Fan
Noise drops, voice less natural- Overall
- 1.54 → 2.87
- Background
- 1.64 → 4.08
- Naturalness
- 0.92 → 2.85
These are 24 real recordings with noise behind a voice: fans, traffic, a dog, a baby, breathing, kitchen sounds, and typing. Each one is shown as it was recorded and after EraseNoise removed the noise. Across all 24, the background score rose from 2.45 to 4.13 out of 5, overall quality from 2.14 to 3.18, and naturalness from 1.98 to 3.90. Press play on any card, then switch between the original and the cleaned version.
Last updated
Steady noise from a fan, an air conditioner, a dishwasher, or a copier.
Road noise, a passing car, and street wash under a voice.
A dog, a baby, or other people in the room while someone talks.
Breathing, chewing, and crunching close to the microphone.
Water, dishes, scrubbing, and doors in a home recording.
Typing, mouse clicks, and a creaking chair next to the microphone.
Scores run from 1 to 5, and higher is better. Each cell shows the original score, then the score after cleaning.
| Recording | Noise group | Overall | Background | Naturalness | Result |
|---|---|---|---|---|---|
| Keyboard typing | Desk | 2.63 → 2.70 | 3.83 → 3.87 | 2.14 → 2.04 | Little change |
| Fan | Fans and machines | 1.54 → 2.87 | 1.64 → 4.08 | 0.92 → 2.85 | Noise drops, voice less natural |
| Car | Traffic and cars | 2.47 → 2.67 | 3.67 → 4.02 | 2.98 → 3.87 | Clear gain |
| Baby | People and pets | 2.10 → 3.27 | 2.26 → 4.04 | 3.37 → 4.70 | Clear gain |
| Dog barking | People and pets | 1.15 → 3.38 | 1.23 → 4.20 | 1.77 → 4.79 | Clear gain |
| Kitchen | Kitchen and home | 2.90 → 3.30 | 3.59 → 4.16 | 2.76 → 4.66 | Clear gain |
| Street | Traffic and cars | 1.62 → 3.28 | 1.68 → 4.17 | 1.04 → 4.32 | Clear gain |
| Door | Kitchen and home | 1.82 → 2.80 | 2.17 → 4.02 | 0.95 → 4.05 | Clear gain |
| Eating chips | Breath and mouth | 2.24 → 3.12 | 2.21 → 4.07 | 1.86 → 3.94 | Clear gain |
| Breathing | Breath and mouth | 2.35 → 3.61 | 2.15 → 4.25 | 3.03 → 4.90 | Clear gain |
| Dishwasher | Fans and machines | 2.57 → 3.14 | 2.89 → 4.14 | 1.94 → 4.18 | Clear gain |
| Air conditioner | Fans and machines | 2.99 → 3.37 | 3.70 → 4.20 | 3.48 → 4.55 | Clear gain |
| Running water | Kitchen and home | 1.10 → 2.91 | 1.14 → 4.11 | 1.13 → 2.51 | Little change |
| Mouse clicks | Desk | 3.13 → 3.43 | 3.94 → 4.22 | 2.72 → 4.04 | Clear gain |
| Chair | Desk | 2.88 → 3.44 | 3.50 → 4.22 | 3.01 → 4.59 | Clear gain |
| Munching | Breath and mouth | 2.62 → 3.39 | 2.73 → 4.23 | 2.39 → 4.62 | Clear gain |
| Copier | Fans and machines | 2.50 → 3.07 | 2.81 → 4.10 | 2.23 → 3.03 | Noise drops, voice less natural |
| Traffic | Traffic and cars | 2.22 → 3.33 | 2.10 → 4.17 | 2.14 → 4.09 | Clear gain |
| Utensils | Kitchen and home | 2.01 → 3.28 | 1.96 → 4.13 | 1.84 → 4.29 | Clear gain |
| Scrubbing | Kitchen and home | 1.10 → 3.41 | 1.17 → 4.20 | 0.74 → 4.54 | Clear gain |
| Laughing | People and pets | 1.72 → 2.86 | 1.67 → 4.03 | 1.22 → 2.94 | Noise drops, voice less natural |
| Yelling | People and pets | 2.16 → 2.79 | 3.29 → 4.05 | 1.37 → 2.44 | Little change |
| Distant fan | Fans and machines | 1.52 → 3.56 | 1.48 → 4.24 | 0.82 → 4.56 | Clear gain |
| Car close by | Traffic and cars | 2.13 → 3.35 | 1.97 → 4.11 | 1.69 → 3.07 | Noise drops, voice less natural |
| Average | 2.14 → 3.18 | 2.45 → 4.13 | 1.98 → 3.90 |
Not every recording improved the same way. These are the cases to check closely on your own file.
Overall quality barely moved (2.63 → 2.70), and naturalness fell slightly (2.14 → 2.04). Keystrokes that land on a word can stay.
Another voice is speech too. In the yelling and laughing samples the background dropped, but naturalness ended at 2.44 and 2.94.
The noise dropped a lot (background 1.14 → 4.11), but the voice ended at 2.51 for naturalness and can sound processed.
The fan, the copier, and the car close by all dropped, but naturalness ended near 3. Listen for a thin voice.
All 24 come from the real recordings in the Microsoft DNS Challenge test set: people talking with real noise around them. There is no clean studio version to compare against, which is why the scores below need no reference.
Each file was processed with the same speech model EraseNoise uses for uploads, at the reference settings: 5-second segments, 50% overlap, full precision. The site runs a faster setting that scored the same in our tests. Nothing was edited by hand.
Overall and Background are DNSMOS scores, which predict how listeners rate overall quality and how clean the background sounds. Naturalness is the NISQA score, which predicts how natural the voice sounds. All three run from 1 to 5.
Little change means overall quality rose by less than 0.1, or naturalness ended under 2.6. Noise drops, voice less natural means naturalness ended under 3.1. Every other recording is labeled Clear gain.
Source: Microsoft DNS Challenge
Yes. They are real recordings from the Microsoft DNS Challenge test set, not mixed in a studio. EraseNoise processed them without hand edits.
Across these 24 recordings, the DNSMOS background score rose from 2.45 to 4.13 out of 5 on average, and overall quality from 2.14 to 3.18. The biggest gains were scrubbing, a dog, and a distant fan. Typing under a word barely changed.
Typing on top of the words, other people shouting or laughing, and running water. In those samples the noise dropped less, or the voice came out less natural.
It depends on how loud the noise is and how close it is to the mic. Upload your own file and hear a free 30-second preview before you pay.
Upload a file and compare a free 30-second preview, original and cleaned, before you pay.
Upload your recording