Commit
·
887e50c
1
Parent(s):
590b29f
update
Browse files- config.py +218 -0
- inference_webui.py +2 -2
- requirements.txt +2 -1
- tools/AP_BWE_main/24kto48k/readme.txt +11 -0
- tools/AP_BWE_main/LICENSE +21 -0
- tools/AP_BWE_main/README.md +91 -0
- tools/AP_BWE_main/datasets1/__init__.py +1 -0
- tools/AP_BWE_main/datasets1/dataset.py +108 -0
- tools/AP_BWE_main/models/__init__.py +1 -0
- tools/AP_BWE_main/models/model.py +464 -0
- tools/assets.py +70 -0
- tools/audio_sr.py +50 -0
- weight.json +1 -0
config.py
ADDED
@@ -0,0 +1,218 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import re
|
3 |
+
import sys
|
4 |
+
|
5 |
+
import torch
|
6 |
+
|
7 |
+
from tools.i18n.i18n import I18nAuto
|
8 |
+
|
9 |
+
i18n = I18nAuto(language=os.environ.get("language", "Auto"))
|
10 |
+
|
11 |
+
|
12 |
+
pretrained_sovits_name = {
|
13 |
+
"v1": "pretrained_models/s2G488k.pth",
|
14 |
+
"v2": "pretrained_models/gsv-v2final-pretrained/s2G2333k.pth",
|
15 |
+
"v3": "pretrained_models/s2Gv3.pth", ###v3v4还要检查vocoder,算了。。。
|
16 |
+
"v4": "pretrained_models/gsv-v4-pretrained/s2Gv4.pth",
|
17 |
+
"v2Pro": "pretrained_models/v2Pro/s2Gv2Pro.pth",
|
18 |
+
"v2ProPlus": "pretrained_models/v2Pro/s2Gv2ProPlus.pth",
|
19 |
+
}
|
20 |
+
|
21 |
+
pretrained_gpt_name = {
|
22 |
+
"v1": "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt",
|
23 |
+
"v2": "pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt",
|
24 |
+
"v3": "pretrained_models/s1v3.ckpt",
|
25 |
+
"v4": "pretrained_models/s1v3.ckpt",
|
26 |
+
"v2Pro": "pretrained_models/s1v3.ckpt",
|
27 |
+
"v2ProPlus": "pretrained_models/s1v3.ckpt",
|
28 |
+
}
|
29 |
+
name2sovits_path = {
|
30 |
+
# i18n("不训练直接推v1底模!"): "pretrained_models/s2G488k.pth",
|
31 |
+
i18n("不训练直接推v2底模!"): "pretrained_models/gsv-v2final-pretrained/s2G2333k.pth",
|
32 |
+
# i18n("不训练直接推v3底模!"): "pretrained_models/s2Gv3.pth",
|
33 |
+
# i18n("不训练直接推v4底模!"): "pretrained_models/gsv-v4-pretrained/s2Gv4.pth",
|
34 |
+
i18n("不训练直接推v2Pro底模!"): "pretrained_models/v2Pro/s2Gv2Pro.pth",
|
35 |
+
i18n("不训练直接推v2ProPlus底模!"): "pretrained_models/v2Pro/s2Gv2ProPlus.pth",
|
36 |
+
}
|
37 |
+
name2gpt_path = {
|
38 |
+
# i18n("不训练直接推v1底模!"):"pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt",
|
39 |
+
i18n(
|
40 |
+
"不训练直接推v2底模!"
