Updated October 2026 · Speech recognition and on-device transcription
Background noise makes speech recognition worse because the model has to separate the voice from the noise before it can read the words. Steady noise, overlapping voices and echo each cause different errors. Improve the recording with a closer microphone and a quieter room, and clean lightly before transcribing.
Ask a speech recognition tool to transcribe a clean studio recording and the result is usually strong. Ask it to transcribe a recording made in a busy café, on a windy street or in a room with an air conditioner, and the errors multiply. The reason is that the model has to separate the speech from everything else before it can read the words, and the separation is never perfect.
This post explains how different kinds of noise cause different errors, what you can do to improve the recording and when cleaning helps more than it hurts. The advice applies to automatic transcripts, live captions and any tool that works from sound.
A steady sound, such as a fan, a fridge or a hum, is one of the easier problems for a model to handle, because it stays the same from moment to moment. Even so, steady noise lowers the clarity of the speech, and quieter consonants become harder to hear. The model may drop soft sounds or confuse similar words, especially in the quiet parts of the recording.
Reducing steady noise at the source, by switching off the fan or moving the microphone, is the most effective fix. A light cleanup afterwards can help further, but a heavy one can remove the speech along with the noise.
When two or more people talk at once, the model has to decide which words belong to which speaker, and it often fails. Crosstalk produces errors that are harder to fix than steady noise, because the sounds genuinely overlap. A single microphone shared by several speakers makes the problem worse.
The best fix is one microphone per speaker, placed at a steady distance. Where that is not possible, ask speakers to take turns, and expect more correction in the transcript.
Echo smears the speech over time, so each sound lingers and blurs into the next. A model can struggle to decide where one word ends and another begins. Echo is often caused by a hard room, and it can be reduced by moving the microphone closer and adding soft surfaces around the speaker. The clap test helps you find a room that sounds dry.
A door slam, a cough or a passing siren can cause a short burst of errors. The model may misread the words around the sound, or insert text where there was none. Note these moments when you check the transcript, and consider re-recording a short section if the sound covered an important line.
The further the microphone is from the speaker, the more room sound and noise it picks up relative to the voice. A quiet speaker at a distance can be the worst case. Move the microphone closer, and check the level so the voice is clear without clipping. A good level helps more than most cleanup.
Light cleanup can improve a transcript when the noise is steady and the voice is clear. Heavy cleanup can make speech watery and remove the soft sounds that the model needs. Compare a transcript made from the original with one made from the cleaned version, and keep the one that is more accurate. Do not assume that cleaner audio always gives a better transcript.
Before a long interview or a live session, record a short test in the actual room with the actual microphones. Transcribe it, read the result, and look for the errors that matter most. If the test transcript is poor, the setup is the problem, and no tool will fix it later. Change the position, the level or the room, and repeat the test until the transcript is usable.
Keep a note of the setup that worked. Recording the microphone, the distance, the room and the level gives you a template for the next session.
Noise will always make transcription harder, and no tool can fully overcome a poor recording. Results vary between models, languages and accents. Test on your own material, and treat the transcript as a draft that needs checking.
A clean signal is the best help a model can have. The Vocal Checker shows clipping, noise and level problems in a recording you own. The Noise-Free Voice Recorder can clean background noise from a recording. For how transcription errors are measured, read the word error rate post in this series.
Record voice-overs with live noise removal, or clean an existing file. Runs in your browser, nothing is uploaded.
Not always. Heavy cleanup can make speech harder to recognise. Compare the results.
Overlapping voices and echo are often the hardest, because the sounds genuinely mix.
If the noise covered an important line, yes. Re-record a short section rather than trying to fix it.
Yes. A microphone close to the speaker gives a much clearer signal than one at a distance.
