Skip to content

Wishlist

Anna Xambó edited this page Apr 26, 2021 · 19 revisions

MIRLCa

Training mode

  • Make training incremental and persistent across sessions. Ideally, some accounting about the files downloaded should be in place so to avoid repeated downloads biasing the training.
  • Create subfolders for the "ok" and "ko" sounds.
  • Instead of random sounds, can we also have the same 130 sounds from the manual training.
  • In training by ID, add a second argument with the good/bad label e.g. a.trainid(3333, "ok").

Performance mode

  • Load as many models as wanted. It can be an argument of the instantiation.
  • Be able to configure what candidate of the list of best candidates to retrieve: random, 2nd and 3rd, etc.

Overall

  • Solve the problem with directories: models vs sounds vs trained sounds.

Bugs

  • Sometimes in the process of trainning a sound remains in loop, even if a new sound is called to evaluate. Is there a way to stop a sound?

MIRLC2.0

Additional features

  • Combine 2 features: pitch and duration, for example (combined search and content-based search).
  • Add parameters/filters to a random query: e.g. giving a range for the min-max duration.
  • Replace a sound (instead of adding a sound).
  • Remove a sound.
  • Copy an existing group.
  • Add a free method that frees the dictionary and buffers.
  • Trim the initial silence in downloaded recordings.
  • Be able to specify the playback mode (sequential/parallel) before the sounds start playing (trying to do so now, causes an error).
  • Ability to incorporate batch file processing so that you could slice a sample up into multiple snippets and plays those samples back randomly, a sort of granular approach.
  • Avoid downloading duplicated sounds.
  • API key as argument of the constructor (same for MIRLCa).
  • Check if a sound has been already downloaded, do not download it again.
  • Add the possibility of multichannel (or controlling the output channels (bus numbers)).
  • Facilitate more control of filtering from Freesound e.g. a subset of sounds.
  • Look into showing a progress bar when sounds are downloading (it only happens in Linux?).
  • Write the decimator from scratch to avoid sc3-plugins as a requirement.
  • Write a distortion effect that does not require sc3-plugins.
  • Add flanger effect.
  • Add limiter effect.
  • Mute certain amount of sounds.
  • Filter non-derivative works.
  • Playup, and playdown&up.

Bugs

  • Debug sequence (check the logic of the model).
  • Check how to save credit files using Atom or not SC IDE.

Documentation / Video Tutorials

  • Create a video tutorial of training by ID or combined (id + random).
  • Create a video tutorial of how to search sounds by ID (either using MIRLC or Freesound).
  • Create a video tutorial of where to locate the files (setting up).
Clone this wiki locally