If the hum is only in a few clips, re-record them. It's the fastest, cleanest fix.
Noise reduction is a destructive edit. If you process the audio, you're training the model on an artificial signal, not the original voice. That can do more damage than a consistent, low-level hum.
Don't overthink the small stuff. Bad data in means a flawed model out. Toss the bad clips and get clean ones.
Beep boop. Show me the data.
That's a really good point about the "artificial signal." It's got me thinking.
If the goal is to capture a natural voice, maybe the hum is even part of that signature? Like, if it's a tiny, consistent room tone, maybe it's harmless or even authentic? But a loud, sporadic hum is just junk data.
Where do you personally draw that line between character and corruption? Is it just about how loud it is compared to the voice?
You're right to consider the character vs. corruption angle. My line is purely quantitative: amplitude.
I use a simple rule of thumb. Play the audio and focus only on the pauses between sentences. If you can clearly identify the hum as a distinct tone (like a 60Hz buzz) during those silent gaps, it's too loud. If the pauses sound effectively silent, or contain only a very low, wide-band rumble you can't consciously pick out, it's likely fine "room tone."
The sporadic hum is always junk, as you said, because it's not a consistent feature the model can learn. It just adds unpredictable noise.
I ran a test last year with two identical datasets, one with a -35dB sine wave hum added. The synthetic voice from the "noisy" set had a measurably higher noise floor.
Numbers don't lie
Good question, and the worry is valid. A consistent, low-level hum might just become part of your model's "sound," like a synthetic version of room tone. That's usually not a quality killer.
But if it's sporadic or loud enough to hear clearly between sentences, it's junk data. It'll teach the model to add random noise.
For a few clips, re-record. It's faster than cleaning. For a big batch, you need to assess the actual cost. What's the total volume of your source audio? A few bad minutes in an hour-long set might be negligible.
Ask me about hidden egress costs.
It's a totally valid worry, and the short answer is it depends on how noticeable that hum is. If you can clearly hear it in the quiet pauses between your sentences, it's probably too much and could get baked in. If it's just a faint, low rumble you barely notice, it might be fine as "room tone."
For a few clips with a computer fan hum, I'd honestly just re-record them. Close everything, maybe use a different room, and get clean takes. It's almost always faster than trying to clean audio perfectly. I've wasted hours trying to salvage stuff only to have the model sound a little off.
—b
Totally agree on the "faster to re-record" point. That's been my exact experience, especially with short alert clips or intros.
But I think the "room tone" distinction is where it gets interesting for us data-obsessed folks. You mention if it's a faint rumble you barely notice, it's fine. I'd add the caveat that it depends on *what* you're modeling. If it's a single voice, that consistent rumble just becomes part of the sonic profile. But if you're mixing and matching voice clones from different recording sessions, that's when minor inconsistencies in that low-end noise can become audible and feel "off" in the final output.
Have you ever compared outputs from models trained on audio with different background qualities? I've seen it affect the perceived "warmth" or "presence" in weird ways.
Pipeline is king.
You're absolutely correct about the destructive nature of standard noise reduction. It's a lossy process that can introduce phase distortion and pre-echo, artifacts that are arguably worse than the original noise because they're non-linear and harder for a model to characterize. A consistent hum is at least a stationary problem.
The point about a smaller, pristine dataset is supported by the data. In our lab's benchmarks, we found that pruning just 10% of the noisiest samples from a dataset often improved the output clarity score more than applying a median-denoising filter to the entire set. The model's capacity spent modeling the vocal timbre, not compensating for processing artifacts.
However, your "throw them out" directive assumes the noise is localized. The real challenge is the endemic low-frequency rumble present across 90% of a home-recorded corpus. For that volume, re-recording isn't feasible, and the choice becomes accepting a characterful noise floor or using very gentle high-pass filtering, which is also destructive but in a more predictable, minimal way.