/insights · VoiceLab
Test your microphone before you record: the five things that ruin a take
Level and distance, background noise, clipping, sibilance and room bleed: the five measurable reasons a voice take sounds wrong, how to check each one in twelve seconds, and why the fix is usually free.
Most bad voice recordings are not caused by a bad microphone. They are caused by five things that are all measurable in under a minute, and four of them are free to fix.
→ MicLab scores all five from twelve seconds of speech, free on Android
1. Level and distance
The single most common problem, and the cheapest fix. Too far and you record the room instead of your voice; the signal-to-noise ratio drops and every later processing step amplifies the mess. Too close and you get plosives, boom from the proximity effect, and clipping on consonants.
A useful target for a dynamic or condenser mic on speech: a hand’s width, slightly off-axis, with peaks landing well under full scale. If you can only change one thing today, change this one.
How to check: record the same sentence at three distances and compare the numbers. The difference between 5 cm and 20 cm is usually larger than the difference between a EUR 20 mic and a EUR 300 mic.
2. Background noise
Every room has a floor: fridge, fans, traffic, a computer. What matters is not the noise level but the gap between your voice and that floor. Under roughly 40 dB of separation, noise becomes audible in the quiet moments and noise reduction starts eating your consonants.
How to check: measure a few seconds of silence, then speech. The ratio is the number to improve. Turning off a fan usually beats any plugin.
3. Clipping
Digital clipping is not “a bit loud”, it is the waveform hitting the ceiling and going flat. It cannot be undone. Anything that touches 0 dBFS is damaged; a take that peaks around -6 dBFS has room to breathe and loses nothing.
How to check: look for samples at or near full scale, not at the meter’s average. A single clipped burst on a plosive is enough to make a sentence unusable.
4. Sibilance
The harsh “s” and “t” sounds that stab at you on headphones. They live between about 5 and 9 kHz and they are a combination of your voice, your mic’s presence peak, and your distance. Some microphones are simply bright.
How to check: compare the energy in the sibilance band against the rest of your voice. If it is disproportionate, move slightly off-axis, or apply a narrow cut where the peak actually is, which is not always where a preset guesses.
5. Room bleed
The reverb tail of your room, riding along with every word. It is the reason a take can be perfectly loud, perfectly clean, and still sound like a stairwell. Room bleed is also the one problem you cannot fix in post: once the reflections are in the recording, they stay.
How to check: measure how much of the decay ends up in your voice. If room echo is your worst result, the microphone is not your problem, the room is. Measure the room itself with a clap and treat the first reflection points, or record somewhere smaller and softer.
Test the microphone you actually record with
A phone can measure its own microphone, but you probably record with something else: a USB interface, an XLR mic through a recorder, a headset. Two ways to close that gap:
- Pick the input. On Android you can select the built-in mic, a wired headset, a USB interface or a Bluetooth mic, and measure that one.
- Import a WAV. Record the same passage in your own setup, export it, and analyse the file. Same checks, same score, no phone microphone in the path.
And a warning worth repeating: your phone applies automatic gain control and noise suppression to voice by default. Measured through that processing, everything looks better than it is. Any honest measurement has to request the raw, unprocessed input, and should show you what your device does to the sound before you trust the number.
What good looks like
| Check | Aim for |
|---|---|
| Level | peaks around -6 dBFS, nothing at 0 |
| Distance | consistent, a hand’s width, slightly off-axis |
| Noise | 40 dB or more between voice and floor |
| Clipping | none |
| Sibilance | present but not dominant |
| Room bleed | short decay, no audible tail |
Twelve seconds of speech is enough to measure all six. MicLab does it on-device, keeps nothing, lets you listen back to the take, and can turn its own measurement into an EQ preset for Wavelet or Equalizer APO that you hear on your own voice before you load it. Free to check on Android.
More in VoiceLab
-
AI voice infrastructure in 2026: what actually works
The honest map of voice infrastructure in 2026: TTS, ASR, voice cloning, dubbing, real-time conversion, QA tooling. What ships in production, what is overpromised, and where the unbuilt opportunities sit.
-
Voice clone detection in 2026: what works, what doesn’t, what’s coming
A working overview of how synthetic-speech detection has evolved against modern voice cloners. Spectral artefacts, prosody fingerprints, watermarking, the regulatory landscape, and the honest limits of current detectors.
-
Audio naturalness: what it is and why it is harder to measure than loudness
Audio naturalness is how human a synthesized or processed voice sounds: prosody, spectral balance, breathing, micro-timing. Why PESQ, ViSQOL and MOS fall short, and which signals you can actually measure to gate bad voice output.