alternative audio data structures / storing ops instead of applying immediately

[dy]

Following https://github.com/audiojs/audio-buffer-list/issues/5.

The current API approach is covered by a lot of similar components, it is destined to insignificant competition and questionable value. The main blocker and drawback is the core - audio-buffer-list component, which does not bring a lot of value, compared to just storing linked audio-buffers.

Alternately, audio could be focused on storing editing-in-process, rather than data wrapper with linear API, similar to XRay RGA.

Principle

• storing operations rather than applying them to data
  • :+1: no precision loss
  • :+1: faster insertion/removal
  • :+1: allows for collaborative editing
  • :+1: allows for faster re/adjusting params of applied control/envelope
  • :-1: possibly somewhat slower playback due to applied transforms stack, hopefully having heavy-duty fx is not a part of editing process
    • ! possibly compiling fx program dynamically, akin to regl
    • ! pre-rendering audio for faster playback
• undo/redo history methods store operations, not full binary replica every step
• branching - allows for alternatives

Pros

• :+1: makes audio unique
• :+1: makes it suitable for editors

[dy]

Reference structures:

• https://github.com/coast-team/mute-structs/
• https://github.com/Chat-Wane/LSEQTree
• https://github.com/atom/xray/tree/master/memo_core

In fact, git seems to be suitable for that too.

Note also that the class technically should allow to utilize any underlying technology: time series, STFT, formants, HPR, HPS SPS etc models (https://github.com/MTG/sms-tools/tree/master/software/models), wavelets etc.

:+1: In the case of formants, for example, transforms are theoretically times faster than the time series. :+1: Abstract interface would discard sampleRate param and make Audio just a time-series data wrapper, with possibly even uncertain/irregular stops. We may want to engage a separate time-series structure for that, which seems to be broadly demanded, from animation/gradient/colormap stops to compact storing of observations.

[dy]

Lifecycle

1. Initialize data model
2. Input data source
   • Convert input data source to target model
3. Modify data source
   • Create stack of modifiers/reducers/transforms
   • Modifiers can possibly be applied real-time
4. Play data source
   • Apply stack of transforms, play / apply transforms per-buffer
5. Get stats
   • Should model include stat params straight ahead?
6. Output data source
   • Apply stack of transforms, output

Plan

• [ ] Collect set of concerns, use-cases of responsibility
• [ ] Come up with ideal API covering all these cases
• [ ] Create baseline/edge/real tests for the cases
• [ ] Fix tests

Stores

• [ ] time-series store
• [ ] web-assembly store
• [ ] stft store
• [ ] harmonic model + residual store
• [ ] formants store
• [ ] wavelets store
• see ref for other stores (Bercelona Institute)