|
41 |
+
): "pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt",
|
42 |
+
i18n("不训练直接推v3底模!"): "pretrained_models/s1v3.ckpt",
|
43 |
+
}
|
44 |
+
SoVITS_weight_root = [
|
45 |
+
"SoVITS_weights",
|
46 |
+
"SoVITS_weights_v2",
|
47 |
+
"SoVITS_weights_v3",
|
48 |
+
"SoVITS_weights_v4",
|
49 |
+
"SoVITS_weights_v2Pro",
|
50 |
+
"SoVITS_weights_v2ProPlus",
|
51 |
+
]
|
52 |
+
GPT_weight_root = [
|
53 |
+
"GPT_weights",
|
54 |
+
"GPT_weights_v2",
|
55 |
+
"GPT_weights_v3",
|
56 |
+
"GPT_weights_v4",
|
57 |
+
"GPT_weights_v2Pro",
|
58 |
+
"GPT_weights_v2ProPlus",
|
59 |
+
]
|
60 |
+
SoVITS_weight_version2root = {
|
61 |
+
"v1": "SoVITS_weights",
|
62 |
+
"v2": "SoVITS_weights_v2",
|
63 |
+
"v3": "SoVITS_weights_v3",
|
64 |
+
"v4": "SoVITS_weights_v4",
|
65 |
+
"v2Pro": "SoVITS_weights_v2Pro",
|
66 |
+
"v2ProPlus": "SoVITS_weights_v2ProPlus",
|
67 |
+
}
|
68 |
+
GPT_weight_version2root = {
|
69 |
+
"v1": "GPT_weights",
|
70 |
+
"v2": "GPT_weights_v2",
|
71 |
+
"v3": "GPT_weights_v3",
|
72 |
+
"v4": "GPT_weights_v4",
|
73 |
+
"v2Pro": "GPT_weights_v2Pro",
|
74 |
+
"v2ProPlus": "GPT_weights_v2ProPlus",
|
75 |
+
}
|
76 |
+
|
77 |
+
|
78 |
+
def custom_sort_key(s):
|
79 |
+
# 使用正则表达式提取字符串中的数字部分和非数字部分
|
80 |
+
parts = re.split("(\d+)", s)
|
81 |
+
# 将数字部分转换为整数,非数字部分保持不变
|
82 |
+
parts = [int(part) if part.isdigit() else part for part in parts]
|
83 |
+
return parts
|
84 |
+
|
85 |
+
|
86 |
+
def get_weights_names():
|
87 |
+
SoVITS_names = []
|
88 |
+
for key in name2sovits_path:
|
89 |
+
if os.path.exists(name2sovits_path[key]):
|
90 |
+
SoVITS_names.append(key)
|
91 |
+
for path in SoVITS_weight_root:
|
92 |
+
if not os.path.exists(path):
|
93 |
+
continue
|
94 |
+
for name in os.listdir(path):
|
95 |
+
if name.endswith(".pth"):
|
96 |
+
SoVITS_names.append("%s/%s" % (path, name))
|
97 |
+
if not SoVITS_names:
|
98 |
+
SoVITS_names = [""]
|
99 |
+
GPT_names = []
|
100 |
+
for key in name2gpt_path:
|
101 |
+
if os.path.exists(name2gpt_path[key]):
|
102 |
+
GPT_names.append(key)
|
103 |
+
for path in GPT_weight_root:
|
104 |
+
if not os.path.exists(path):
|
105 |
+
continue
|
106 |
+
for name in os.listdir(path):
|
107 |
+
if name.endswith(".ckpt"):
|
108 |
+
GPT_names.append("%s/%s" % (path, name))
|
109 |
+
SoVITS_names = sorted(SoVITS_names, key=custom_sort_key)
|
110 |
+
GPT_names = sorted(GPT_names, key=custom_sort_key)
|
111 |
+
if not GPT_names:
|
112 |
+
GPT_names = [""]
|
113 |
+
return SoVITS_names, GPT_names
|
114 |
+
|
115 |
+
|
116 |
+
def change_choices():
|
117 |
+
SoVITS_names, GPT_names = get_weights_names()
|
118 |
+
return {"choices": SoVITS_names, "__type__": "update"}, {
|
119 |
+
"choices": GPT_names,
|
120 |
+
"__type__": "update",
|
121 |
+
}
|
122 |
+
|
123 |
+
|
124 |
+
# 推理用的指定模型
|
125 |
+
sovits_path = ""
|
126 |
+
gpt_path = ""
|
127 |
+
is_half_str = os.environ.get("is_half", "True")
|
128 |
+
is_half = True if is_half_str.lower() == "true" else False
|
129 |
+
is_share_str = os.environ.get("is_share", "False")
|
130 |
+
is_share = True if is_share_str.lower() == "true" else False
|
131 |
+
|
132 |
+
cnhubert_path = "pretrained_models/chinese-hubert-base"
|
133 |
+
bert_path = "pretrained_models/chinese-roberta-wwm-ext-large"
|
134 |
+
pretrained_sovits_path = "pretrained_models/s2G488k.pth"
|
135 |
+
pretrained_gpt_path = "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
|
136 |
+
|
137 |
+
exp_root = "logs"
|
138 |
+
python_exec = sys.executable or "python"
|
139 |
+
|
140 |
+
webui_port_main = 9874
|
141 |
+
webui_port_uvr5 = 9873
|
142 |
+
webui_port_infer_tts = 9872
|
143 |
+
webui_port_subfix = 9871
|
144 |
+
|
145 |
