Does AI Noise Removal Affect Voice Quality? We Measured 56,993 Files

All three of our audio cleanup tools run DeepFilterNet3, an open-source AI noise model. When a job finishes, the server compares the input audio with the output over the same 20-millisecond slices, called frames, and takes two readings:

Background change. How much quieter the quietest fifth of the input's frames became. Those frames are the pauses, where the noise sits with little speech on top of it.

Voice change. The median change across the loudest half of the input's frames. That is where the speech is.

Both are in decibels (dB), a ratio scale. A 10 dB drop leaves one tenth of the sound energy and is usually heard as about half as loud. A 20 dB drop leaves one hundredth, about a quarter as loud. Files under about 4 seconds get no reading, and long files are sampled in up to six 30-second windows spread across the file.

What counts as speech damage. A voice change worse than minus 10 dB. That line has two limits, and both matter. On our labelled test clips, it caught 5 of the 8 real failures and every file it flagged was a real failure, so on clean test material it undercounts. On real files it can also fire when the noise was louder than the voice: then the loudest frames are noise, and removing them is the job. The counters cannot tell those cases apart. Read 28.5% as "the loud part of the audio lost more than 10 dB", not as "28.5% of files were ruined".

The most common result was 12 to 18 dB, in 41.0% of files. 61.1% of files had their pauses cut by more than 12 dB, which sounds less than half as loud. Only 4.7% changed by 3 dB or less, which is close to no audible change.

How far the model may go depends on the setting. On Remove Background Noise, the default, Balanced, limits how far any frequency can be pulled down to 20 dB. Light limits it to 12 dB, and Maximum has no limit.

Across the 55,067 files with a voice reading:

• 34.1% dropped by 1.5 dB or less, or got louder, and 46.8% by 3 dB or less.
• 24.7% dropped by 3 to 10 dB. A drop in this range often means noise was removed from underneath the speech, which is the point of the tool, rather than speech being removed.
• 28.5%, or 15,690 files, dropped by more than 10 dB. That includes 5.8% that dropped by more than 20 dB.

For the reasons above, 28.5% does not mean that many files were ruined. It does mean you should listen before you publish.

Production counters cannot say what kind of noise was in a file, so we also ran a controlled test. We mixed clean read speech from the public LibriSpeech corpus with seven kinds of noise at four loudness levels, 112 runs in all, and measured each result against the clean original. At the Balanced setting, with the noise 5 dB below the voice:

• Hiss: 19.5 dB removed, no speech frames destroyed.
• Keyboard clicks: 19.7 dB removed, 0.9% of speech frames destroyed.
• Background music: 18.3 dB removed, 0.7% destroyed.
• Hum and street noise: 6 to 10 dB removed, none destroyed.
• Crowd babble: 13.1 dB removed, 24.8% destroyed.
• A second person talking: 0.5 dB removed, 43.7% destroyed.

The pattern is clear. Steady, non-speech noise comes out cleanly. A second voice barely comes out at all, crowd babble costs about a quarter of the speech frames, and in both cases the speech you want to keep suffers. DeepFilterNet3 was never trained on audio with a second voice mixed in, so it has no way to decide which voice to keep.

1. Compare before you publish. Play one loud spoken line and one pause in the cleaned file, then the same moments in your original. Words with gaps in them, or words that fade out mid-sentence, are the damage described above. Keep your original: we delete both files about 90 minutes after processing.
2. Run it again on Light. Light limits the reduction to 12 dB per frequency. In our second-voice test, that cut the share of speech frames destroyed from 43.7% to 22.1%. You will hear more of the background in exchange.
3. If the background is a person or a crowd, expect little help from any setting. That is a limit of the model, not of your file. Re-record if you can. If you cannot, use Light and accept some background.
4. If the noise is steady and still audible, try Maximum. On hiss, Maximum removed 31.0 dB instead of 19.5 dB, at the cost of 5.4% of speech frames.
5. If the voice is clean but too quiet or uneven, use Repair Audio. Repair Audio runs noise removal with no limit, like Maximum, and also normalizes the volume to -16 LUFS. If you want a gentler blend instead, Enhance Speech has its own strength setting.

Anonymous counters only. Each finished job adds one to a fixed set of running totals: tool, length bucket, audio or video, setting, background bucket, voice bucket, and the damage flag. No file, file name, account, IP address or job ID is stored with them, so no total can be traced back to an upload.

Window. Readings started shortly after 19:21 UTC on September 18, 2026, when the measurement was released. The totals here were read on November 4, 2026 at 17:51 UTC. 56,993 jobs carried a reading. 54,145 had a usable background reading and 55,067 a usable voice reading; the rest had none. The flag uses the exact voice reading and the buckets a rounded one, so the buckets past 10 dB sum to 15,687, three short of the flag.

What is pooled. The readings pool Remove Background Noise, Repair Audio and Enhance Speech, and every strength setting. Remove Background Noise produced 13,751 of the 15,690 flagged files (87.6%). The counters do not record each tool's share of measured jobs, so we cannot give a damage rate per tool.

Our own testing. Because the counters record no account, our own tests are included. We estimate our automated tests ran at most about 670 of these jobs, assuming three per backend commit, each on a short clip of clean synthetic speech. Those clips can only lower the flagged share, and removing all of them would raise it from 28.5% to 28.8%. We also test by hand, and those uploads look like anyone else's. Allowing for 500 of them either way, the flagged share stays between 27.8% and 29.1%.

The aggregate counts are available as a CSV download and licensed CC BY 4.0. Feel free to reuse them with credit to VidClean. Questions about the method are welcome at hello@vidclean.net.