How to Transcribe Audio with Background Noise (and When to Clean It First)

Short answer: transcribe the original audio first, without cleaning it: current models tolerate a fair amount of noise, and aggressive noise reduction can make the result worse. If parts of the transcript come out badly, then try cleaning the audio (Audacity or a "speech enhancer") and transcribe only that version again. Next time, move the microphone closer: it helps more than anything else.

What kind of noise do you have?

Not all problems are the same, and each has its own fix:

Problem	Examples	Impact	What to do
Constant noise	Air conditioning, fans, distant traffic	Small, if the voice is clear	Transcribe as is; clean only if it fails
Variable noise	Cafés, streets, background voices	Considerable: the model may "hear" other people's words	Clean and review carefully
Echo or reverb	Large halls, churches, empty classrooms	High: syllables blur together	Hard to fix afterwards; prevent it
Background music	Videos, podcasts with a theme, events	Variable; loud music hurts a lot	Lower the music when editing or separate the voice
Distant or quiet voice	Recorder across the room	High	Boost and clean; prevent it
Overlapping voices	Several people at once	One voice gets lost	Only avoidable by asking people to take turns

Step 1: try the original

1. Upload the file to the transcription tool.
2. Choose the language manually: with noise, automatic detection can get the first seconds wrong.
3. Press Transcribe and review the result, listening to doubtful passages in the editor.

If the text is good except for a few moments, fixing those by hand is fastest. If it fails throughout, go to step 2.

Step 2: clean the audio

• Audacity (free): select a stretch with only noise → Effect → Noise Reduction → "Get Noise Profile"; then select everything and apply a moderate reduction (6–12 dB). Push it too far and the voice sounds "metallic" and results get worse.
• AI speech enhancers: tools like Adobe Podcast's Enhance Speech (free with limits) isolate the voice from noise and echo. They work very well with one speaker; with several, check nobody gets "eaten".
• Voice/music separators: if music is the problem, stem-separation tools let you keep just the vocals.
• Normalize the volume if the voice is very quiet (Audacity → Effect → Normalize).

Export as WAV, FLAC or good-quality MP3 and transcribe that version. Compare it with the first transcript and keep the better one.

Step 3: review where it fails most

With noise, errors cluster around proper names, numbers and uncommon words. Check those especially, and mark anything you can't understand even after several listens as [inaudible 00:12:45] instead of guessing.

How to record so it doesn't happen again

• Move the microphone closer. This matters most: the phone less than a meter (3 ft) from the speaker, ideally a hand's width away for a single speaker.
• Pick the room: small, with furniture, curtains or rugs; avoid empty rooms and cafés.
• Switch off noisy things: fans, projectors, notifications.
• On video calls, use a headset and record the call audio, not the speaker.
• Do a 30-second test and listen with headphones before you start.

More tips by scenario in interview transcription, lecture transcription and meeting transcription.

FAQ

Why does the transcript make up sentences during silences or music? Speech models can "fill" stretches without clear speech with plausible-sounding text. If you see sentences nobody said during music or noise, delete them when reviewing.

Does a phone call count as noisy audio? It's narrowband audio: it transcribes well, but check names and numbers more carefully. There's a guide to call recordings.

Is it worth paying to clean the audio? Only if the audio can't be recorded again and the result matters (a key interview, evidence). Otherwise, try the original and the free options first.