+
api_port = 9880
|
146 |
+
|
147 |
+
|
148 |
+
# Thanks to the contribution of @Karasukaigan and @XXXXRT666
|
149 |
+
def get_device_dtype_sm(idx: int) -> tuple[torch.device, torch.dtype, float, float]:
|
150 |
+
cpu = torch.device("cpu")
|
151 |
+
cuda = torch.device(f"cuda:{idx}")
|
152 |
+
if not torch.cuda.is_available():
|
153 |
+
return cpu, torch.float32, 0.0, 0.0
|
154 |
+
device_idx = idx
|
155 |
+
capability = torch.cuda.get_device_capability(device_idx)
|
156 |
+
name = torch.cuda.get_device_name(device_idx)
|
157 |
+
mem_bytes = torch.cuda.get_device_properties(device_idx).total_memory
|
158 |
+
mem_gb = mem_bytes / (1024**3) + 0.4
|
159 |
+
major, minor = capability
|
160 |
+
sm_version = major + minor / 10.0
|
161 |
+
is_16_series = bool(re.search(r"16\d{2}", name)) and sm_version == 7.5
|
162 |
+
if mem_gb < 4 or sm_version < 5.3:
|
163 |
+
return cpu, torch.float32, 0.0, 0.0
|
164 |
+
if sm_version == 6.1 or is_16_series == True:
|
165 |
+
return cuda, torch.float32, sm_version, mem_gb
|
166 |
+
if sm_version > 6.1:
|
167 |
+
return cuda, torch.float16, sm_version, mem_gb
|
168 |
+
return cpu, torch.float32, 0.0, 0.0
|
169 |
+
|
170 |
+
|
171 |
+
IS_GPU = True
|
172 |
+
GPU_INFOS: list[str] = []
|
173 |
+
GPU_INDEX: set[int] = set()
|
174 |
+
GPU_COUNT = torch.cuda.device_count()
|
175 |
+
CPU_INFO: str = "0\tCPU " + i18n("CPU训练,较慢")
|
176 |
+
tmp: list[tuple[torch.device, torch.dtype, float, float]] = []
|
177 |
+
memset: set[float] = set()
|
178 |
+
|
179 |
+
for i in range(max(GPU_COUNT, 1)):
|
180 |
+
tmp.append(get_device_dtype_sm(i))
|
181 |
+
|
182 |
+
for j in tmp:
|
183 |
+
device = j[0]
|
184 |
+
memset.add(j[3])
|
185 |
+
if device.type != "cpu":
|
186 |
+
GPU_INFOS.append(f"{device.index}\t{torch.cuda.get_device_name(device.index)}")
|
187 |
+
GPU_INDEX.add(device.index)
|
188 |
+
|
189 |
+
if not GPU_INFOS:
|
190 |
+
IS_GPU = False
|
191 |
+
GPU_INFOS.append(CPU_INFO)
|
192 |
+
GPU_INDEX.add(0)
|
193 |
+
|
194 |
+
infer_device = max(tmp, key=lambda x: (x[2], x[3]))[0]
|
195 |
+
is_half = any(dtype == torch.float16 for _, dtype, _, _ in tmp)
|
196 |
+
|
197 |
+
|
198 |
+
class Config:
|
199 |
+
def __init__(self):
|
200 |
+
self.sovits_path = sovits_path
|
201 |
+
self.gpt_path = gpt_path
|
202 |
+
self.is_half = is_half
|
203 |
+
|
204 |
+
self.cnhubert_path = cnhubert_path
|
205 |
+
self.bert_path = bert_path
|
206 |
+
self.pretrained_sovits_path = pretrained_sovits_path
|
207 |
+
self.pretrained_gpt_path = pretrained_gpt_path
|
208 |
+
|
209 |
+
self.exp_root = exp_root
|
210 |
+
self.python_exec = python_exec
|
211 |
+
self.infer_device = infer_device
|
212 |
+
|
213 |
+
self.webui_port_main = webui_port_main
|
214 |
+
self.webui_port_uvr5 = webui_port_uvr5
|
215 |
+
self.webui_port_infer_tts = webui_port_infer_tts
|
216 |
+
self.webui_port_subfix = webui_port_subfix
|
217 |
+
|
218 |
+
self.api_port = api_port
|
inference_webui.py
CHANGED
@@ -29,7 +29,7 @@ logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
|
|
29 |
logging.getLogger("multipart.multipart").setLevel(logging.ERROR)
|
30 |
warnings.simplefilter(action="ignore", category=FutureWarning)
|
31 |
|
32 |
-
version = model_version = os.environ.get("version", "
|
33 |
|
34 |
from config import change_choices, get_weights_names, name2gpt_path, name2sovits_path
|
35 |
|
@@ -88,7 +88,7 @@ cnhubert.cnhubert_base_path = cnhubert_base_path
|
|
88 |
|
89 |
import random
|
90 |
|
91 |
-
from
|
92 |
|
93 |
|
94 |
def set_seed(seed):
|
|
|
29 |
logging.getLogger("multipart.multipart").setLevel(logging.ERROR)
|
30 |
warnings.simplefilter(action="ignore", category=FutureWarning)
