Measuring centre-channel vocal separation
How much of a centred signal actually disappears, and what survives.
In short
Method
- 01
Synthesise a 3-second 44.1 kHz stereo WAV containing two tones: 440 Hz at equal amplitude in both channels (a stand-in for a centred lead vocal), and 2 kHz in the left channel only (a stand-in for a hard-panned instrument).
- 02
Process the file with the site’s own separation engine in instrumental mode — mid/side decomposition with the centre cancelled only between 180 Hz and 9 kHz, cascaded 24 dB/oct crossovers.
- 03
Decode the exported WAV back into an AudioBuffer and measure energy at 440 Hz and 2 kHz using a Goertzel filter over the first 2 seconds.
- 04
Compare each against the same measurement taken on the unprocessed source.
Results
| Measurement | Value | Note |
|---|---|---|
| Centred 440 Hz tone | −29.5 dB | about 97% removed |
| Hard-panned 2 kHz tone | −0.5 dB | essentially untouched |
| Output duration | 3.00 s | unchanged |
| Output channels | 2 | stereo preserved |
| Processing time | < 1 s | in-browser, no upload |
Findings
The cancellation is close to total, but only for perfectly centred content
A tone with identical amplitude and phase in both channels is the ideal case, and it loses 29.5 dB. Real vocals are never quite this centred — small differences in reverb, doubling and stereo widening leave residue. The measurement therefore describes the ceiling of the technique rather than what a commercial record will give you.
Band-limiting is what keeps the result usable
Naive L−R cancellation removes everything centred at every frequency, including kick and bass. Restricting cancellation to 180 Hz–9 kHz leaves those bands intact, which is why the processed file still reads as music instead of a thin mono artefact. The 0.5 dB loss on the side-panned tone confirms the untouched bands pass through essentially unaltered.
Stereo width predicts the outcome before processing
Because the technique depends entirely on the difference between channels, a file’s side-to-mid energy ratio predicts whether it can work at all. A mono file has no difference to exploit and returns near-silence, which is why the tool measures this first and refuses to run rather than producing a useless file.
Limitations
- Synthetic tones are the best case; commercial masters are narrower and more processed
- Reverb and delay tails on a vocal are not centred and survive cancellation
- Doubled or hard-panned backing vocals survive by design
- Measured on one browser engine; other implementations of the Web Audio filters may differ slightly
Measuring centre-channel vocal separation
How much of a centred signal actually disappears, and what survives.