At Loudness.ai, we tested freshly created speech examples with several kinds of added noise. The examples below let you compare the noisy recording, the cleaned output, and the clean speech reference.
The one-minute example processed in 32.3 seconds, about 1.9 times faster than its playback duration. The shorter examples finished in approximately seven to nine seconds.
Hear the before and after examples
These controlled tests use new synthesized narration and generated background noise. Each pair contains the same words and has the same duration. Keep your playback volume consistent as you switch between the noisy and cleaned versions.
Simulated fan noise
A steady wash of simulated ventilation noise underneath a garden-planning narration. Listen to the spaces between phrases and the softer endings of words.
Electrical hum and hiss
Electrical hum and background hiss beneath a spoken meeting reminder. Listen to how the low background sound changes while following the full sentence.
Changing rumble and clicks
Changing low rumble and intermittent clicks beneath a short travel story. Follow the words through the changing background sounds.
A complete one-minute example
The longer test follows a newly written narration about a visit to a community garden, with a mixture of steady and changing background noise. The file includes speech, pauses, and background sound at the beginning and end. The narration plays through once.
How quickly the recordings processed
We measured processing after initial setup, with upload time outside the measurement. The files were processed in full, including their pauses and background noise.
| Example | Audio duration | Processing time |
|---|---|---|
| Simulated fan noise | 14.3 seconds | 7.9 seconds |
| Electrical hum and hiss | 16.7 seconds | 6.8 seconds |
| Changing rumble and clicks | 17.2 seconds | 8.6 seconds |
| Mixed noise with longer narration | 60.0 seconds | 32.3 seconds |
A 60-second file takes one minute to play. In this test, its processing finished in 32.3 seconds. The cleaned recording keeps its original pace; faster processing shortens the wait for the output.
Check the voice as well as the background
A quieter background is one part of a useful result. The words should remain easy to follow, including quiet phrases, consonants, and the ends of sentences.
For these examples, we kept a clean speech reference before adding the noise. That gives you a third version to compare with the noisy input and the cleaned result. You can listen for how closely the processed voice follows the reference, as well as how much background sound remains.
The reference-based intelligibility score increased for all four processed examples. For the three shorter recordings, the scores changed from 0.824 to 0.961, 0.965 to 0.989, and 0.805 to 0.948. Higher scores indicate better estimated intelligibility on this measure. Use the audio comparisons to judge the voice and background sound for yourself.
When reviewing your own recording, keep the original file and compare both versions at the same playback volume. Listen through a complete sentence and its surrounding pauses. Check the portions you intend to use in your final edit.
Make the next edit easier
A shorter cleanup wait helps you move from recording to the decisions that shape an edit: choosing a useful quote, placing a voiceover against a video, or preparing a lesson for an audience.
Clean the spoken track before adding background music so you can judge the voice clearly. Keep a copy of the original recording as you work, and make a short microphone test before a longer recording session.
To work on your next voice recording, open the Loudness.ai Audio Enhancer. Choose Noise Removal for unwanted background sound or Voice Enhance for clearer speech and more even levels. Upload your recording, compare the result, and download the cleaned file when you are happy with the voice.
Comparison with Loudness.ai Audio Enhancer
We processed the same four noisy recordings with both modes in the Loudness.ai Audio Enhancer: Noise Removal and Voice Enhance. The players below let you compare those outputs with the cleaned examples in this article. Playback files keep their resulting output levels.
Simulated fan noise
Hum and hiss
Changing rumble and clicks
One-minute mixed noise
Observed processing times
Each option used its tested settings. These file-processing measurements exclude upload and queue time.
| Recording | Tested cleanup | Noise Removal | Voice Enhance |
|---|---|---|---|
| Simulated fan noise | 7.95 s | 2.63 s | 1.94 s |
| Hum and hiss | 6.79 s | 1.27 s | 1.90 s |
| Changing rumble and clicks | 8.61 s | 1.40 s | 1.98 s |
| One-minute mixed noise | 32.27 s | 2.57 s | 4.02 s |
For the one-minute file, Noise Removal finished in 2.57 seconds and Voice Enhance in 4.02 seconds. The tested cleanup took 32.27 seconds.