|
31 |
|
32 |
+
version = model_version = os.environ.get("version", "v2ProPlus")
|
33 |
|
34 |
from config import change_choices, get_weights_names, name2gpt_path, name2sovits_path
|
35 |
|
|
|
88 |
|
89 |
import random
|
90 |
|
91 |
+
from module.models import Generator, SynthesizerTrn, SynthesizerTrnV3
|
92 |
|
93 |
|
94 |
def set_seed(seed):
|
requirements.txt
CHANGED
@@ -33,4 +33,5 @@ torch==2.4
|
|
33 |
pydantic<=2.10.6
|
34 |
torchmetrics<=1.5
|
35 |
fast_langdetect
|
36 |
-
split_lang
|
|
|
|
33 |
pydantic<=2.10.6
|
34 |
torchmetrics<=1.5
|
35 |
fast_langdetect
|
36 |
+
split_lang
|
37 |
+
peft
|
tools/AP_BWE_main/24kto48k/readme.txt
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
For the inference of the v3 model, if you find that the generated audio sounds somewhat muffled, you can try using this audio super-resolution model.
|
2 |
+
对于v3模型的推理,如果你发现生成的音频比较闷,可以尝试这个音频超分模型。
|
3 |
+
|
4 |
+
put g_24kto48k.zip and config.json in this folder
|
5 |
+
把g_24kto48k.zip and config.json下到这个文件夹
|
6 |
+
|
7 |
+
download link 下载链接:
|
8 |
+
https://drive.google.com/drive/folders/1IIYTf2zbJWzelu4IftKD6ooHloJ8mnZF?usp=share_link
|
9 |
+
|
10 |
+
audio sr project page 音频超分项目主页:
|
11 |
+
https://github.com/yxlu-0102/AP-BWE
|
tools/AP_BWE_main/LICENSE
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
MIT License
|
2 |
+
|
3 |
+
Copyright (c) 2023 Ye-Xin Lu
|
4 |
+
|
5 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
6 |
+
of this software and associated documentation files (the "Software"), to deal
|
7 |
+
in the Software without restriction, including without limitation the rights
|
8 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
9 |
+
copies of the Software, and to permit persons to whom the Software is
|
10 |
+
furnished to do so, subject to the following conditions:
|
11 |
+
|
12 |
+
The above copyright notice and this permission notice shall be included in all
|
13 |
+
copies or substantial portions of the Software.
|
14 |
+
|
15 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
16 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
17 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
18 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
19 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
20 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
21 |
+
SOFTWARE.
|
tools/AP_BWE_main/README.md
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Towards High-Quality and Efficient Speech Bandwidth Extension with Parallel Amplitude and Phase Prediction
|
2 |
+
### Ye-Xin Lu, Yang Ai, Hui-Peng Du, Zhen-Hua Ling
|
3 |
+
|
4 |
+
**Abstract:**
|
5 |
+
Speech bandwidth extension (BWE) refers to widening the frequency bandwidth range of speech signals, enhancing the speech quality towards brighter and fuller.
|
6 |
+
This paper proposes a generative adversarial network (GAN) based BWE model with parallel prediction of Amplitude and Phase spectra, named AP-BWE, which achieves both high-quality and efficient wideband speech waveform generation.
|
7 |
+
The proposed AP-BWE generator is entirely based on convolutional neural networks (CNNs).
|
8 |
+
It features a dual-stream architecture with mutual interaction, where the amplitude stream and the phase stream communicate with each other and respectively extend the high-frequency components from the input narrowband amplitude and phase spectra.
|
9 |
+
To improve the naturalness of the extended speech signals, we employ a multi-period discriminator at the waveform level and design a pair of multi-resolution amplitude and phase discriminators at the spectral level, respectively.
