FreeDV
/
phasenn
mirror of https://github.com/drowe67/phasenn.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Commit Graph

  • 6171230b92 updated README master David 2020-11-07 17:01:52 +1030
  • 123213d16d converted phasenn_test8 to tf.keras, this script has a custom loss function David 2020-11-07 16:50:48 +1030
  • 479a2580fb updated README, phasenn_test9 ported ed to tf.keras David 2020-11-07 16:10:16 +1030
  • cfb446155d moved amplitude specific simulations to a new repo David 2019-12-23 12:58:08 +1030
  • 041bb29ec3 first pass at eband->{Am}, getting quite reasonable SD results and test plots for 10-14 length vectors David 2019-12-22 13:06:43 +1030
  • 628dca7638 restored to original NN model David 2019-12-21 14:04:28 +1030
  • d2039fd4b7 cleaned up and example cmd lines David 2019-12-21 13:56:12 +1030
  • b65abfe150 added per frame linear const est David 2019-12-20 13:25:33 +1030
  • 9076bd73fb building up script to explore interpolation/decimation of rate K vectors using NNs David 2019-12-18 05:58:06 +1030
  • 900b286bcc adding results, work in progress David 2019-12-15 18:34:21 +1030
  • 2ff2166344 comparing with conventionally decoded vectors, some issues as conventional has benign errors at end points David 2019-12-15 18:34:06 +1030
  • 5c4367d8a1 Revert "quite reaonable results on visible inspection of a few frames" David 2019-12-13 17:36:59 +1030
  • fc79915bab quite reaonable results on visible inspection of a few frames David 2019-12-13 17:36:47 +1030
  • d372e37903 quite reaonable results on visible inspection of a few frames David 2019-12-13 17:36:32 +1030
  • 76ffb675e4 first pass at newamp1_train David 2019-12-13 16:36:41 +1030
  • aa2087ebdd add phase0 model to compare David 2019-12-13 13:32:33 +1030
  • 712b1d5198 using voiced/unvoiced model in synth David 2019-12-13 09:08:34 +1030
  • a2832ecb6a training on just high energy V frames, didn't improve randomness of UV speech David 2019-12-13 08:15:34 +1030
  • 837de6367c tests with synthetic speech to attempt to track down UV issueson real speech David 2019-12-11 05:32:54 +1030
  • c1484a76c1 fixed custom loss function - good matches and even overfitting - but still probs with UV David 2019-12-09 20:22:05 +1030
  • 5c19633779 adding vanilla codec2 as a comparison David 2019-12-08 17:46:27 +1030
  • 1b5b854c61 phasenn_train.py used for initial synthesis David 2019-12-08 17:21:40 +1030
  • 67adf33ebe building up synthesis tools David 2019-12-08 17:18:04 +1030
  • 62c5474cf6 first pass at synth script to put everything back together David 2019-12-08 12:55:17 +1030
  • 03722ac424 simulated V/UV speech, used to test phasenn_train.py, with this signal, UV strain sup well (nice and random) David 2019-12-08 12:54:40 +1030
  • de7b7ea394 amplitude waiting doing OK for a few test frames of voiced speech, UV too clicky, suspect random phases being averaged, 22 minute train_8k David 2019-12-07 14:28:57 +1030
  • 0b34b92906 first pass phase training, no t overly impressed by results, but basic frameowkr doing sensible things, e.g. overtraining on small, all_8k database David 2019-12-07 11:58:21 +1030
  • fd14956b60 first pass with all_8k, converges _too_ well, firs UV phase perfectly David 2019-12-07 10:37:26 +1030
  • 6c8a2052ff added better visualisation of phase and amplitude together in PNGs David 2019-12-06 12:29:12 +1030
  • f0ee4a7684 added CLI, ability to work with internal and external n0 removal David 2019-12-06 09:10:52 +1030
