Keep the lower tone
Compare A and B at the same position. The 440 Hz tone remains; the 1,760 Hz tone is strongly reduced.
SOUND, MADE LEGIBLE.
LOCAL · FREE · NO ACCOUNT009 / A SOUND INSPECTION LAB
Hear what changes.
Find out what a frequency filter keeps—and what it takes away. Compare at the same position, then keep the sound and its recipe.
A FIRST LISTEN / TWO MINUTES
Start with a change you can explain.
Choose a task, confirm the recording change, then press Play and switch between A and B. Processing never starts playback.
Reduce a steady 1 kHz layer
Listen for the steady layer to recede. The moving sweep is also reduced as it crosses the rejected band: a filter cannot identify a source.
CURRENT RECORDING · REVISION 0
Two tones, one signal
4.000 s / 16,000 Hz / MONO
Original synthetic reference
Your audio stays in this browser’s memory. No audio upload, account or automatic save. Reloading closes the session. WAV only: ≤32 MiB, ≤30 seconds, ≤3 million channel samples, 8–96 kHz.
INSPECT / ORIGINAL ↔ PROCESSED
Follow the signal.
7.813 Hz / bin
8.0 kHz4.0 kHz0 Hz
0 s2.00 s4.00 s
-120 dB0 dB
PSD · dB relative to 1 FS²/Hz. Every analysis frame contributes; time columns average linear power. Display rows show the maximum bin power within their frequency range. Values below −120 dB use the darkest color.
LISTEN / SAME-POSITION A–B
Seeking pauses. A/B switching keeps the same position with a short linear crossfade.
Common A/B headroom: 0.00 dB. Same gain, not loudness-matched. Quieter does not necessarily mean better.
Playback may resample to the device context and uses a short edge fade. Analysis and WAV export retain the source rate without playback fades. Digital peaks are not true peaks or a hearing-safety guarantee.
MEASURE / FULL RECORDING
Numbers behind the sound.
Unweighted sample statistics.
Digital levels—not calibrated sound pressure.
| Signal / channel | Sample peak | RMS | Mean amplitude | At / beyond ±1 |
|---|---|---|---|---|
| A · Original / mono | -6.09 dB | -11.23 dB | ≈ 0 | 0 |
| B · Processed / mono | -9.08 dB | -12.13 dB | ≈ 0 | 0 |
POWER IN SELECTED BAND
80.91% → 100.00%
Fraction of integrated, window-weighted PSD between 300 and 800 Hz. Whole bins whose centers lie inside the edges; not the full-recording RMS.
LARGEST AVERAGED PSD BIN
437.50 Hz
Processed signal. A spectral bin is not a pitch estimate or an automatic identification of a sound.
UNDER THE FILTER
A response, not a promise.
Odd-length, Hamming-windowed sinc filtering, independently for each channel. The center-cropped result keeps 64,000 frames at the original rate.
256 samples (16.00 ms) at each end can be affected by zero extension. Longer filters sharpen transitions but widen this boundary region.
0 → 8.0 kHz-100 to +6 dB
TAKE THE RESULT WITH YOU
Export with the tradeoffs visible.
PCM16 WAV keeps the source rate and channel count. No dither; original metadata is not copied. The numeric report contains settings and spectra, not audio. A processing recipe can reapply the settings after you reopen the exact original WAV or named reference. Keep your original WAV: neither JSON download contains audio.
Gain: 0.000 dB · saturated samples: 0
WAV recipes match original file bytes, not the filename. A renamed copy works; re-encoding or changing the WAV metadata does not. Recipes are not audio backups or a cross-device bit-for-bit guarantee.
READ THE METHOD
Transparent by design.
01 / Preserve the source
RIFF WAV is parsed directly, not through the browser’s resampling decoder. PCM 8/16/24/32-bit and IEEE float 32/64-bit are supported, including a restricted mono/stereo extensible subset. Nonfinite or extreme source values are rejected.
Microsoft · WAV format ↗02 / Account for power
Periodic Hann windows, half-window hops, centered zero extension, no detrending. Each frame’s one-sided PSD is normalized by the squared window weights on actual samples. Frame and channel powers are averaged before conversion to dB; opposite-phase stereo does not vanish.
SciPy · Spectral analysis ↗03 / Keep limits in view
Cutoffs mark finite transitions, approximately half amplitude—not perfect walls. Pass mode normalizes at the band midpoint; reject mode at DC. Centered convolution is not forward–backward filtering. This is an inspection tool, not automatic noise removal or a calibrated acoustic meter.
SciPy · Windowed FIR design ↗