Brandon Thomas
55947770cd
Model outputs audio, but all zeroes right now
2019-12-26 03:39:16 -05:00
Brandon Thomas
346855955c
Useful hacky script for jit containerizing mels/tensors
...
This seems to be the only(?) way of loading single tensors into
libtorch/tch.rs. It's a hack: load the tensor, pack it into a container
module, then define a forward method that returns the wrapped tensor.
2019-12-26 02:02:25 -05:00
Brandon Thomas
b9efb9db60
This might be a new approach...
2019-12-26 01:21:41 -05:00
Brandon Thomas
99d13ed283
light refactor
2019-12-25 17:01:28 -05:00
Brandon Thomas
615595c0d3
Add some mels for testing
2019-12-25 16:49:01 -05:00
Brandon Thomas
194c220ad2
Move melgan scripts to sub-project
2019-12-25 16:32:43 -05:00
Brandon Thomas
282e05340f
Now the two Tacotron projects work with the Rust Workspace
2019-12-25 01:54:02 -05:00
Brandon Thomas
a03966c209
Refactor into "workspace" project. VC sidecar works again!
...
Tensorflow Tacotron and Torch Tacotron+Melgan are broken (the latter
hasn't worked yet anyhow)
2019-12-25 01:31:19 -05:00
Brandon Thomas
56ba5fbaaa
wip (I need to restructure this entire codebase)
2019-12-25 00:05:57 -05:00
Brandon Thomas
da9ffb03bf
closer match py JIT output; can load/eval from rust
2019-12-23 16:40:45 -05:00
Brandon Thomas
5d0d22c5f8
Adds the ability to load PyTorch models (read more)
...
- Python script de-cudas the models and exports them in the
understandable "zip" format. (The models output by MelGan, Tacotron, etc
do not conform to something Rust's libtorch bindings can understand by
default)
- Adds the Rust code to load but not the cargo deps since that needs a
refactor. The version is `tch = "0.1.3"`
2019-12-23 15:01:39 -05:00
Brandon Thomas
b77de37029
Merge branch 'master' of github.com:echelon/voder
2019-12-16 01:33:15 -05:00
Brandon Thomas
01896b7f9a
crop image, add lots of comments
2019-12-14 04:06:45 -05:00
Brandon Thomas
3bf8f3a044
this looks much closer to a spectrogram
2019-12-14 03:28:08 -05:00
Brandon Thomas
4408fd664e
image visualization (WIP)
2019-12-14 03:24:23 -05:00
Brandon Thomas
ed1f63c9a3
Pow_1 looks like a good point - very fast on CPU
2019-12-14 02:43:10 -05:00
Brandon Thomas
d1d3957310
Merge branch 'master' of github.com:echelon/voder
2019-12-13 03:47:06 -05:00
Brandon Thomas
26cb3bdaaf
(temporary GPU build break)
2019-12-13 03:46:46 -05:00
Brandon Thomas
5431902f8d
Merge branch 'master' of github.com:echelon/voder
2019-12-13 03:45:09 -05:00
Brandon Thomas
011690ce8b
(temporary GPU build break)
2019-12-13 03:44:54 -05:00
Brandon Thomas
e5f50002da
(temporary GPU build break)
2019-12-13 03:35:50 -05:00
Brandon Thomas
1958622d8c
generalize function
2019-12-11 04:24:22 -05:00
Brandon Thomas
95561e2568
signal reconstituted, but a bit loud
2019-12-11 04:14:48 -05:00
Brandon Thomas
774789c607
comment
2019-12-11 03:11:01 -05:00
Brandon Thomas
06d0bbcbba
spacing cleaup (minor), added cleaned sentence
2019-12-11 03:08:30 -05:00
Brandon Thomas
9dd998714a
This runs!
2019-12-11 02:52:47 -05:00
Brandon Thomas
15570be82d
Inputs and tensors appear to be correct - now need to get data/dims
2019-12-11 02:42:33 -05:00
Brandon Thomas
dd9362352e
Add 2nd binary to load tacotron (Keith Ito) model
2019-12-11 01:16:29 -05:00
Brandon Thomas
dee59f5f4c
Add 2nd binary for tacotron
2019-12-10 23:59:50 -05:00
Brandon Thomas
32a54e395d
record before (or right during) activating edge.
...
This keeps the beginning audio segment from being choppy.
2019-11-08 03:24:15 -05:00
Brandon Thomas
354c6a7261
Move sidecar out to own module
2019-11-08 02:29:12 -05:00
Brandon Thomas
37d8c9ec85
Move CPAL mic/speaker into new module
2019-11-08 01:56:57 -05:00
Brandon Thomas
16639c7bed
README multi-gpu, CycleGAN sample rate cleanup
2019-11-08 01:23:30 -05:00
Brandon Thomas
7f325513fb
This is almost realtime and should work in conversation
2019-10-21 04:11:03 -04:00
Brandon Thomas
89d3ae6362
heuristic for pausing the mic
2019-10-21 04:08:09 -04:00
Brandon Thomas
ed05ab1551
well, this is as best as the model does so far...
2019-10-21 03:37:08 -04:00
Brandon Thomas
fe5e15f206
This sounds good coming back down the pipe...
2019-10-21 03:23:01 -04:00
Brandon Thomas
21bc256478
This might be right up until sending back. Might just need an upsample!
...
Original audio - sounds good
Downsampled audio - sounds good
Model output - sounds good (might even be closer to trump, didn't test
enough)
Voder audio - squeaky. needs upsample probably
2019-10-21 03:18:02 -04:00
Brandon Thomas
4155fa4e36
this is a good shape for debugging
2019-10-21 03:12:26 -04:00
Brandon Thomas
f096ae07bf
Changing proto schema. Degradation in quality.
...
I had something really close. That might not have found it into a
commit between now and the last commit :(
2019-10-21 02:58:05 -04:00
Brandon Thomas
af20a57e8a
THIS EVALUATES THE MODEL IN REALTIME +3-6sec delay
...
Still doesn't sound right. Probably a sample rate issue.
2019-10-20 03:10:21 -04:00
Brandon Thomas
d63007ac9f
Experimenting - trying to figure out why audio sounds so bad
2019-10-20 02:50:18 -04:00
Brandon Thomas
7e33d9c7ca
Add my demo_server.py implementation from CycleGAN
2019-10-20 02:18:51 -04:00
Brandon Thomas
b38c96ae2f
Protobufs UP and DOWN. Sounds great!
2019-10-20 02:04:41 -04:00
Brandon Thomas
05b703068a
THIS SOUNDS FANTASTIC (no transform yet, 1:1 audio)
2019-10-20 01:51:55 -04:00
Brandon Thomas
6df1e8b072
Two way works over protobuf! Woo! This will be so much better...
2019-10-20 01:36:12 -04:00
Brandon Thomas
dc9983e45b
Protos in python
2019-10-20 01:17:25 -04:00
Brandon Thomas
45a8dc446e
Better proto pipeline
2019-10-20 00:48:18 -04:00
Brandon Thomas
1ae8aafd5a
Add protos. Temporarily commented out TensorFlow pieces
2019-10-20 00:38:04 -04:00
Brandon Thomas
47022cdeeb
Copy relevant CycleGAN code here.
...
master @ 1ca8b078c12d915bca1f57e822859fa20bd439c8
Date: Tue Oct 15 10:50:12 2019 -0400
2019-10-20 00:11:53 -04:00
Brandon Thomas