My Phone Recording Was 80% Subway Noise — Then Free AI Denoising and Voice Cloning Fixed It

Last Tuesday I recorded a two-minute voice memo on my phone while waiting for the 6 train, and when I played it back at home it sounded less like a person talking and more like a washing machine full of cutlery. Tires screeching. An announcement about a delay I already knew about. Some guy two seats down watching a video on speaker. My actual voice was buried somewhere underneath, like someone shouting from the bottom of a well.

I almost deleted it. Then I got stubborn.

Here's the thing — I've been burned by "free" audio tools enough times that my expectations start at floor level. You upload a file, wait four minutes, and get back a clip with a robotic voice saying "Generated with..." over the last three seconds. Or you hit a 30-second cap and suddenly there's a credit card form. So when I decided to find out whether free AI denoising and voice cloning could actually rescue that subway recording, I went in assuming I'd waste an afternoon and write it off.

I didn't. But it wasn't all smooth either, and the rough parts are worth knowing about before you try it.

"Free" usually means watermarked, capped, or useless

I've tested maybe a dozen free AI voice generator tools over the past two years, and the pattern is depressingly consistent.

The polished ones give you 1,000 characters, then hit you with a $22/month wall the second you want to download anything. The cheap ones sound like a GPS unit from 2009 — every sentence flat, every pause in the wrong place. And a surprising number want you to "verify your business" or connect a LinkedIn account before they'll let you export your own audio.

What I actually wanted was boring: free text to speech that doesn't sound like a hostage recording, online voice cloning that runs in a browser tab, and no registration text to speech so I'm not getting newsletters until 2031.

That last part matters more than people admit. If a tool makes me create an account before I can hear a single second of output, I close the tab. Every time. I don't care how good it supposedly is.

So my bar was: no account, no ads screaming at me, no counter ticking down in the corner, and no arbitrary limit that kicks in exactly when the audio starts getting good.

Step one: throw the noise at the denoiser

I dragged the subway file into the browser's denoise module and braced for mud.

What came back was clean. Not "clean for a free tool" — clean, period. The tire screech was gone. The announcement was gone. The guy's phone speaker was gone. My voice was sitting in the middle of the mix with no hiss underneath it, and the whole thing took maybe six or seven seconds.

No login screen. No popup about upgrading to Pro. No watermark stapled onto the tail end.

That was the moment my skepticism started cracking. I've paid for noise reduction plugins that did less. I've spent forty minutes in Audacity hand-drawing silence over breaths and room hum, and this did it in the time it takes to refill a coffee.

If you want to poke at the same thing, the free voice cloning tool is where I started. Upload, wait, download. That's the entire flow.

Eight seconds of clean audio was all it needed

Next step: cloning.

I clipped eight seconds out of the cleaned-up recording — just me talking, no background noise — and fed it to the model.

Under a minute later I had a voice. Not a "close enough" approximation. It had my pauses. It had the little uptick I do at the end of questions. It even kept the breath I take before a long sentence, which is a weirdly specific thing to notice, but once you hear it you can't unhear it.

This is the part where I expected the catch. Clone your voice free almost always means "clone your voice for free, then pay to export it." Not here. No email. No phone number. No captcha gauntlet.

The output was a realistic AI voice online — I'd genuinely struggle to pick it out of a lineup next to my real recording. Which is exciting, and also a little unsettling. I'll come back to that.

Then I carried the cloned voice over to the online text-to-speech page and pasted in my script.

Feeding the script to the online TTS

Paste the text, pick the cloned voice from the dropdown, hit generate. Full voiceover in a few seconds.

Speed and pitch sit on sliders, so I nudged the pace down a notch — my cloned voice talks faster than I thought I did, which was humbling. There's a pause control too, and once you find the right setting your script stops sounding like a list of sentences and starts sounding like a person.

No code. No API key. No command line. If you've ever tried to wire up a free voice clone API by hand, you know that's an afternoon of documentation, rate limits, and a 401 error you can't explain. This was a text box and a button.

Whole pipeline — denoise, clone, generate — took under three minutes. For a two-minute recording that was destined for the trash.

Where this really earns its keep is short-form video. If you're pumping out YouTube Shorts or TikTok clips and you hate the sound of your own recorded voice — which, statistically, you do — cloning it once and generating the rest is a genuinely different workflow. Same story for e-learning modules and audiobook drafts, where you need forty minutes of narration and you have neither the time nor the vocal stamina.

If you want to go deeper on settings and edge cases, the tutorials and notes section has other people's write-ups, which is where I found the 44.1 kHz tip below.

Where it actually tripped me up

Not flawless. Three things bit me.

  • Long scripts get auto-split into chunks, and you can hear the seams. The tone shifts slightly at each break, like a new paragraph read by someone in a marginally different mood. Keep your paragraphs short and it's barely noticeable.
  • Sample rate matters more than you'd think. I uploaded a 48 kHz file, it got resampled down, and the voice came back noticeably duller — muffled, like someone threw a blanket over the mic.
  • Wifi hiccups kill the render. Two of my attempts just stopped at 80% and sat there until I refreshed.

The fixes are easy once you know them: record at 44.1 kHz mono, keep your sample clips under 30 seconds because longer ones get truncated, and don't attempt this on airport wifi.

One more thing I ran into — if you're doing free AI dubbing for a video, break your script into short paragraphs before you generate. The chunking is much less painful that way, and syncing to your footage is easier to manage.

No ads, no signup wall, no credit counter

The thing that surprised me most wasn't the audio quality. It was the absence of friction.

No banner ads. No "you have 3 free generations remaining" counter in the corner. No modal popping up after thirty seconds. No forced account creation, no email confirmation loop, no "we've sent you a magic link" that never arrives.

I downloaded the WAV and it was just... a WAV.

That's rarer than it should be. Unlimited free TTS sounds like a marketing line until you actually find a tool that doesn't meter you — and then you notice how much of your workflow was built around rationing.

Two caveats, because I'd rather you hear them from me than find out the hard way.

First, rights. Check the terms before you use a cloned voice commercially. In the US, voice is increasingly treated as a protected personal attribute — Tennessee passed the ELVIS Act, and there's been real litigation over AI voice clones in the last couple of years. Cloning your own voice for your own podcast is one thing. Cloning someone else's without written permission is a legal minefield, and honestly it should be.

Second, language. Multilingual TTS free options live on the same page, which matters if you publish to more than one audience. I ran a Spanish paragraph through with the same cloned voice and it held up — the accent wasn't native-perfect, but it was clearly intelligible, and for a free speech synthesis pass that's a fair trade.

I sent the clip to my podcast friends and nobody noticed

Last night I dropped the finished voiceover into a group chat with three friends who run podcasts. Didn't say anything about how it was made. Just the file.

First reply: "who's the new host?" Second: "did you get a booth?" Third one asked me what microphone I bought.

Nobody said AI. Nobody said synthetic. One of them asked if I'd hired a voice actor, and I felt a little bad taking so long to answer.

That's the whole review, really.

I've spent actual money on plugins that did less than a browser tab did for free. I've spent hours editing breaths out of recordings that this cleaned in seconds. And I've definitely paid for voiceover work that sounded less like me than the clone does.

Next time I record in a bad room — and I will, because my kitchen has the acoustics of a parking garage — I'm going straight to denoise, then clone, then generate. No editing marathon.

The one habit I'm keeping: record a clean ten-second sample somewhere quiet before you need it. It costs you nothing, and it's the difference between a voice clone that sounds like you and one that sounds like you shouting through a wall.