BEATS WITH KEVHOW IT WORKS
What gets measured, how it’s graded, what to fix first, and the real DSP running underneath. In plain language.
WHY THIS EXISTS
Ears get less reliable over time. After a few hours on a mix, ears adjust and stop being a reliable judge. Signal Check runs the same loudness, dynamics, stereo, frequency, and true-peak checks a mastering engineer would reach for by hand, so a read doesn’t depend on how tired the ears doing the listening are.
THE 5 READS
Five reads cover a mix. Each one has a headline number and a second, closer look underneath it.
- Loudness. Integrated level in LUFS against streaming and genre targets, so the number means something instead of just being big or small. The loudest section gets its own reading, because that’s the part a limiter and a streaming platform react to first.
- Dynamics. How much a mix breathes versus how flat it’s been pushed. Crest factor covers the whole track. PSR covers the loud parts, which is where dynamics get crushed.
- Stereo. Width and mono compatibility, so a wide master doesn’t collapse or phase out in mono playback. The low end is checked separately, since bass spread wide in stereo is bass that thins out on a phone speaker.
- Frequency. Band-by-band tonal balance against a genre tilt: too much bass, not enough top, or graded as balanced. Resonant harshness in the 2 kHz to 5 kHz range gets called out on its own, with the frequency of the worst offender.
- True peak. Clipping risk, and how often a track rides right at the ceiling. You get the count of clipped samples, not just a yes or no.
GRADED IN CONTEXT
A rough mix and a finished master get held to different numbers. Before the grade lands, you say what the track is right now: rough mix, mixed, mastered, or reference. The loudness targets and the true-peak ceiling both move with that choice, so a mix in progress gets graded as a mix in progress.
Reference turns grading off. It describes what a track is doing instead of scoring it, which is how you read a release you admire without holding it to your own targets.
THE DIAGNOSIS
Every analysis ends with two things: a shape and a sentence.
- Five axes, the same five reads: loudness, balance, stereo, dynamics, peaks.
- A bigger shape means fewer problems. A point turns red where a reading needs a fix, so one look tells you where the track is weak.
- Every takeaway gets ranked by how far off it is. The worst one gets pulled to the top of the summary and written as one plain sentence.
- A track with six things flagged still has one place to start.
SHOW THE RECEIPTS
- Open the receipts and every measured value sits next to the reference line that graded it: integrated loudness, the loudest section, true peak, clipped samples, PSR, crest factor, and stereo or mono.
- You get to see why a reading was called red, rather than taking the grade on faith.
- Change something, bounce it, run it again. Signal Check keeps your last check and shows the before and after: old value, arrow, new value, for every number that moved.
- That’s how you find out whether the fix landed, before you commit to it.
THE SCIENCE
- Gated integrated loudness in LUFS, K-weighted per
ITU-R BS.1770-4, the same standard streaming platforms use for normalization. - Short-term loudness runs on a 3-second window with a 100 ms hop. The loudest section is the 95th percentile of those windows, so a single freak moment can’t define it.
- Graded against genre-specific target ranges, not one fixed number for every track.
- Crest factor:
20 × log10(true peak ÷ RMS), in dB. A wide gap means the mix still breathes; a narrow one means it’s been pushed hard. - PSR is the gap between the peaks and the short-term loudness around them, taken as the median across the loudest quarter of the track. It catches a squashed drop that a quiet intro would otherwise hide in the whole-track average.
- Per-band correlation from the cross-spectrum of the left and right channels, not just one overall mono-sum check.
- Width and mono-compatibility loss are both derived from that same per-band correlation data.
- The low band gets its own mono-compatibility read, because that’s the range where a wide image costs you the most on a single speaker.
- Hann-windowed FFT, analyzed across the full track and averaged, not a single snapshot.
- Balance is graded against a genre-specific tonal tilt, not one flat curve every genre gets held to.
- Harshness comes from hunting resonant peaks in the 2 kHz to 5 kHz band and reporting the worst one by frequency.
- True peak isn’t the loudest sample. It’s the loudest point the waveform reaches between samples, which is what actually clips a D/A converter or codec on playback.
- Found via 4x-oversampled inter-sample peak detection with windowed-sinc interpolation, the method defined in
ITU-R BS.1770-4.
STAYS LOCAL
Nothing leaves your browser. No upload. No account. No server ever touches the audio. Every measurement runs locally, and the one previous check kept for the before-and-after comparison is stored in your own browser too.
EXPORT THE BREAKDOWN
Every analysis can export as a 10-slide carousel: the numbers, the verdict, and a plain read of what’s happening in the mix, ready to post or send to a collaborator.
WHO IT’S FOR
For producers finishing a mix without a mastering engineer on call, checking a bounce before it goes to a label or playlist curator, or just wanting a second opinion that isn’t a friend’s ears at 2am.
RUN A CHECK
Drop a track in and get the full read in seconds.