Commits · 2ec31cc5ccb92e1b472cf3da1d26490412805a79 · Xiph.Org / Opus

Sep 20, 2023
- added FARGAN adversarial training script · 2ec31cc5
  Jan Buethe authored 1 year ago
  
  Unverified
  
  2ec31cc5
Sep 13, 2023
- removed trailing whitespace in fargan · 82f48d36
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  82f48d36
- integrated JM's FFT ada conv · e7beaec3
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  e7beaec3
- Remove --has-gain and --passthrough-size args · b24c7b43
  Jean-Marc Valin authored 1 year ago
  
  Couldn't be used anymore
  Verified
  
  b24c7b43
- Add conditioning interpolation, fwconv layer · 2f8b36d6
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  2f8b36d6
- Only use one frame of pre-loading · 72c5ea41
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  72c5ea41
- Randomly double the training sequence length · 108b75c4
  Jean-Marc Valin authored 1 year ago
  
  Helps with stability with little cost in training speed
  Verified
  
  108b75c4
- Adds skip connections · d54b9fb4
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  d54b9fb4
- 5-tap pitch predictor · fb570ed8
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  fb570ed8
- Also use previous frame · 2e0c1ad3
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  2e0c1ad3
- explicit signal gain, explicit pitch predictor · 4f63743f
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  4f63743f
- FARGAN initial commit in Opus · 1b13f631
  Jean-Marc Valin authored 1 year ago
  
  Copied/adapted from LPCNet repo
  Verified
  
  1b13f631
Sep 12, 2023
- fix · 4f4b6242
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  4f4b6242
- added requirements.txt to osce · 5467539c
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  5467539c
- added more enhancement stuff · 2f290d32
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  2f290d32
Sep 05, 2023
- added copyright headers · 7b8ba143
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  7b8ba143
- added LPCNet torch implementation · 35ee397e
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  35ee397e
- brought NoLACE up to date · 90a171c1
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  90a171c1
Sep 02, 2023
- Fixes vnni macro redefinition with clang · d431c321
  Michael Klingbeil authored 1 year ago and Jean-Marc Valin committed 1 year ago
  
  Verified
  
  d431c321
Sep 01, 2023
- renamed ShapeNet to NoLACE · 4a47b1a1
  Jan Buethe authored 1 year ago
  
  Signed-off-by: Jan Buethe <jbuethe@amazon.de>
  Unverified
  
  4a47b1a1
Aug 09, 2023
- Changing DRED exp. ID so we can reserve 127 · 1fbdec87
  Jean-Marc Valin authored 1 year ago
  
  There was a suggestion to use 127 for extending the extension ID space
  Verified
  
  1fbdec87
- Don't allocate DRED bits if we don't have enough · bbe4dcc4
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  bbe4dcc4
- Making it easier to remove DRED experimental ID · 0886828e
  Jean-Marc Valin authored 1 year ago
  
  When ready, change DRED_EXTENSION_ID to the final ID, remove DRED_EXPERIMENTAL_VERSION completely, and change DRED_EXPERIMENTAL_BYTES to zero (eventually remove it).
  Verified
  
  0886828e
Aug 08, 2023
- improved auto-scaling in wexchange · 5160d7fd
  Jan Buethe authored 1 year ago
  
  Unverified
  
  5160d7fd
Aug 04, 2023
- Add fwgan_load_model() · 6cba42f9
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  6cba42f9
- int version for convenience · 022f2b7e
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  022f2b7e
- Adding some comments · f7ee7137
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  f7ee7137
- Continuation now working · bd23d911
  Jean-Marc Valin authored 1 year ago
  
  Added required filtering/delay/memory to get continuation
  Verified
  
  bd23d911
Aug 03, 2023
- More scaling/filtering to run_fwgan_subframe() · 6a184fc7
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  6a184fc7
- More FWGAN refactoring · b1a601fb
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  b1a601fb
- FWGAN refactoring · 5179896b
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  5179896b
- Applying continuation after first subframe · b0e1a2eb
  Jean-Marc Valin authored 1 year ago
  
  Continuation matches Python code now
  Verified
  
  b0e1a2eb
Aug 02, 2023
- Avoid sin()/cos() calls for pitch reference · 3eac8c12
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  3eac8c12
- Don't compute linear activation in-place · 83e95a5c
  Jean-Marc Valin authored 1 year ago
  
  saves a few cycles
  Verified
  
  83e95a5c
Aug 01, 2023
- Add FWGAN to lpcnet_demo · bf5eb5bf
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  bf5eb5bf
- C implementation of FWGAN · e62fd5c5
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  e62fd5c5
- Skeleton for FWGAN code · 155367d2
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  155367d2
- Handle float matrices with multiple of 8 rows · e9f8402a
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  e9f8402a
- Add Gated Linear Unit (GLU) · 5eaa4a50
  Jean-Marc Valin authored 1 year ago
  
  Verified
  
  5eaa4a50
- updated dump_model_weights.py · 5e045405
  Jan Buethe authored 1 year ago
  
  Unverified
  
  5e045405