GunnarThor
commited on
Commit
•
c885237
1
Parent(s):
aa1d68c
add model files
Browse files- README.md +9 -0
- checkpoint-1500000steps.pkl +3 -0
- config.yml +167 -0
- stats.h5 +3 -0
README.md
CHANGED
@@ -1,3 +1,12 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
+
|
5 |
+
This model was trained by Gunnar Thor Örnólfsson in 2023
|
6 |
+
|
7 |
+
It uses a combination of the Talrómur 1 and 2 corpora, totalling 44 speakers (with 4 speakers being held back for evaluation.
|
8 |
+
|
9 |
+
All 8 voices in Talrómur 1 were used.
|
10 |
+
36 out of the 40 voices in Talrómur 2 were used.
|
11 |
+
|
12 |
+
The model was trained for 1.500.000 steps using Tomoki Hayashi's implementation of Style-MelGAN: https://github.com/kan-bayashi/ParallelWaveGAN/
|
checkpoint-1500000steps.pkl
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0fb6b455b89a754472d208e2e33d8536ed426de6c597ae48f27ed91de98425a4
|
3 |
+
size 114014636
|
config.yml
ADDED
@@ -0,0 +1,167 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
allow_cache: false
|
2 |
+
batch_max_steps: 22528
|
3 |
+
batch_size: 32
|
4 |
+
config: conf/style_melgan.v1.yaml
|
5 |
+
dev_dumpdir: dump/dev_1and2/norm
|
6 |
+
dev_feats_scp: null
|
7 |
+
dev_segments: null
|
8 |
+
dev_wav_scp: null
|
9 |
+
discriminator_adv_loss_params:
|
10 |
+
average_by_discriminators: false
|
11 |
+
discriminator_grad_norm: -1
|
12 |
+
discriminator_optimizer_params:
|
13 |
+
betas:
|
14 |
+
- 0.5
|
15 |
+
- 0.9
|
16 |
+
lr: 0.0002
|
17 |
+
weight_decay: 0.0
|
18 |
+
discriminator_optimizer_type: Adam
|
19 |
+
discriminator_params:
|
20 |
+
discriminator_params:
|
21 |
+
bias: true
|
22 |
+
channels: 16
|
23 |
+
downsample_scales:
|
24 |
+
- 4
|
25 |
+
- 4
|
26 |
+
- 4
|
27 |
+
- 1
|
28 |
+
kernel_sizes:
|
29 |
+
- 5
|
30 |
+
- 3
|
31 |
+
max_downsample_channels: 512
|
32 |
+
nonlinear_activation: LeakyReLU
|
33 |
+
nonlinear_activation_params:
|
34 |
+
negative_slope: 0.2
|
35 |
+
out_channels: 1
|
36 |
+
pqmf_params:
|
37 |
+
- - 1
|
38 |
+
- None
|
39 |
+
- None
|
40 |
+
- None
|
41 |
+
- - 2
|
42 |
+
- 62
|
43 |
+
- 0.267
|
44 |
+
- 9.0
|
45 |
+
- - 4
|
46 |
+
- 62
|
47 |
+
- 0.142
|
48 |
+
- 9.0
|
49 |
+
- - 8
|
50 |
+
- 62
|
51 |
+
- 0.07949
|
52 |
+
- 9.0
|
53 |
+
repeats: 4
|
54 |
+
use_weight_norm: true
|
55 |
+
window_sizes:
|
56 |
+
- 512
|
57 |
+
- 1024
|
58 |
+
- 2048
|
59 |
+
- 4096
|
60 |
+
discriminator_scheduler_params:
|
61 |
+
gamma: 0.5
|
62 |
+
milestones:
|
63 |
+
- 200000
|
64 |
+
- 400000
|
65 |
+
- 600000
|
66 |
+
- 800000
|
67 |
+
discriminator_scheduler_type: MultiStepLR
|
68 |
+
discriminator_train_start_steps: 100000
|
69 |
+
discriminator_type: StyleMelGANDiscriminator
