250117_test model used for paper curves

drowe67 · Jan 20, 2025 · d5f1383 · d5f1383
1 parent c1899e6
commit d5f1383
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/250117_test/checkpoints/checkpoint_epoch_200.pth b/250117_test/checkpoints/checkpoint_epoch_200.pth
diff --git a/README.md b/README.md
@@ -422,8 +422,8 @@ A log of models trained by the author.
 | model19_check3 | model19 but loss function weighting for data symbols reduced fom 1/18 to 0.5/18, which reduced vocoder feature loss with just a small impact on BER.  Loss at various op points and channels very close to model17 | Fs | - |
 | model20 | model19_check3 but loss function weighting for pitch and corr doubled, attempt to improve rick samples.  Didn't help. |  Fs | - |
 | model21 | very basic fixed EbNodB model like model02 but at 20dB at epoch 30, produced good quality speech on rick, loss 0.02, but not a practical solution |  Rs | - |
-| model22 | very basic
- fixed EbNodB model like model02 but at 10dB at epoch 30, rick sample starting to get tonal/pitch shift artefact, in between, loss 0.045 |  Rs | - |
+| model22 | very basic fixed EbNodB model like model02 but at 10dB at epoch 30, rick sample starting to get tonal/pitch shift artefact, in between, loss 0.045 |  Rs | - |
+| 250117_test | Experimental model for joint PAPR/bandwidth optimisation (no pilots/CP) | Rs | - |
 
 Note the samples are generated with `evaluate.sh`, which runs inference at rate Fs. even if (e.g model 05), trained at rate Rs.