GramFrame — Spectrograph Player Trial

Choose a WAV recording from this computer. It is analysed in your browser and never leaves this machine.

or drag a .wav file onto this box

Analysis

FFT size

Window overlap

Frame averaging

Top frequency (Hz)

Window (seconds on screen)

Display

Background normalisation

Darkest percentile

Brightest percentile

What each control does

FFT size: How many samples go into each transform, and so how finely the frequency axis is divided: doubling it halves the width of a column in hertz. This is the single biggest reason a gram can look coarser than you expect — at 1024 on a 16 kHz recording each column is 15.6 Hz wide, so the whole 0–200 Hz band is thirteen columns and closely spaced tonals are merged before they are ever drawn. The cost is time: a longer window smears a sound that changes quickly.
Window overlap: How far each transform is stepped on from the last. More overlap gives more rows for the same FFT size — a smoother picture and a better chance of catching a short event — at proportionally more analysis time.
Frame averaging: Averages several transforms into each painted row. A single transform of noise is a rough estimate, which is what makes an un-averaged gram look speckled; averaging four of them halves that speckle and lets a steady tonal show through it. The row then covers proportionally more time, so a brief sound is diluted. Averaging also cuts the row count, which is what lets a long recording keep a small FFT size.
Background normalisation: Rather than painting the measured level, paint how far it stands above the background around it, so a weak tonal reads the same wherever it lies. Split window estimates the background across frequency, from bins either side with a gap left around the bin itself: a tonal never contributes to the background it is judged against, so one that runs the whole recording survives. It cannot show broadband structure — anything as wide as the window counts as background. Per bin median estimates it over time instead, and does the opposite: it removes a fixed spectral shape cleanly and keeps broadband events, but a tonal present throughout is background to it, and is flattened away.
Darkest / brightest percentile: Which slice of the levels the colours are spread across. Everything below the first is painted the darkest colour and everything above the second the brightest — and that clipping happens when the picture is made, so the contrast sliders on the player cannot bring it back. Lowering the darkest percentile towards 0 is what returns the quietest part of a recording to the picture.
Colour map: The row of buttons under the player's own bar, below the contrast sliders, repaints the picture with a different colour map. Grey is black for the quietest level and white for the loudest, for comparing with a legacy display that only ever drew grey. Colour is the table the player always used: its hue boundaries act as extra contours, so more features stand out than in grey, but its brightness is not in level order — yellow is lighter than the red above it. The other four are matplotlib's perceptually uniform maps: they keep the hue boundaries with brightness that rises strictly with level, so brighter always means louder. Inferno and Magma start from black, so the background stays dark as on a legacy display; Viridis starts from dark blue and runs through green; Plasma has no dark end at all. Each is a repaint of the same analysis: instant, no reading changes, and the choice survives a Re-analyse.