|
70 |
+
distributed: false
|
71 |
+
eval_interval_steps: 1000
|
72 |
+
fft_size: 1024
|
73 |
+
fmax: 7600
|
74 |
+
fmin: 80
|
75 |
+
format: hdf5
|
76 |
+
generator_adv_loss_params:
|
77 |
+
average_by_discriminators: false
|
78 |
+
generator_grad_norm: -1
|
79 |
+
generator_optimizer_params:
|
80 |
+
betas:
|
81 |
+
- 0.5
|
82 |
+
- 0.9
|
83 |
+
lr: 0.0001
|
84 |
+
weight_decay: 0.0
|
85 |
+
generator_optimizer_type: Adam
|
86 |
+
generator_params:
|
87 |
+
aux_channels: 80
|
88 |
+
bias: true
|
89 |
+
channels: 64
|
90 |
+
dilation: 2
|
91 |
+
gated_function: softmax
|
92 |
+
in_channels: 128
|
93 |
+
kernel_size: 9
|
94 |
+
noise_upsample_activation: LeakyReLU
|
95 |
+
noise_upsample_activation_params:
|
96 |
+
negative_slope: 0.2
|
97 |
+
noise_upsample_scales:
|
98 |
+
- 11
|
99 |
+
- 2
|
100 |
+
- 2
|
101 |
+
- 2
|
102 |
+
out_channels: 1
|
103 |
+
upsample_mode: nearest
|
104 |
+
upsample_scales:
|
105 |
+
- 2
|
106 |
+
- 2
|
107 |
+
- 2
|
108 |
+
- 2
|
109 |
+
- 2
|
110 |
+
- 2
|
111 |
+
- 2
|
112 |
+
- 2
|
113 |
+
- 1
|
114 |
+
use_weight_norm: true
|
115 |
+
generator_scheduler_params:
|
116 |
+
gamma: 0.5
|
117 |
+
milestones:
|
118 |
+
- 100000
|
119 |
+
- 300000
|
120 |
+
- 500000
|
121 |
+
- 700000
|
122 |
+
- 900000
|
123 |
+
generator_scheduler_type: MultiStepLR
|
124 |
+
generator_type: StyleMelGANGenerator
|
125 |
+
global_gain_scale: 1.0
|
126 |
+
hop_size: 256
|
127 |
+
lambda_adv: 1.0
|
128 |
+
lambda_aux: 1.0
|
129 |
+
log_interval_steps: 100
|
130 |
+
num_mels: 80
|
131 |
+
num_save_intermediate_results: 4
|
132 |
+
num_workers: 2
|
133 |
+
outdir: exp/train_1and2_style_melgan.v1
|
134 |
+
pin_memory: true
|
135 |
+
pretrain: ''
|
136 |
+
rank: 0
|
137 |
+
remove_short_samples: false
|
138 |
+
resume: ''
|
139 |
+
sampling_rate: 22050
|
140 |
+
save_interval_steps: 50000
|
141 |
+
stft_loss_params:
|
142 |
+
fft_sizes:
|
143 |
+
- 1024
|
144 |
+
- 2048
|
145 |
+
- 512
|
146 |
+
hop_sizes:
|
147 |
+
- 120
|
148 |
+
- 240
|
149 |
+
- 50
|
150 |
+
win_lengths:
|
151 |
+
- 600
|
152 |
+
- 1200
|
153 |
+
- 240
|
154 |
+
window: hann_window
|
155 |
+
train_dumpdir: dump/train_1and2/norm
|
156 |
+
train_feats_scp: null
|
157 |
+
train_max_steps: 1500000
|
158 |
+
train_segments: null
|
159 |
+
train_wav_scp: null
|
160 |
+
trim_frame_size: 1024
|
161 |
+
trim_hop_size: 256
|
162 |
+
trim_silence: false
|
163 |
+
trim_threshold_in_db: 60
|
164 |
+
verbose: 1
|
165 |
+
version: 0.6.0
|
166 |
+
win_length: null
|
167 |
+
window: hann
|
stats.h5
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:54f1528f410ec426951fc6e30ec4cc266805f8477599c6ad2a26f07e2797bb71
|
3 |
+
size 4736
|