How to clean a noisy recording before you transcribe it
To get a better transcript from a noisy recording, clean it lightly rather than as much as possible: import it into Denoise, keep AI · Auto at its default Denoise strength, tap Denoise, then give the cleaned file to your transcription app. In our test with a speech recognizer, audio cleaned at full strength lost more words than the noisy original in moderate noise, and the fewest words were lost when part of the original sound was mixed back in, which the default strength also does.
Last updated · By Routes Software, the makers of Denoise
Prepare a recording for transcription in three steps
Import the recording
Tap + › Import from Files (or Import from Gallery for a video), or share it to Denoise from Voice Memos or your recorder app.
Clean it at the default strength
Keep AI · Auto and leave Denoise strength where it starts, not at the maximum, then tap Denoise.
Transcribe the cleaned file
Tap Share Cleaned to send the file to your transcription app, or Save to Files… and import it there. If words are still missing, transcribe the original as well and compare: in moderate noise, it can do as well.
Hear the difference
Good to know
- Our test: synthetic English and German speech under office noise, transcribed by the open Parakeet speech recognizer (NVIDIA), with the noise removed by DPDFNet, the model behind AI · Auto on iPhone and iPad. With the speech 5 dB quieter than the noise, the recognizer got 36 % of the English words wrong in the fully cleaned audio, 20 % in the noisy original, and 6 % when 30 % of the original was mixed back into the cleaned audio.
- In very loud noise, cleaning helped: with the speech 10 dB quieter than the noise, 56 % of the English words were wrong in the cleaned audio and 76 % in the original. Summed over all eight cases (four noise levels, two languages), the mix lost the fewest words, and any share of the original from 15 % to 50 % did about equally well.
- At its default position, Denoise strength keeps roughly 35 to 40 % of the original sound mixed in, inside the range that did best.
- The test used synthetic voices, one kind of noise and one recognizer. Apple's, Google's and Whisper-based transcription weren't measured, so compare both versions on your own recording when the words matter.
- Don't use AI · Sidon for a transcript: it regenerates the voice and, when too much of it is missing, can produce sounds that were never said, which a transcription app would write down.
Frequently asked questions
Does removing background noise improve transcription?
Often, but not always. In very loud noise, cleaning helped the speech recognizer in our test; in moderate noise, fully cleaned audio lost more words than the original, because the clean-up's own artefacts confused it. A lighter clean that keeps part of the original did best overall.
Which transcription app should I use?
Any app or service that accepts an audio file. Denoise itself doesn't make transcripts on iPhone or Android.
Which format should I save in?
M4A, the default, works with most transcription apps. Choose WAV if the app asks for uncompressed audio.
Is it free?
Importing, cleaning and listening are free, with no account. Saving or sharing the result is paid: 3 exports cost $1.99, and Denoise Pro $2.99 a month or $14.99 once ($11.99 on Google Play). These are US prices from October 2026; the store shows yours before you pay.
Is my recording uploaded?
Not by Denoise: the AI model runs on your phone. Your transcription app may upload the file it receives, so check its own privacy policy.