Audio analyzer measurements
We measured each complete exported file for integrated loudness, true peak, and RMS level. We also measured the same speech-free opening from 0.00 to 0.25 seconds in every version to compare remaining background sound.
| Recording | Version | Integrated LUFS | True peak (dBTP) | RMS (dBFS) | Opening noise RMS (dBFS) |
|---|---|---|---|---|---|
| Simulated fan noise | Noisy input | -21.37 | -3.63 | -21.32 | -27.32 |
| Simulated fan noise | Tested cleanup | -22.20 | -4.20 | -22.31 | -78.41 |
| Simulated fan noise | Loudness.ai Noise Removal | -23.24 | -5.50 | -23.39 | -77.44 |
| Simulated fan noise | Loudness.ai Voice Enhance | -16.27 | -1.37 | -16.51 | -68.93 |
| Hum and hiss | Noisy input | -22.08 | -4.46 | -21.30 | -27.55 |
| Hum and hiss | Tested cleanup | -22.12 | -5.23 | -22.33 | -73.18 |
| Hum and hiss | Loudness.ai Noise Removal | -22.65 | -5.39 | -22.85 | -81.08 |
| Hum and hiss | Loudness.ai Voice Enhance | -15.97 | -1.19 | -16.45 | -66.61 |
| Changing rumble and clicks | Noisy input | -20.05 | -1.95 | -19.47 | -33.17 |
| Changing rumble and clicks | Tested cleanup | -22.14 | -4.21 | -22.38 | -68.90 |
| Changing rumble and clicks | Loudness.ai Noise Removal | -23.49 | -4.99 | -24.01 | -66.98 |
| Changing rumble and clicks | Loudness.ai Voice Enhance | -16.24 | -0.87 | -16.78 | -44.88 |
| One-minute mixed noise | Noisy input | -21.42 | -2.88 | -21.29 | -28.12 |
| One-minute mixed noise | Tested cleanup | -21.48 | -3.36 | -22.28 | -77.28 |
| One-minute mixed noise | Loudness.ai Noise Removal | -22.30 | -3.89 | -23.16 | -76.16 |
| One-minute mixed noise | Loudness.ai Voice Enhance | -16.09 | -1.23 | -17.03 | -66.64 |
Integrated LUFS describes overall loudness; a value closer to zero is louder. True peak shows the largest reconstructed peak, and RMS describes signal level. A more negative opening-noise reading means less signal energy in that particular speech-free section. The Loudness.ai LUFS Meter provides loudness and peak readings for your own exports.
Reference-based intelligibility comparison
We compared each version with its clean speech reference using STOI, an estimated intelligibility score. Higher values indicate better estimated intelligibility on this measure.
| Recording | Noisy input | Tested cleanup | Noise Removal | Voice Enhance |
|---|---|---|---|---|
| Simulated fan noise | 0.824 | 0.961 | 0.936 | 0.926 |
| Hum and hiss | 0.965 | 0.989 | 0.977 | 0.966 |
| Changing rumble and clicks | 0.805 | 0.948 | 0.903 | 0.897 |
| One-minute mixed noise | 0.807 | 0.954 | 0.921 | 0.914 |
The tested cleanup had the highest STOI score on each of these four inputs. For the hum example, Noise Removal produced the quietest measured opening among the processed versions, at -81.08 dBFS. Voice Enhance produced louder, levelled recordings, measuring between -16.27 and -15.97 LUFS.
These readings describe different parts of the result: background residue, estimated intelligibility, and playback level. Voice Enhance intentionally changes level and tone, so its loudness readings and the audio comparisons are useful alongside the reference scores. You can download the full measurements, including reference-fidelity scores, to examine the results in more detail.