Turn a recording into a table. Every frame of the file becomes one row of descriptors, the ones that classifiers, similarity searches and plots are built on, with a time column in front. Drop a file and the measurement starts; the download is CSV or JSON.
The columns
RMS is the frame’s level, zero-crossing rate how often the waveform crosses zero per second, a cheap measure of noisiness and brightness. Spectral centroid is the spectrum’s centre of mass in Hz, spread its width around that, rolloff the frequency below which 85 percent of the energy lies, flatness how noise-like the spectrum is between 0 and 1, and flux how much the spectrum rose since the previous frame, scaled by the frame’s energy. The 13 MFCCs describe the spectral envelope compactly; the first is overall level, the rest shape.
Frame size
The window sets the trade between time and frequency. 1024 samples follows fast changes and smears low frequencies; 4096 resolves bass notes and smears transients. 2048 is the usual middle. The hop is a quarter of the window in every case.
Output
CSV has a header row and one row per frame, with values rounded to what the measurement supports. JSON carries the hop and the column names with an array of frames. The means over the whole file are shown on the page for a first look.
Limits
The file is mixed to mono before analysis, so stereo width is not in the table; the stereo analyzer measures that. Rows pool pairwise past 20,000, which keeps an hour of audio to a few megabytes.