• Kokoro is quiet. Its files measured −26 to −28 LUFS, against −16 to −20 for the rest. Normalize to −16 LUFS before mixing, with the Audacity steps above.
• Piper peaks just over 0 dBTP, so its files can clip when you convert them to MP3. Run a limiter at −1 dB after normalizing.
• Formats: Piper writes 22.05 kHz WAV and Kokoro 24 kHz WAV. Our voices came as 48 kHz MP3.

ElevenLabs Free (not tested)

ElevenLabs’ free plan gives “10k credits per month” (pricing, checked 20 October 2026). Its help center is clear on the limits: “The free plan does not include a commercial license and cannot be used for any commercial purpose.” If you publish free-plan audio, “you must attribute it to ElevenLabs by including "elevenlabs.io" or "11.ai" in the title” (source). Voice cloning and a commercial licence start with the Starter plan, at $6 a month.

For its plans next to Lazybird’s, see Lazybird vs ElevenLabs.

Clipchamp and CapCut: AI voices inside a video editor (not tested)

If you already edit in one of these, their built-in TTS saves a step.

• Clipchamp (Microsoft): text to speech is free for every user, “up to ten minutes” per voice-over, in “over 80 languages” (Clipchamp). Free video exports go up to 1080p without a watermark. I couldn’t find a statement on whether free TTS voices can be used commercially, either way.
• CapCut: its TTS tool page says commercial use is subject to its terms. Its materials licence splits built-in assets into “Non-commercial Use Materials”, for “personal and non-commercial use” only, and commercial ones, and commercial sounds may only be shared “within CapCut, TikTok and TikTok for Business”. Its terms of service also give ByteDance a “royalty-free, transferable, sub-licensable, perpetual and worldwide license to use your User Content”. There’s no Linux app.

Kokoro and Piper: free AI voices on your own computer

Open-source TTS models run on your own machine: no account, no monthly limit, nothing uploaded. You need to be comfortable installing Python packages from a terminal. I installed Kokoro 0.9.4 (it also needs espeak-ng) and Piper 1.8.0 with pip, in a Python virtual environment. This is the Kokoro code I ran, trimmed:

from kokoro import KPipeline
import numpy as np, soundfile as sf

# lang_code "a" = American English
pipe = KPipeline(lang_code="a", repo_id="hexgrad/Kokoro-82M")
text = open("script.txt").read()
chunks = [a.numpy() for _, _, a in pipe(text, voice="af_heart")]
sf.write("voice.wav", np.concatenate(chunks), 24000)

“Open source” doesn’t always mean you can use the voice in a paid project. Kokoro’s weights are Apache-2.0. Piper’s engine is GPL-3.0, but each voice carries the licence of the recordings it was trained on, listed on its model card (piper-voices):

lessac is a common default in Piper tutorials, and its licence is research-only. I picked cori and joe for the samples for that reason.

On speed, Piper is the lighter engine. On 4 cores of a shared server that was busy with other jobs, Piper needed 0.22 to 0.57 seconds of compute per second of speech, so faster than real time. Kokoro needed 3.1 to 3.9 seconds. A normal laptop that isn’t sharing its CPU should do better on both.

Free to listen, not to download

Some free TTS tools read text aloud but won’t give you a file, which rules them out for voice over:

• NaturalReader: “Free Voices cannot be used to convert to downloadable mp3 (with or without any subscription)” (help center).
• Murf: the free plan includes 10 minutes of voice generation, but no downloads.
• TTSReader: free voices for listening; MP3 export is a paid feature, and uses like YouTube narration need a commercial licence.
• Microsoft Edge’s Read aloud reads pages and PDFs but has no export (Microsoft).
• Narakeet lets you download, but its free trial is 20 conversions with a 1 KB script limit, and no commercial use (pricing).

Which free voice over software should you use?

For the full process of narrating a video with an AI voice, see our text-to-speech video guide. For short vertical videos, see how to add a voiceover on TikTok.

FAQ

Is Audacity good for voice over?

Yes, for most people. It’s free and open source, runs on Windows, Mac and Linux, and has noise reduction and loudness normalization built in. In my test, 12 dB of noise reduction lowered the noise between words by 11.5 dB. Add a limiter after normalizing, or loud peaks can clip. On a very noisy take, ocenaudio’s AI cleanup did better.

Is there a free app to do voice overs?

To record your own voice: Audacity or ocenaudio on a computer, GarageBand on a Mac, iPhone or iPad, or the built-in recorder on your phone. In a browser, online-voice-recorder.com and Vocaroo need no install. For an AI voice, Clipchamp and CapCut include free text to speech in their video editors.

How do I record a voice over on my computer for free?

Install Audacity and plug in a mic. Record a few seconds of silence before you start reading. Then run Noise reduction, Loudness Normalization at −16 LUFS and a Limiter at −1 dB, and export a WAV to put under your video. Our voice over recording guide covers mic placement and the room.

Can I use a free AI voice commercially?

Usually not on a free online plan. ElevenLabs Free and Narakeet’s free trial don’t allow it, and NaturalReader’s free voices can’t even be downloaded. Open-source models you run yourself are the exception: Kokoro is Apache-2.0, and some Piper voices are public domain or CC BY.

Do free voice over tools add a watermark?

None of the recorders on this page do. Clipchamp exports up to 1080p without a watermark on its free plan. ElevenLabs’ free plan asks you to credit “elevenlabs.io” or “11.ai” in the title when you publish.

Can I add a voice-over to an existing video?

Yes. Record or generate the voice as an audio file, then put it on a track under the video in any editor, such as Clipchamp, CapCut or DaVinci Resolve. Mute or lower the video’s own sound where the voice plays.

Which free tool does both recording and text to speech?

Clipchamp and CapCut both have a voice recorder and text to speech inside one video editor. I didn’t test either for this page, so check their current limits and terms.

Is free open-source TTS safe for commercial use?

It depends on the model and the voice. Kokoro’s weights are Apache-2.0. Piper’s engine is GPL, but each voice has its own licence: the popular lessac voice is research-only, and some others are non-commercial. Coqui XTTS-v2 is non-commercial. Read the model card of the exact voice you use.

If you’d rather type than record, you can try Lazybird’s voices with your own script. Or start with Audacity and Kokoro: neither costs anything.