|
10 |
+
Experimental results demonstrate that our proposed AP-BWE achieves state-of-the-art performance in terms of speech quality for BWE tasks targeting sampling rates of both 16 kHz and 48 kHz.
|
11 |
+
In terms of generation efficiency, due to the all-convolutional architecture and all-frame-level operations, the proposed AP-BWE can generate 48 kHz waveform samples 292.3 times faster than real-time on a single RTX 4090 GPU and 18.1 times faster than real-time on a single CPU.
|
12 |
+
Notably, to our knowledge, AP-BWE is the first to achieve the direct extension of the high-frequency phase spectrum, which is beneficial for improving the effectiveness of existing BWE methods.
|
13 |
+
|
14 |
+
**We provide our implementation as open source in this repository. Audio samples can be found at the [demo website](http://yxlu-0102.github.io/AP-BWE).**
|
15 |
+
|
16 |
+
|
17 |
+
## Pre-requisites
|
18 |
+
0. Python >= 3.9.
|
19 |
+
0. Clone this repository.
|
20 |
+
0. Install python requirements. Please refer [requirements.txt](requirements.txt).
|
21 |
+
0. Download datasets
|
22 |
+
1. Download and extract the [VCTK-0.92 dataset](https://datashare.ed.ac.uk/handle/10283/3443), and move its `wav48` directory into [VCTK-Corpus-0.92](VCTK-Corpus-0.92) and rename it as `wav48_origin`.
|
23 |
+
1. Trim the silence of the dataset, and the trimmed files will be saved to `wav48_silence_trimmed`.
|
24 |
+
```
|
25 |
+
cd VCTK-Corpus-0.92
|
26 |
+
python flac2wav.py
|
27 |
+
```
|
28 |
+
1. Move all the trimmed training files from `wav48_silence_trimmed` to [wav48/train](wav48/train) following the indexes in [training.txt](VCTK-Corpus-0.92/training.txt), and move all the untrimmed test files from `wav48_origin` to [wav48/test](wav48/test) following the indexes in [test.txt](VCTK-Corpus-0.92/test.txt).
|
29 |
+
|
30 |
+
## Training
|
31 |
+
```
|
32 |
+
cd train
|
33 |
+
CUDA_VISIBLE_DEVICES=0 python train_16k.py --config [config file path]
|
34 |
+
CUDA_VISIBLE_DEVICES=0 python train_48k.py --config [config file path]
|
35 |
+
```
|
36 |
+
Checkpoints and copies of the configuration file are saved in the `cp_model` directory by default.<br>
|
37 |
+
You can change the path by using the `--checkpoint_path` option.
|
38 |
+
Here is an example:
|
39 |
+
```
|
40 |
+
CUDA_VISIBLE_DEVICES=0 python train_16k.py --config ../configs/config_2kto16k.json --checkpoint_path ../checkpoints/AP-BWE_2kto16k
|
41 |
+
```
|
42 |
+
|
43 |
+
## Inference
|
44 |
+
```
|
45 |
+
cd inference
|
46 |
+
python inference_16k.py --checkpoint_file [generator checkpoint file path]
|
47 |
+
python inference_48k.py --checkpoint_file [generator checkpoint file path]
|
48 |
+
```
|
49 |
+
You can download the [pretrained weights](https://drive.google.com/drive/folders/1IIYTf2zbJWzelu4IftKD6ooHloJ8mnZF?usp=share_link) we provide and move all the files to the `checkpoints` directory.
|
50 |
+
<br>
|
51 |
+
Generated wav files are saved in `generated_files` by default.
|
52 |
+
You can change the path by adding `--output_dir` option.
|
53 |
+
Here is an example:
|
54 |
+
```
|
55 |
+
python inference_16k.py --checkpoint_file ../checkpoints/2kto16k/g_2kto16k --output_dir ../generated_files/2kto16k
|
56 |
+
```
|
57 |
+
|
58 |
+
## Model Structure
|
59 |
+

|
60 |
+
|
61 |
+
## Comparison with other speech BWE methods
|
62 |
+
### 2k/4k/8kHz to 16kHz
|
63 |
+
<p align="center">
|
64 |
+
<img src="Figures/table_16k.png" alt="comparison" width="90%"/>
|
65 |
+
</p>
|
66 |
+
|
67 |
+
### 8k/12k/16/24kHz to 16kHz
|
68 |
+
<p align="center">
|
69 |
+
<img src="Figures/table_48k.png" alt="comparison" width="100%"/>
|
70 |
+
</p>
|
71 |
+
|
72 |
+
## Acknowledgements
|
73 |
+
We referred to [HiFi-GAN](https://github.com/jik876/hifi-gan) and [NSPP](https://github.com/YangAi520/NSPP) to implement this.