  • 329bbf710f changed cost function for est_n0, getting reasonable results for simulated speech and real speech vectors make some sort of sense (hard to tell with phase spectra) David 2019-12-04 16:21:54 +1030
  • 3a70ff4677 Merge branch 'master' of github.com:drowe67/phasenn David 2019-12-04 14:26:43 +1030
  • 4f8c7ea4fb added some tests for est_n0, working OK for basic impulse case David 2019-12-04 14:26:24 +1030
  • 9ab6a221dc debugging plot_n0.py/est_n0 C code David 2019-12-03 07:07:59 +1030
  • 56abe1afd2 utility to plot n0 for Codec 2 model files David 2019-12-01 18:41:14 +1030
  • 0a39182a84
         Update README.md drowe67 2019-12-01 11:08:43 +1030
  • 12ccecab68
         Update README.md drowe67 2019-12-01 11:07:45 +1030
  • 57ccb4c122
         Update README.md drowe67 2019-12-01 11:06:23 +1030
  • 900fdbfa3f
         Update README.md drowe67 2019-12-01 11:03:14 +1030
  • b3a25900a1
         Update README.md drowe67 2019-12-01 10:55:10 +1030
  • 290323bf81
         Update README.md drowe67 2019-12-01 10:27:46 +1030
  • 0d6129d500
         Update README.md drowe67 2019-12-01 10:25:23 +1030
  • 0b605d3af5 swapped example names David 2019-12-01 10:23:42 +1030
  • cb8cb985ac example images David 2019-12-01 10:05:23 +1030
  • 3ef98f55a7 codec 2 model record I/O test working OK David 2019-12-01 09:50:58 +1030
  • 63a1db65e6 still now working David 2019-11-30 13:53:18 +1030
  • 7b448657bc added V/UV model and actually doing a reasonable job David 2019-11-29 18:21:10 +1030
  • 39827bfdf1 merge David 2019-11-29 18:21:05 +1030
  • 3173bb3961 building up V/UV model David 2019-11-24 10:22:48 +1030
  • bca28cc34a docu David 2019-11-24 10:06:52 +1030
  • a3d92a4d29 added 2nd formant simulation David 2019-11-24 10:06:07 +1030
  • c2ac3d520c training Am->phim using n0 +/-1 acurracy, working reasonably David 2019-11-23 15:05:50 +1030
  • 1eeb6864e8 attempt to remove linear component of phase, unsucessful David 2019-11-23 13:35:24 +1030
  • 43b4979784 extracting n0 using DSP, leave residual linear term that greatly affects phase at high freqs, which will make training difficult David 2019-11-23 12:52:24 +1030
  • 2b09ac0e98 building up n0+disp sim David 2019-11-23 09:50:45 +1030
  • 70bfc1185e DSP based n0 estimation David 2019-11-22 17:41:29 +1030
  • 4fe30e210b more realistic range of filters David 2019-11-22 17:05:22 +1030
  • 02bc840667 correct n0 values David 2019-11-22 17:01:03 +1030
  • d3e0778cb6 broken at point we try to match in the DFT domain... David 2019-11-22 15:28:59 +1030
  • e01a050516 broken... David 2019-11-22 15:28:14 +1030
  • 037394fdef using e_rect as input (ex phase) rather than phase_rect (ex+dispersive) David 2019-11-22 12:59:33 +1030
  • a82226f9da work in progress combination of n0 est and dispersive part David 2019-11-22 12:43:59 +1030
  • 371d73d4d4 updated README David 2019-11-22 12:42:55 +1030
  • 2d95bedd05 test9 n0 scaled by P_max, sampling excitation on width length DFT bins rather than m*Wo David 2019-11-22 12:41:36 +1030
  • d48c50dbe2 still not working ..... David 2019-11-22 11:54:21 +1030
  • 52fc623291 cleaned up n0_dft test David 2019-11-21 19:36:15 +1030
  • 13f29a2ab4 sampling e on 0..width-1 to reduce inaccuracy at nigh n0 values, building up test for layer David 2019-11-21 19:29:34 +1030
  • d9a3d57396 building up test91, still some bugs David 2019-11-20 20:33:17 +1030
  • e921bc8f1e building up test91, still some bugs David 2019-11-20 20:33:05 +1030
  • 2850ede820 first pass at README.md David 2019-11-17 12:30:07 +1030
  • 99da10f0d4 moved out of codec2 repo David 2019-11-17 12:05:58 +1030