Audio Spectrum Analyzer

What it is for

Sometimes you need numbers, not ears. Which frequency is that whine in the recording? Is this "pink noise" file really pink? Is the recording clipped, or does it have a DC offset? Is a hum at 50 or 100 Hz? This analyzer reads an audio file and answers with an averaged FFT spectrum chart, the strongest peak frequencies, the spectral centroid and roll-off, the spectral slope, the levels per channel and a waveform overview.

WAV files (PCM 8 to 32-bit or float) are read directly. MP3, FLAC, Ogg/Opus, M4A and AIFF are first decoded with ffmpeg to 32-bit float. The tool analyses the first 10 minutes. It does not measure loudness in LUFS (use the Audio Loudness Normalizer for that), and it does not detect musical pitch or chords.

Worked example

We analysed the 4-second room hum file made with the Ambience Generator (120 Hz hum, 22.05 kHz stereo), using an FFT size of 8192.

The summary: "WAV file, 4 s, 22050 Hz, stereo: strongest peak at 120 Hz (−17.5 dBFS), spectral centroid 167 Hz, RMS −20 dBFS, peak −12.9 dBFS." The peak table lists 120, 360, 240, 600 and 480 Hz: the fundamental and its harmonics, with the third harmonic slightly stronger than the second, as in transformer hum. The octave band around 125 Hz is the loudest, and everything above 2 kHz is more than 30 dB lower.

Then we analysed 3 seconds of pink noise at 44.1 kHz. The summary said "no distinct peaks (noise-like spectrum)", the spectral slope was −3 dB/octave, and all octave bands were within 0.3 dB of each other. That is pink noise as it should be. The two channels measured −18.7 and −18.8 dBFS RMS.

How to read the result

• Spectrum chart: frequency on a logarithmic axis, level in dBFS. Red markers show the listed peaks. A sine wave of amplitude 0.5 reads about −6 dBFS at its peak.
• Peak frequencies are refined between FFT bins, so they are more precise than the bin width shown in the assumptions. A peak is only listed when it clearly stands out from its neighbourhood. Noise therefore shows none.
• Spectral centroid: the power-weighted average frequency. High values mean bright or hissy sound, low values mean dull or bassy sound.
• Spectral slope: about 0 for white noise, −3 for pink, −6 for brown. Music usually falls between −3 and −6.
• RMS, peak, crest factor: average level, highest sample and their difference. The DC offset and clipped samples produce warnings when they matter.
• Waveform overview: the minimum and maximum per time slice, for spotting silence, fades and clipping.

Tips and common mistakes

• FFT size trades frequency detail for averaging. 8192 or 16384 separates close tones, and 1024 or 2048 suits short sounds such as claps.
• If you get the warning that the sound is short compared with the FFT size, choose a smaller size. Otherwise the spectrum is hardly averaged.
• Keep the Hann window for almost everything. Rectangular is only for test tones that fall exactly on an FFT bin.
• The spectrum uses the mono mix. Check the per-channel table to find a channel that is quieter or silent.
• An MP3's spectrum often stops sharply at 16–19 kHz. That is the encoder's low-pass, not your microphone.

FAQ

Why is the strongest peak not the note I hear?

Peaks are the strongest spectral lines. A voice or instrument can have a stronger harmonic than its fundamental, and your ear may still hear the fundamental.

Can it measure LUFS loudness?

No. RMS here is a plain signal average. For loudness to broadcast or streaming standards use the Audio Loudness Normalizer.

Are my files stored?

No. Uploads are processed in a private temporary folder and discarded. Only the numbers and charts are shown.

What is the largest file I can analyse?

Up to 50 MB, of which the first 10 minutes are analysed.