Noise removal has a built-in tradeoff: remove too little and the distraction remains; remove too much and the speaker can sound robotic or incomplete. A trustworthy demo should make that tradeoff audible instead of hiding it behind a marketing score.
Our public review method centers on paired listening. We compare the same moment before and after processing, document where the source came from, and state what the example can—and cannot—prove.
Judge the voice and the noise together
Switch versions without changing the listening task. Pay attention to intelligibility, residual room sound, consonants, pauses, and whether the speaker still sounds natural.
Source: Real Estate Coaching 2023 Kickoff Zoom Call · CC BY 3.0, edited from the original
What this testing is—and is not
The comparisons on this site are product demonstrations. They help visitors hear a workflow and decide whether it is worth testing on their own material.
They are not an independent laboratory benchmark, a universal quality score, or a promise that every microphone, speaker, language, and noise condition will behave the same way.
Our five listening checks
1. Speech intelligibility
Are the words easier to understand? We listen especially to quiet consonants, word endings, and moments where the noise overlaps speech.
2. Residual noise
Has the distracting sound been reduced enough for the intended use? Total silence is not the target. A small amount of stable ambience can sound more natural than a voice surrounded by digital voids.
3. Voice character
Does the speaker still sound like the same person? We listen for metallic resonance, watery movement, thinning, unstable pitch, and exaggerated sibilance.
4. Transitions and pauses
Processing artifacts often become obvious when speech starts or stops. We check breaths, pauses, and the transition between noisy and quiet sections.
5. Workflow reliability
A good output also needs to be usable. We check that the task completes, the result can be previewed and downloaded, and—in video tests—that the soundtrack remains synchronized with the picture.
How we prepare a comparison
For each public sample, the process should record:
- the source and its usage rights;
- the selected excerpt and duration;
- the original file format;
- the published processed result;
- any trimming, format conversion, or other edit;
- the publication or review date.
The original and cleaned controls should point to the corresponding versions of the same excerpt. We do not want a louder, shorter, or selectively edited result to win by presentation rather than cleanup quality.
Why several types of recordings matter
A podcast, phone recording, video soundtrack, and remote call do not present the same problem. Testing only one clean studio speaker would say little about everyday uploads.
The public demo set therefore includes different recording contexts. It is still a small demonstration set, not complete coverage of every language, voice, microphone, codec, or acoustic environment.
How to run your own useful test
The most relevant evaluation uses your recording and your publishing requirements:
- choose 20–30 seconds containing both typical speech and the main noise;
- listen on headphones and the device your audience commonly uses;
- compare at similar playback volume;
- review the hardest overlap, not only a quiet pause;
- keep the original file;
- do not publish the result until you have reviewed the full output.
How we will improve the evidence
The useful next step is broader, repeatable coverage: more noise types, more microphones, multiple languages, and documented review conditions. When we publish measurements, they should include the test set and method rather than appear as an isolated percentage.
Until then, the before-and-after control remains the clearest evidence on the page: the same listener, the same excerpt, and an explicit invitation to decide with your own ears.
Common questions
Frequently asked questions
Are the public demos independent scientific benchmarks?
No. They are transparent product demonstrations using paired excerpts. We publish sources and limitations so visitors can judge the audible result directly.
Why not publish one quality score?
A single number can hide the tradeoff between reducing noise and preserving natural speech. We review intelligibility, residual noise, voice character, transitions, and processing reliability separately.
Do you edit the cleaned sample after processing?
Published comparisons should disclose any additional edits. The goal is to compare aligned source and result clips, not to secretly improve only one side.
What should I listen for in a denoise comparison?
Listen for easier-to-understand words, lower distracting noise, natural consonants and breaths, stable word endings, and the absence of metallic or watery artifacts.