Avdpro commited on
Commit
9218cca
·
verified ·
1 Parent(s): 2ef316c

Add native MLX RVC voice-training initialization

Browse files
NOTICE.md CHANGED
@@ -6,5 +6,7 @@
6
  - The bundled functional voice is trained from synthetic audio generated with
7
  the Qwen3-TTS `serena` preset. It is marked as synthetic in the checkpoint
8
  manifest and must not be represented as a recording of a real person.
 
 
9
  - The Package performs local voice conversion only. It does not grant rights
10
  to source recordings supplied by a user.
 
6
  - The bundled functional voice is trained from synthetic audio generated with
7
  the Qwen3-TTS `serena` preset. It is marked as synthetic in the checkpoint
8
  manifest and must not be represented as a recording of a real person.
9
+ - The native MLX training initialization is a safetensors conversion of the
10
+ pinned upstream RVC v2/48 kHz generator and discriminator checkpoints.
11
  - The Package performs local voice conversion only. It does not grant rights
12
  to source recordings supplied by a user.
README.md CHANGED
@@ -10,10 +10,11 @@ tags:
10
  # MLX-RVC Serena e70
11
 
12
  Native MLX RVC v2/48 kHz composite checkpoint for AI2Apps. It contains the
13
- MLX-trained e70 synthetic Serena test voice, its exact retrieval index, and the
14
- shared ContentVec and RMVPE inference components. The checkpoint contains only
15
- safetensors and JSON; it does not require Torch, torchaudio, FAISS, or pickle at
16
- runtime.
 
17
 
18
  The voice training corpus was generated with the `serena` preset of Qwen3-TTS
19
  and is intended as a built-in functional voice, not as the identity of a real
 
10
  # MLX-RVC Serena e70
11
 
12
  Native MLX RVC v2/48 kHz composite checkpoint for AI2Apps. It contains the
13
+ MLX-trained e70 synthetic Serena test voice, its exact retrieval index, the
14
+ shared ContentVec and RMVPE components, and the converted RVC v2/48 kHz
15
+ generator/discriminator initialization used by the native MLX trainer. The
16
+ checkpoint contains only safetensors and JSON; it does not require Torch,
17
+ torchaudio, FAISS, or pickle at runtime.
18
 
19
  The voice training corpus was generated with the `serena` preset of Qwen3-TTS
20
  and is intended as a built-in functional voice, not as the identity of a real
ai2apps-checkpoint.json CHANGED
@@ -6,7 +6,8 @@
6
  "components": {
7
  "voice": "voice",
8
  "contentvec": "contentvec",
9
- "rmvpe": "rmvpe"
 
10
  },
11
  "files": [
12
  {
@@ -59,15 +60,30 @@
59
  "size": 441,
60
  "sha256": "c5087cfb5ed6395e6e151f639cc87eb6bad0994bf06c71d7b765ed83690f5b14"
61
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
62
  {
63
  "path": "README.md",
64
- "size": 881,
65
- "sha256": "2cc683272ac5b4de6d77a6b5376f1066302c1adf8c6e50ca03dce367d259eea7"
66
  },
67
  {
68
  "path": "NOTICE.md",
69
- "size": 521,
70
- "sha256": "a36935e4d1a988f0ae58e4e244db1723ce8eb91d9c97475254c0300c4fac0100"
71
  },
72
  {
73
  "path": "LICENSE",
 
6
  "components": {
7
  "voice": "voice",
8
  "contentvec": "contentvec",
9
+ "rmvpe": "rmvpe",
10
+ "training": "training"
11
  },
12
  "files": [
13
  {
 
60
  "size": 441,
61
  "sha256": "c5087cfb5ed6395e6e151f639cc87eb6bad0994bf06c71d7b765ed83690f5b14"
62
  },
63
+ {
64
+ "path": "training/generator.safetensors",
65
+ "size": 150433888,
66
+ "sha256": "d2979ac5acf184ead3b5c62232d083e72ba7592441509ca88f6a496acca00573"
67
+ },
68
+ {
69
+ "path": "training/discriminator.safetensors",
70
+ "size": 285553012,
71
+ "sha256": "6672bcb014b6b8d1ba87f9bf9caffc4c1cb6f73375184ce4156d13c13437441c"
72
+ },
73
+ {
74
+ "path": "training/manifest.json",
75
+ "size": 905,
76
+ "sha256": "6d3ecfc37a9f8da68577d3a8f1d6b77c90da2497d2be52d2845a1287469ff916"
77
+ },
78
  {
79
  "path": "README.md",
80
+ "size": 970,
81
+ "sha256": "e5cd710ab7ac2329fa200cb57b310ccd3e2d6081284085ff977b46ac7444b1af"
82
  },
83
  {
84
  "path": "NOTICE.md",
85
+ "size": 670,
86
+ "sha256": "1520786160efa529c76e1d614e97f4d22d02af184fc073466f394b290eb0d5fe"
87
  },
88
  {
89
  "path": "LICENSE",
training/discriminator.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6672bcb014b6b8d1ba87f9bf9caffc4c1cb6f73375184ce4156d13c13437441c
3
+ size 285553012
training/generator.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d2979ac5acf184ead3b5c62232d083e72ba7592441509ca88f6a496acca00573
3
+ size 150433888
training/manifest.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "ai2apps.mlx-rvc-training-checkpoint/v1",
3
+ "version": "v2",
4
+ "sample_rate": 48000,
5
+ "uses_f0": true,
6
+ "generator": {
7
+ "source": {
8
+ "name": "f0G48k.pth",
9
+ "sha256": "b5d51f589cc3632d4eae36a315b4179397695042edc01d15312e1bddc2b764a4"
10
+ },
11
+ "weights": {
12
+ "name": "generator.safetensors",
13
+ "sha256": "d2979ac5acf184ead3b5c62232d083e72ba7592441509ca88f6a496acca00573",
14
+ "tensor_count": 423,
15
+ "parameter_count": 37597570,
16
+ "dtype": "float32"
17
+ }
18
+ },
19
+ "discriminator": {
20
+ "source": {
21
+ "name": "f0D48k.pth",
22
+ "sha256": "2269b73c7a4cf34da09aea99274dabf99b2ddb8a42cbfb065fb3c0aa9a2fc748"
23
+ },
24
+ "weights": {
25
+ "name": "discriminator.safetensors",
26
+ "sha256": "6672bcb014b6b8d1ba87f9bf9caffc4c1cb6f73375184ce4156d13c13437441c",
27
+ "tensor_count": 110,
28
+ "parameter_count": 71385417,
29
+ "dtype": "float32"
30
+ }
31
+ }
32
+ }