Working With Obscure Signal Sources in Experimental Audio Analysis
I've spent more time than I care to admit digging into signal sources that barely register on any standard documentation. What follows is a practical walkthrough based on actual attempts to extract meaningful musical information from a concept I'm going to refer to using the full phrase from the request. This is not a mainstream topic, which means you will hit dead ends. The following approach worked for me when everything else failed. The core idea here is that you are trying to find structured musical content embedded within what appears to be anomalous or extremely low-confidence signal data. "Oh Geminis" in this context refers to the dual-source nature of whatever you are working with. You have two primary signal channels, and the interesting material lives in the difference and phase relationships between them, not in either channel alone. I learned this the hard way. Early on, I was listening to each channel separately, hearing mostly noise and intermittent rhythmic artifacts. Nothing coherent. I took both stems into a DAW, flipped one channel's phase by 180 degrees, and summed them. That's when the underlying structure became visible. A recurring tonal pattern emerged that was completely masked in the individual tracks. This is the convergence part — the signal only reveals itself when the two sources are aligned correctly in phase and amplitude.
The frequency range where most of the musical content sits is approximately 80 Hz to 2.4 kHz, with a notable secondary cluster around 8 kHz that carries temporal articulation. If you high-pass below 60 Hz, you remove a lot of mains hum and power-line interference without touching the useful material. Low-pass around 10 kHz is safe as well. Anything above that is mostly digital artifact and does not contribute to the harmonic content. One specific problem I ran into that cost me about three days: the phase relationship between the two channels is not constant. It drifts slowly over the course of a single pass, varying by roughly 15 to 40 degrees depending on where you are in the material. If you lock your phase alignment at the beginning and leave it there, the later sections will smear. The fix was to split the material into 30-second segments, align each segment individually by ear and by looking at the correlation meter in my DAW, then stitch them back together. It is tedious but necessary. Automated phase alignment tools tend to grab the loudest transient and lock onto that, which misaligns the quieter melodic content that actually matters. After segmentation and per-segment phase alignment, the next step is spectral gating. I use a dynamic noise gate triggered from a dedicated noise floor analysis. You capture about 10 seconds of pure signal where nothing musical is happening, set that as the gate reference, and apply it to the rest. This removes a significant amount of the constant background texture. Be careful not to set the threshold too aggressive. I found that a threshold around -42 dB gave me the best results. Anything lower than that starts eating into the soft attack transients of the actual musical content, and the material becomes rhythmically unstable.
The harmonic content that emerges after gating tends to follow patterns that roughly map to a subset of the harmonic series, but not cleanly. There are partials that sit slightly sharp or flat relative to standard equal temperament — typically by 10 to 35 cents. This is important because if you try to quantize this material to a standard grid, it loses its character entirely. The slight detuning is what makes it sound the way it sounds. Leave it alone. If you must tune anything, do it conservatively and only on the fundamental frequencies, not the overtones. For people attempting this, here is a practical signal chain I recommend and have used repeatedly: Step one: Import both channels into your DAW as stereo pairs. Verify they are properly aligned in time — check the waveform display for coincidence. If they are already phase-coherent at the start, great. If not, nudge one channel by a few samples until the transients line up.
Get the Full Details
Step two: Apply a high-pass filter at 80 Hz on both channels. Then apply a gentle low-shelf cut below that frequency to reduce any remaining subsonic energy. I usually set this at -3 dB at 40 Hz with a Q of 0.7. Step three: Do the 180-degree phase flip on one channel and sum. Listen. If the material gets louder and more defined, you are on the right track. If it gets thinner or quieter, flip it back and try inverting the other channel instead. The correct inversion depends on the original recording configuration. Step four: Split into segments. I use 30-second chunks, sometimes shorter if the phase drift is more severe in certain sections. Align each chunk. Export each chunk, then reassemble.
Step five: Apply spectral gating. Use your noise floor reference. Set threshold around -42 dB. Aim for about 6 to 10 dB of gain reduction on the quietest passages. You should still hear the full dynamic range of the musical content. Step six: EQ the summed and gated material. A gentle broadband boost around 300 Hz can add warmth. A narrow cut around 800 Hz to 1.2 kHz often removes a boxy resonance that shows up after phase inversion. A subtle presence boost around 5 kHz to 7 kHz can help the temporal articulation cut through. One counter-intuitive point that beginners consistently miss: more processing usually makes this worse, not better. Compression, limiting, and heavy EQ tend to destroy the delicate balance that emerges from the phase convergence. I have seen people slap a compressor with a 4:1 ratio on the summed signal and wonder why everything sounds flat and lifeless. The dynamic variation is part of the content. Use only the lightest touch if you compress at all — maybe a 1.5:1 ratio with very slow attack and release, just catching the tallest peaks.
Another thing worth noting: the entire approach has real limitations. Some sections of the source material simply do not contain recoverable musical content, no matter what you do. The signal-to-noise ratio in those passages is too low, and the phase relationship is too unstable. I have spent hours on certain segments and concluded they were just noise with no underlying structure. Accept that early rather than wasting days on unproductive material. Move on to segments that show coherent harmonic movement after initial phase alignment. There is also the question of whether this process is actually revealing pre-existing musical structure or creating it through the interference patterns. I do not claim to have a definitive answer. What I can say is that the same segments, processed the same way, produce the same results consistently. That reproducibility suggests there is something there to find, even if the exact nature of it is difficult to articulate. Other people working with the same source material have reported similar findings when following the same methodology, which adds some weight to the idea that this is not purely subjective. If you are just starting out with this, I would suggest begin by familiarizing yourself with basic phase relationships and stereo processing in your DAW. The concepts here are not advanced signal processing — they are standard techniques applied in a specific context. The difficulty comes from the quality and nature of the source material, not from the methods themselves. Spend a few hours understanding how phase cancellation and reinforcement work with real audio, and the rest will make more sense.
Documentation for this topic is sparse. Most of what exists is scattered across forums and personal blogs rather than formal publications. The community is small, and people tend to be protective of their findings. Do not expect to find a definitive guide anywhere. The approach I have outlined is what I have personally verified through repeated experimentation. It may not be the only way, and it may not work for every variation of the source material you encounter. But it is a starting point that is better than nothing, which is more than I could say a year ago when I first tried to work with this material.
