Spectrogram from CSV (time-frequency plot)

The sweep shows as a bright line that climbs from 5 Hz at the left to 50 Hz at the right, over the dark floor of the noise. Above 50 Hz there is only noise, so typing a frequency range of 0 to 60 Hz fills the heat map with the sweep. Each slice’s loudest frequency follows the sweep to within a frequency step, although the sweep moves 1.5 Hz during one 2-second segment.

The same in Python

With the sample file downloaded, these lines give SciPy’s answer for the same settings (skiprows=[1] skips the units row):

import numpy as np, pandas as pd
from scipy import signal

x = pd.read_csv("megarows-sample-sweep.csv", skiprows=[1])["Storm_Pressure_Sweep"].to_numpy()
f, t, S = signal.spectrogram(x, fs=500, window=("tukey", 0.25), nperseg=1024, noverlap=512)
print(f"{len(t)} time steps x {len(f)} frequencies")
j = 28
print(f"slice at {t[j]:.2f} s: peak at {f[np.argmax(S[:, j])]:.2f} Hz")
print(f"strongest {10 * np.log10(S.max()):.1f} dB")

They print:

57 time steps x 513 frequencies
slice at 29.70 s: peak at 27.34 Hz
strongest 4.5 dB

We ran them on 8 October 2026 with NumPy 2.1.3, SciPy 1.14.1, pandas 2.2.3 and Python 3.12. A test runs the same calculation through MegaRows on every change to the site and checks that every slice’s loudest frequency is SciPy’s and that its values match SciPy’s to about 12 significant digits. Download the sample file (CSV, 400 KB).

Reading a spectrogram

• A horizontal line is a steady tone: a fan, a hum, a resonance that keeps ringing.
• A slanted line is a frequency that changes: a sweep, a machine speeding up, the sample above.
• A vertical stripe is a short event with power at many frequencies at once: a knock, an impact, a gust.
• The colour is the power spectral density in dB: every 10 dB is ten times the power.

The settings, and what they change

Window, segment length and overlap

The signal is cut into segments of the chosen length, each tapered by the window and turned into a PSD: one column of the heat map. A longer segment gives finer frequency steps (the sample rate divided by the segment length) but coarser time steps; a shorter one the opposite. The overlap sets how far apart the slices are: at 50 % a new one starts every half segment. Tukey (0.25), SciPy’s default for spectrograms, keeps most of each segment at full weight; Hann leaks less between nearby frequencies.

Colour scale in dB

The colours run from the strongest value down by the colour range, 80 dB unless you change it, with the scale beside the heat map. A smaller range brings out the strongest parts; a larger one shows quieter detail.

The values are a power spectral density in the channel’s unit squared per hertz, and the colours 10·log10 of it, with 1 unit²/Hz as 0 dB: a signal in Pa gives Pa²/Hz, in g gives g²/Hz. Give each column its unit in a units row under the header, as the sample does.

Frequency range

Empty, the heat map shows every frequency from 0 Hz to half the sample rate. Type a range, such as 0 to 60 Hz, to keep only those frequencies: the heat map uses its full height for them.

For measured data, not audio

This tool opens data files with a time column and signals saved as numbers: CSV, TSV and TXT, Excel workbooks, JSON, MATLAB .mat, Parquet, HDF5 and LabVIEW TDMS files, from data loggers, test rigs, sensors and simulations. It does not open WAV, MP3 or other sound files. The sample rate comes from the time column; without one, you type it in. It works on files with millions of rows.

Sampling and gaps

The spectrogram assumes evenly spaced samples: each row is one sample, one step of the sample rate after the one before. Before you read the frequency axis, check the line under the X axis in the settings: it gives the sample rate, or says that the steps are uneven.

• Sample rate. It is 1 ÷ the median step of the time column, so a few late or early rows do not change it.
• Uneven steps. When any step is more than 1 % away from the median, the line says what share of the steps differ, how many gaps there are and which sample rate the results assume. Nothing is resampled: each row stays one sample, as if the steps were even.
• Gaps. A step longer than 1.5 times the median counts as a gap. The rows on either side are joined, so after a gap the heat map’s time axis, in seconds from the first row at the assumed rate, runs behind the clock.
• Empty cells. A missing value in the signal is filled with a straight line between its neighbours, and a missing first or last value takes the nearest one, before the spectrogram is calculated.
• Duplicate times. Rows with the same time stay separate samples, and the line counts the steps that do not go forward. When most steps repeat or go back, there is no sample rate to read, and the line says so.
• Time units. Dates and times, and durations such as 00:01:30, count in seconds. A number column counts as seconds unless its unit says ms, µs, ns, min or h: give it its unit in a units row under the header, as the sample does.

For exact frequencies from a file logged unevenly or with gaps, resample it to one fixed rate first, as the settings suggest.

Checked against SciPy

Private by design: nothing is uploaded

Measurements are often confidential. This tool reads your file and calculates the spectrogram in your browser, so the file and its name never leave your device; you can check it in the Network tab of your browser’s developer tools.