The biggest improvement comes before you press record. Get the microphone close to the person speaking, switch off anything that hums, and record in a room with soft surfaces. After that, choose a larger speech model instead of reaching for noise reduction software. Modern speech-to-text models are trained on messy real-world audio, and aggressive filtering often strips out the consonants they depend on. Clean up with software only when the noise is loud and steady.
Here’s how to do each of those, with specific settings for phone recordings.
Why does mic distance matter more than anything else? #
Sound from your mouth falls off quickly with distance. Each time you halve the distance between mouth and mic, your voice arrives about 6 dB louder, while the fan, traffic or café chatter in the room stays roughly the same. That ratio of voice to noise (the signal-to-noise ratio) decides how much the transcription engine has to guess.
In practice:
- Dictation and voice notes: hold the phone 15 to 30 cm (6 to 12 inches) from your mouth, slightly off to the side so your breath doesn’t hit the mic directly.
- Interviews: put the phone between you and the other person, closer to whoever speaks more quietly. A wired clip-on mic on the interviewee is even better.
- Meetings: put the phone in the middle of the table, not at one end next to the person running the call.
A cheap mic close to the mouth usually beats an expensive one across the room. Our guide to improving speech recognition with better microphones covers when upgrading the hardware is worth it.
A two-minute checklist before you record #
- Turn off steady noise sources. Fans, air conditioning, dishwashers and the TV in the next room are all easy to switch off and hard to remove later.
- Close windows facing a street or construction.
- Silence notifications. A phone buzzing on a wooden table sounds like a drum to the mic. Airplane mode stops calls and alerts, and an offline transcription app keeps working in it.
- Put the phone on something soft. A folded cloth or notebook absorbs taps, cup clinks and people leaning on the table.
- Keep the mic uncovered. Check that your case or your hand isn’t over the bottom mic opening.
- Record a 10-second test and play it back. If you can hear the fridge, the model can too.
How to fix the room itself #
Hard surfaces bounce your voice back to the mic a fraction of a second late. That smeared echo confuses speech recognition almost as much as noise does. Kitchens, bathrooms, glass meeting rooms and empty offices are the worst places to record.
- Choose rooms with carpet, curtains, sofas or full bookshelves.
- For solo dictation, a closet full of clothes is a surprisingly good vocal booth.
- If you’re stuck in a hard room, sit close to a wall hanging or put a coat over the back of your chair behind you.
- Outdoors, wind is the main enemy. Turn your back to it, shield the phone with your body, or use a foam windscreen on an external mic.
Phone-specific settings that help #
Phones are good recorders if you avoid a few traps.
- Use the built-in mic or a wired mic rather than Bluetooth earbuds. Many earbuds switch to a lower-quality call audio mode when their microphone is active, which loses detail that transcription needs.
- Don’t record through a speakerphone call if you can record the room directly instead.
- Keep the phone still. Handling noise from a phone moving in your hand is loud at close range.
- Record long sessions with the screen off, so you aren’t touching it. Check that your app keeps recording when the screen locks.
Private Transcribe is built for this kind of recording. It keeps recording when the screen locks or you switch apps, and transcribes on the phone itself with Whisper, so a noisy café recording never gets uploaded anywhere to be “cleaned.” Its model tiers matter here. The Accurate tier (190 MB) handles background noise and accents noticeably better than the smaller ones, and Professional or Best is the tier to use for a recording you can’t redo.
Does noise reduction software improve speech-to-text accuracy? #
Sometimes. It depends on the kind of noise.
| Noise type | Fix before recording | Does noise reduction help afterwards? |
|---|---|---|
| Steady hum (fan, AC, fridge, electrical buzz) | Turn it off | Often, if applied gently |
| Traffic, distant street noise | Close windows, move rooms | Sometimes |
| Other people talking (café, open office) | Move closer to the speaker | Rarely. Filters can’t tell voices apart reliably |
| Wind | Shield the mic, windscreen | Rarely |
| Echo in a hard room | Soft furnishings, closer mic | Rarely |
| Clicks, bumps, table knocks | Soft surface under the phone | Mostly no |
The reason for all those “rarely” answers is that noise reduction and speech recognition solve the same problem, deciding what is speech. A filter that works hard enough to remove café chatter usually also removes the “s,” “f” and “t” sounds the recognizer needs to tell words apart. Whisper-based models learned from hundreds of thousands of hours of imperfect audio and generally cope better with honest noise than with over-processed audio. We explain the mechanics in why offline speech-to-text accuracy drops with noise.
If you do filter, do it this way:
- Cut a one-minute sample of the worst section.
- Make a second copy with light noise reduction. In Audacity, which runs locally, use the Noise Reduction effect with a low reduction setting and a noise profile taken from a pause with no speech.
- Transcribe both copies and compare.
- Process the full file only if the filtered copy clearly wins.
Web-based enhancers such as Adobe Podcast’s Enhance Speech can sound impressive, but they work by uploading your audio. Skip them for confidential interviews, patient notes or anything you wouldn’t email to a stranger.
Choose the right model for the conditions #
When the audio is imperfect, a bigger model is the most reliable fix. Larger Whisper models have more capacity to recognize words through noise and unfamiliar accents. They’re slower, but on a recent phone or laptop the difference is minutes, not hours.
Two more settings help with noisy recordings:
- Pin the language instead of letting the model guess. Noise at the start of a file can make auto-detection pick the wrong language.
- Add custom vocabulary. Names, product terms and jargon are the first words to go wrong in noise. Private Transcribe has a custom vocabulary list in Settings, and whisper.cpp accepts the same kind of hint with
--prompt.
For recurring recordings like a weekly podcast, the specific tweaks in how to improve private speech-to-text accuracy for podcasts apply too.
What to do with a recording that’s already noisy #
You can’t recover what the mic never captured, but you can get the most out of what it did:
- Run the largest model you have before anything else.
- Trim long stretches of pure noise or silence. Speech models occasionally invent text during long gaps.
- Try gentle noise reduction on a sample, as above, only if the noise is steady.
- Re-record the key parts if you can. Reading back an important quote or figure takes a minute.
- Proofread against the audio for anything you’ll quote or act on. A transcript with timestamps makes this fast, because you can jump straight to the doubtful line.
Local and cloud engines fail in similar ways on bad audio. There’s a fuller comparison in private transcription accuracy vs online services.
Frequently asked questions #
Should I always use noise reduction before transcribing? #
No. If you can understand the recording comfortably, transcribe it as is. Noise reduction helps mainly with loud, steady noise like a fan or electrical hum. With voices, wind or echo it often lowers accuracy, so test it on a short sample first.
What’s the best microphone for transcription on a phone? #
For one speaker, a wired clip-on (lavalier) mic about 15 to 20 cm below the mouth is the biggest upgrade. For a group, the phone’s built-in mic in the middle of the table works well in a quiet room. Bluetooth earbud mics are usually the weakest option.
Can AI remove background voices from a recording? #
Voice-isolation tools can reduce background chatter, but they also remove parts of the main speaker’s voice and can create artifacts that confuse transcription. Moving the mic closer to the speaker before recording works far better than separating voices afterwards.
Does recording in airplane mode affect quality? #
It doesn’t change the audio, but it stops calls and notifications from interrupting or buzzing during the recording. With an app that transcribes on the device, you can record and transcribe entirely in airplane mode.
How quiet does the room need to be? #
Quiet enough that you can hear the speaker clearly on playback without straining. Ordinary office noise is fine for modern models if the mic is close. Problems start when background voices are nearly as loud as the person you’re recording.