Answer · Hard audio
How do I transcribe noisy audio accurately?
Upload the original recording rather than a heavily noise-reduced copy, choose the language yourself, and add names and terms before upload. Steady noise like a fan or traffic is usually fine; music, other conversations and echo cause more errors. Next time, record with the microphone close to the speaker, ideally one microphone per person.
Try it on your own file
Free for 3 files a day, up to 30 minutes each. Speaker labels, timestamps, and exports for Word, SRT and NVivo.
Drop audio or video here
MP3, MP4, M4A, WAV, MOV, WEBM, OGG · up to 5 hours
Why noise reduction can make it worse
Noise reduction makes audio pleasant to listen to by smoothing it. That smoothing also removes small details in speech that speech recognition uses to tell similar words apart. Strong noise reduction, or the "voice isolation" setting on some apps, can lower accuracy.
Try the original first. If the result is poor, try a lightly cleaned version and compare the same passage. Keep whichever gives better text, not whichever sounds nicer.
Not all noise is equal
Constant noise such as air conditioning, traffic hum or a car engine is the easiest kind. The voice stands out against it.
Background speech, music with lyrics, clattering cups and echo in a large empty room are harder. They overlap the voice in the same frequencies.
Distance matters most. A speaker one metre from the phone in a café is much harder than a speaker 20 centimetres away in the same café.
The details
Get the microphone close
Halving the distance to the speaker does more than any software. A phone on the table next to the person beats a laptop across the room.
Separate tracks if you can
Lapel mics, a podcast recorder or a call app that records each person on their own track give each voice a clean channel.
Choose the language
Auto-detection can be thrown off by noise at the start. Picking the language yourself avoids a wrong-language transcript.
Use the names and terms box
In noisy audio, rare names are the first words to go. Listing them gives the models a hint and fixes close misspellings.
Expect gaps, not invented text
Words that cannot be heard may be missing. Check quotes against the audio before you use them.
Step by step
Upload the original file
Use the recording as it came off the phone or recorder.
Set the language and terms
Choose the spoken language and type names, places and jargon in the terms box.
Find the weak spots
Read through and note lines that don't make sense. They are usually where noise peaked.
Replay and fix
Click each weak line to play it, slow it to 0.75x, and correct it in place.
Common mistakes
- Running several noise-reduction passes before upload.
- Recording in a big echoey room when a smaller one is available.
- Putting music under a recording you want transcribed.
- Converting to a very low bitrate to save space. Upload the original; video and large files are shrunk to audio in your browser anyway.
Questions people also ask
Should I clean up the audio before uploading?
Usually not. Try the original first. Heavy noise reduction often makes text worse even when the audio sounds better. If you do clean it, compare the same passage both ways.
Can it transcribe a recording from a loud restaurant?
Often yes, if the speaker was close to the phone. Expect more errors in parts where other tables are loud or the speaker turned away.
Does speaking louder help?
Speaking clearly and close to the microphone helps more than speaking louder. Shouting distorts the sound on many phone microphones.
Is there a way to test before paying?
Yes. The free plan covers 3 files a day up to 30 minutes each, so you can check a noisy file before you buy hours.