Custom Noise without having to use Mix

[XenoxYT]

I suggest implementing custom noise with NoiseNN. White Noise would be used as base and noiseNN gets applied as a filter. and the custom noise could be used for the spin effect as base noise, noisebeat (like mixbeat but noise gets used instead of mixinput), noisepulse (Like Mixpulse but noise gets used as base).

It would work like a filter curve. (Some audio programs have a filter curve eq as feature). Values are entered in decibel.

example:
noise00: e=2 5 3 2 0 0 0 0 0 0 0 0 0 0 0 0 -2 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10
This custom filtered noise

This is how it would look in a sequence:

noise00/20
noise00:spin:300+2/20
noise00:noisebeat:3/20
noise00:noisepulse:20/20

Here is how it would look by using SpinNN and CustomNN:

spin00:noise00:spin:300+2/20
custom00:noise00:noisebeat:4/20
custom00:noise00:noisepulse:4/20

This would make noise customizable without having to rely on external programs for that.

[lm7137]

If the values are in decibels shouldn't they all be negative?

[XenoxYT]

If a value is positive at the start for example the lower frequencies of the noise frequency spectrum gets boosted. (Maybe adding a option to limit the maximum and minimum decibel value). Currently you have to use audio programs that are only available on pc to create custom noise ( White noise as base and it gets filtered using a filter curve eq). This would make SbaGenX noise generation much better on Android. And .sbg files could be shared without having to upload the mix file (Custom filtered noise). If values are negative for example at the end then the part of the noise frequency spectrum there is lower than at the start.

This example would create deeper noise:

noise00: 5 2 0 -1 -1 -1 -1 -3 -3 -3 -3 -3 -5 -5 -5 -10 -10 -10 -10 -10 -10 -10 -10 -20 -20 -20 -30 -30 -30 -30 -30

[lm7137]

Questions that must be nailed down first:

1. What is the x-axis?

• linear frequency bins?
• log-spaced bands?
• octave-like bands?
• fixed count from 0 Hz to Nyquist?

2. How are values interpreted?

• dB gain relative to white noise?
• clipped range?
• allowed positive boost?

3. How is interpolation done?

• linear in frequency?
• linear in log-frequency?
• smoothstep?

4. How is output normalized?

• peak?
• RMS?
• equal perceived loudness is not trivial

5. Is it sample-rate invariant?

• a 44.1 kHz build and a 48 kHz build should not materially change the intended spectrum

[lm7137]

One more important design point:

The new functionality should not be limited to white noise.

If we implement noiseNN, it should define a custom noise spectrum in its own right, not merely “white noise plus a filter” as an implementation detail. Otherwise the feature is framed too narrowly and may constrain later design choices.

The important user-facing concept is:

• noiseNN = user-defined spectral noise shape

not:

• noiseNN = white noise with an attached EQ

That keeps the feature general enough to cover:

• white-like shapes
• pink/brown-like shapes
• band-limited shapes
• unusual custom spectra

without baking “white noise” into the semantics.

[lm7137]

Recommended design answers for noiseNN:

1. What is the x-axis?

• Use fixed log-spaced bands.
• Reason:
  • more natural for audio perception
  • more useful than linear bins for low-frequency shaping
  • easier for users to reason about than FFT-bin-like linear spacing
• Practical recommendation:
  • 32 bands across a fixed audible range, e.g. 20 Hz to 20 kHz

2. How are values interpreted?

• Treat values as dB offsets for band gain.
• Reason:
  • matches how users think about filter curves / EQ
  • gives an intuitive way to attenuate or boost regions
• Practical recommendation:
  • allow a bounded range, e.g. -24 dB to +24 dB
  • clamp outside that range

3. How is interpolation done?

• Interpolate linearly in log-frequency.
• Reason:
  • consistent with the log-spaced band model
  • avoids low-end overcompression that linear-Hz interpolation would cause
• Optional:
  • later allow smoothing modes, but start with one deterministic interpolation rule

4. How is output normalized?

• Normalize to a stable RMS target after shaping.
• Reason:
  • peak normalization is too unstable for noise
  • RMS is the more meaningful baseline for continuous noise sources
• Practical recommendation:
  • normalize shaped noise so that nominal amplitude behaves comparably to existing white/pink/brown sources
  • accept that equal perceived loudness is approximate, not exact

5. Is it sample-rate invariant?

• Yes, it should be designed to be sample-rate invariant.
• Reason:
  • the same noiseNN definition should sound materially the same at 44.1 kHz and 48 kHz
• Practical recommendation:
  • define the curve over absolute frequency bands, not over raw sample indices
  • compute the shaping filter/noise spectrum relative to actual sample rate at runtime

[XenoxYT]

Your recommended answer are very good. Will you implement it in the next alpha release?

[XenoxYT]

Maybe let the amount of bands be set by the amount of numbers entered in NoiseNN. Currently NoiseNN fails when to much or to less values are entered.

[lm7137]

It takes 32 numbers... you can fill out the spec as necessary. I will consider your suggestion however.

Note: noiseNN now also available in the Android App.

[XenoxYT]

I´m currently testing it. Will noisebeat (Uses customNN) and noisepulse (uses CustomNN) get implemented too alongside NoiseNN?

[lm7137]

Please check release https://github.com/lm7137/SBaGenX/releases/tag/v3.9.0-alpha.10 I hope this meets your needs