lengyue233 Asteriski commited on
Commit
89fac6a
1 Parent(s): 1936ce4

Add ALYS + voice provider credits (#6)

Browse files

- Add ALYS + voice provider credits (49a207cd9b3b36417f51889fdff3c21440ee4c2d)


Co-authored-by: Aster <[email protected]>

Files changed (3) hide show
  1. checkpoints/ALYS.ckpt +3 -0
  2. config.yaml +14 -6
  3. configs/ALYS.py +48 -0
checkpoints/ALYS.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:475fb4b9f56f8d14812ee78e4d5b39b2e58f60d8d0350e84587ed21a9ba96fca
3
+ size 409439345
config.yaml CHANGED
@@ -18,7 +18,7 @@ models:
18
  checkpoint: checkpoints/Kiritan.ckpt
19
  readme: |
20
  This model is trained on the Tohoku Kiritan dataset and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
21
- It has a cute, yet powerful voice.
22
  default_speaker: "kiritan"
23
 
24
  - name: "Tohoku Itako (Feminine)"
@@ -26,7 +26,7 @@ models:
26
  checkpoint: checkpoints/Itako.ckpt
27
  readme: |
28
  This model is trained on the Tohoku Itako dataset and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
29
- It has a bright and whispery voice.
30
  default_speaker: "itako"
31
 
32
  - name: "No.7 (Feminine)"
@@ -34,7 +34,7 @@ models:
34
  checkpoint: checkpoints/Seven.ckpt
35
  readme: |
36
  This model is trained on the No.7 dataset and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
37
- It has a strong and sharp voice.
38
  default_speaker: "seven"
39
 
40
  - name: "Yoko (Feminine)"
@@ -66,7 +66,7 @@ models:
66
  checkpoint: checkpoints/Ritsu.ckpt
67
  readme: |
68
  This model is trained on the Namine Ritsu ENUNU Dataset and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
69
- It has a powerful and throaty voice.
70
  default_speaker: "ritsu"
71
 
72
  - name: "S (Masculine)"
@@ -90,5 +90,13 @@ models:
90
  checkpoint: checkpoints/Azure.ckpt
91
  readme: |
92
  This model is trained on a dataset known as Azure Cobalt and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
93
- It has a stable, mature voice.
94
- default_speaker: "azure"
 
 
 
 
 
 
 
 
 
18
  checkpoint: checkpoints/Kiritan.ckpt
19
  readme: |
20
  This model is trained on the Tohoku Kiritan dataset and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
21
+ It has a cute, yet powerful voice. CV: Akaneya Himika
22
  default_speaker: "kiritan"
23
 
24
  - name: "Tohoku Itako (Feminine)"
 
26
  checkpoint: checkpoints/Itako.ckpt
27
  readme: |
28
  This model is trained on the Tohoku Itako dataset and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
29
+ It has a bright and whispery voice. CV: Kido Ibuki
30
  default_speaker: "itako"
31
 
32
  - name: "No.7 (Feminine)"
 
34
  checkpoint: checkpoints/Seven.ckpt
35
  readme: |
36
  This model is trained on the No.7 dataset and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
37
+ It has a strong and sharp voice. CV: Koiwai Kotori
38
  default_speaker: "seven"
39
 
40
  - name: "Yoko (Feminine)"
 
66
  checkpoint: checkpoints/Ritsu.ckpt
67
  readme: |
68
  This model is trained on the Namine Ritsu ENUNU Dataset and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
69
+ It has a powerful and throaty voice. CV: Canon
70
  default_speaker: "ritsu"
71
 
72
  - name: "S (Masculine)"
 
90
  checkpoint: checkpoints/Azure.ckpt
91
  readme: |
92
  This model is trained on a dataset known as Azure Cobalt and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
93
+ It has a stable, mature voice. CV: Aster
94
+ default_speaker: "azure"
95
+
96
+ - name: "ALYS (Feminine)"
97
+ config: configs/ALYS.py
98
+ checkpoint: checkpoints/ALYS.ckpt
99
+ readme: |
100
+ This model is trained on the ALYS DB 001 JPN dataset, originally produced by Voxwave and released under the [GPL-3.0](https://choosealicense.com/licenses/gpl-3.0/) license.
101
+ It has a slightly soft voice. CV: Poucet
102
+ default_speaker: "ALYS"
configs/ALYS.py ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from fish_diffusion.datasets.hifisinger import HiFiSVCDataset
2
+ from fish_diffusion.datasets.utils import get_datasets_from_subfolder
3
+
4
+ _base_ = [
5
+ "./_base_/archs/hifi_svc.py",
6
+ "./_base_/trainers/base.py",
7
+ "./_base_/schedulers/exponential.py",
8
+ "./_base_/datasets/hifi_svc.py",
9
+ ]
10
+
11
+ speaker_mapping = {
12
+ "ALYS": 0,
13
+ }
14
+
15
+ model = dict(
16
+ type="HiFiSVC",
17
+ speaker_encoder=dict(
18
+ input_size=len(speaker_mapping),
19
+ ),
20
+ )
21
+
22
+ preprocessing = dict(
23
+ text_features_extractor=dict(
24
+ type="ContentVec",
25
+ ),
26
+ pitch_extractor=dict(
27
+ type="CrepePitchExtractor",
28
+ keep_zeros=False,
29
+ f0_min=40.0,
30
+ f0_max=1600.0,
31
+ ),
32
+ energy_extractor=dict(
33
+ type="RMSEnergyExtractor",
34
+ ),
35
+ augmentations=[
36
+ dict(
37
+ type="FixedPitchShifting",
38
+ key_shifts=[-5.0, 5.0],
39
+ probability=0.75,
40
+ ),
41
+ ],
42
+ )
43
+
44
+ trainer = dict(
45
+ # Disable gradient clipping, which is not supported by custom optimization
46
+ gradient_clip_val=None,
47
+ max_steps=1000000,
48
+ )