95 commits

Author	SHA1	Message	Date
Jan Buethe	1978cc6094	refactoring	2022-10-21 12:13:38 +00:00
Jan Buethe	d1646a680a	added NFEC decoder C implementation	2022-10-20 17:27:39 +00:00
Jan Buethe	ea4d8f54c3	added statistical model to dump_nfec_model	2022-10-19 17:18:25 +00:00
Jan Buethe	c1b357ed47	first attempt of C implementation of fec encoder (not tested yet due to NEON/DOT_PROD not being separable)	2022-10-18 19:30:23 +02:00
Jean-Marc Valin	9629ea6a70	Fine-tuning the scripts	2022-10-14 01:01:39 -04:00
Jean-Marc Valin	0f7fe64d5a	Compute FEC features based on loss pattern	2022-10-05 01:56:53 -04:00
Jean-Marc Valin	89db314efb	Updating fec_encoder.py for recent changes	2022-10-04 15:46:13 -04:00
Jean-Marc Valin	61459c24e0	Change decoder architecture to be like the encoder	2022-10-04 00:27:36 -04:00
Jean-Marc Valin	79d1a916d0	Weighting loss by 1/sqrt(lambda)	2022-10-02 12:34:42 -04:00
Jean-Marc Valin	0b01863732	Larger range of quantizers	2022-09-30 22:21:30 -04:00
Jan Buethe	524f84800f	removed dump_data delay from total_delay in fec_encoder.py	2022-09-30 16:00:02 +02:00
Jean-Marc Valin	4c82d3b419	Completely move quantization out of encoder and decoder	2022-09-30 03:53:40 -04:00
Jan Buethe	cd0993fd8c	clarifications in help printout	2022-09-29 21:14:34 +02:00
Jan Buethe	97ffa94d5c	fixed decoder bug (non-quantized input)	2022-09-29 21:13:30 +02:00
jbuethe	589e674116	alignment fix in fec_encoder	2022-09-29 17:54:34 +02:00
jbuethe	a866abe3c3	ported debug changes from fec_encoder_torch.py	2022-09-29 14:18:41 +02:00
Jean-Marc Valin	0a2d6dfcb6	Use the encoder state as decoder initial state ... Helps reduce the error on the most recent frames	2022-09-28 15:34:02 -04:00
Jean-Marc Valin	38dda0f950	Oops, forgot to run PVQ quantization for the state	2022-09-28 15:33:20 -04:00
jbuethe	b43f077ba8	corrected offset between decoder output on even/odd latent frames	2022-09-27 16:37:46 +00:00
jbuethe	01baf1a0fc	added missing dead-zone to encode_rdovae.py	2022-09-27 16:31:04 +00:00
jbuethe	be42c3b514	added fec_encoder.py and corresponding fec_packets.[chpy]	2022-09-27 16:29:13 +00:00
Jean-Marc Valin	ef12c29f14	Update encoder/decoder	2022-09-27 02:28:13 -04:00
Jean-Marc Valin	405aa7cf69	WIP: training with different alignment	2022-09-27 02:28:12 -04:00
Jean-Marc Valin	981d06eefd	Refactoring towards multiple offset decoding	2022-09-27 02:28:12 -04:00
Jean-Marc Valin	a4f7c157cf	Stop decimating in the encoder	2022-09-27 02:28:12 -04:00
Jean-Marc Valin	fdd51eb760	RepeatVector no longer likes tensor input	2022-09-27 02:28:11 -04:00
Jean-Marc Valin	8569121f6c	RDO-VAE work in progress	2022-09-27 02:28:11 -04:00
Jean-Marc Valin	d45ab6fcb6	Move back to tanh for frame rate network ... Swish has lower loss, but doesn't seem to improve quality	2022-09-24 03:22:57 -04:00
Jean-Marc Valin	f5c251c5d5	Properly align LPC with lookahead in data loader	2022-09-24 03:21:36 -04:00
janpbuethe	f3c738d45f	removed debug prints in dump_lpcnet.py	2022-09-07 09:10:19 +00:00
Jean-Marc Valin	60450472a6	Merge branch 'plc_challenge' into master	2022-09-07 00:38:55 -04:00
janpbuethe	920300c546	Add lpc weighting and model parameter handling ... Model now stores LPC gamma, look-ahead, and end-to-end. Parameters aren't quite reliable yet, YMMV	2022-09-06 23:14:39 -04:00
Jean-Marc Valin	e034b1096b	Biasing for overestimating the pitch correlation	2022-02-25 13:55:21 -05:00
Jean-Marc Valin	05f02aaa49	Adaptive bias ... Bias the energy for vowels, but not for consonants	2022-02-24 13:03:44 -05:00
Jean-Marc Valin	1ae958acf3	Make LPCNet frame rate network use swish	2022-02-16 01:01:51 -05:00
Jean-Marc Valin	2a8bcf4c0f	3-part pitch loss function	2022-02-13 02:51:22 -05:00
Jean-Marc Valin	2e06c07893	more emphasis on pitch loss	2022-02-11 03:20:20 -05:00
Jean-Marc Valin	b2847687e0	Getting rid of prediction delay	2022-02-09 03:06:41 -05:00
Jean-Marc Valin	dc539a9ce9	WIP non-causal PLC	2022-02-05 22:57:53 -05:00
Jean-Marc Valin	2e18f0d160	Using Burg cepstrum for feature prediction	2022-02-04 22:04:23 -05:00
Jean-Marc Valin	3e2198c6e1	Learning to predict time t+1 from time t ... Instead of t from t	2022-02-02 15:00:24 -05:00
Jean-Marc Valin	cd2e568bb6	Using lost packet file instead of uniform random	2022-02-02 00:22:57 -05:00
Jean-Marc Valin	1db1946f77	Support for biased loss	2022-02-01 02:57:50 -05:00
Jean-Marc Valin	32a63fd31d	WIP: PLC prediction	2022-01-31 23:21:55 -05:00
Jean-Marc Valin	c45963d40a	Code for testing PLC models	2022-01-31 18:49:19 -05:00
Jean-Marc Valin	e1181bcad0	oops, fix band loss	2022-01-30 17:29:33 -05:00
Jean-Marc Valin	c8cbfa7e9b	Adding feature prediction (totally untested)	2022-01-29 02:54:48 -05:00
Jean-Marc Valin	7fdca7f01d	Minor cleanup	2021-10-20 23:35:59 -04:00
Jean-Marc Valin	0e523aa3f4	controllable look-ahead	2021-10-20 23:35:59 -04:00
Jean-Marc Valin	a9bf6cee8a	Don't hardcode the number of bands	2021-10-20 23:35:59 -04:00