|
74 |
+
|
75 |
+
## Citation
|
76 |
+
```
|
77 |
+
@article{lu2024towards,
|
78 |
+
title={Towards high-quality and efficient speech bandwidth extension with parallel amplitude and phase prediction},
|
79 |
+
author={Lu, Ye-Xin and Ai, Yang and Du, Hui-Peng and Ling, Zhen-Hua},
|
80 |
+
journal={arXiv preprint arXiv:2401.06387},
|
81 |
+
year={2024}
|
82 |
+
}
|
83 |
+
|
84 |
+
@inproceedings{lu2024multi,
|
85 |
+
title={Multi-Stage Speech Bandwidth Extension with Flexible Sampling Rate Control},
|
86 |
+
author={Lu, Ye-Xin and Ai, Yang and Sheng, Zheng-Yan and Ling, Zhen-Hua},
|
87 |
+
booktitle={Proc. Interspeech},
|
88 |
+
pages={2270--2274},
|
89 |
+
year={2024}
|
90 |
+
}
|
91 |
+
```
|
tools/AP_BWE_main/datasets1/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
|
tools/AP_BWE_main/datasets1/dataset.py
ADDED
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import random
|
3 |
+
import torch
|
4 |
+
import torchaudio
|
5 |
+
import torch.utils.data
|
6 |
+
import torchaudio.functional as aF
|
7 |
+
|
8 |
+
|
9 |
+
def amp_pha_stft(audio, n_fft, hop_size, win_size, center=True):
|
10 |
+
hann_window = torch.hann_window(win_size).to(audio.device)
|
11 |
+
stft_spec = torch.stft(
|
12 |
+
audio,
|
13 |
+
n_fft,
|
14 |
+
hop_length=hop_size,
|
15 |
+
win_length=win_size,
|
16 |
+
window=hann_window,
|
17 |
+
center=center,
|
18 |
+
pad_mode="reflect",
|
19 |
+
normalized=False,
|
20 |
+
return_complex=True,
|
21 |
+
)
|
22 |
+
log_amp = torch.log(torch.abs(stft_spec) + 1e-4)
|
23 |
+
pha = torch.angle(stft_spec)
|
24 |
+
|
25 |
+
com = torch.stack((torch.exp(log_amp) * torch.cos(pha), torch.exp(log_amp) * torch.sin(pha)), dim=-1)
|
26 |
+
|
27 |
+
return log_amp, pha, com
|
28 |
+
|
29 |
+
|
30 |
+
def amp_pha_istft(log_amp, pha, n_fft, hop_size, win_size, center=True):
|
31 |
+
amp = torch.exp(log_amp)
|
32 |
+
com = torch.complex(amp * torch.cos(pha), amp * torch.sin(pha))
|
33 |
+
hann_window = torch.hann_window(win_size).to(com.device)
|
34 |
+
audio = torch.istft(com, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window, center=center)
|
35 |
+
|
36 |
+
return audio
|
37 |
+
|
38 |
+
|
39 |
+
def get_dataset_filelist(a):
|
40 |
+
with open(a.input_training_file, "r", encoding="utf-8") as fi:
|
41 |
+
training_indexes = [x.split("|")[0] for x in fi.read().split("\n") if len(x) > 0]
|
42 |
+
|
43 |
+
with open(a.input_validation_file, "r", encoding="utf-8") as fi:
|
44 |
+
validation_indexes = [x.split("|")[0] for x in fi.read().split("\n") if len(x) > 0]
|
45 |
+
|
46 |
+
return training_indexes, validation_indexes
|
47 |
+
|
48 |
+
|
49 |
+
class Dataset(torch.utils.data.Dataset):
|
50 |
+
def __init__(
|
51 |
+
self,
|
52 |
+
training_indexes,
|
53 |
+
wavs_dir,
|
54 |
+
segment_size,
|
55 |
+
hr_sampling_rate,
|
56 |
+
lr_sampling_rate,
|
57 |
+
split=True,
|
58 |
+
shuffle=True,
|
59 |
+
n_cache_reuse=1,
|
60 |
+
device=None,
|
61 |
+
):
|
62 |
+
self.audio_indexes = training_indexes
|
63 |
+
random.seed(1234)
|
64 |
+
if shuffle:
|
65 |
+
random.shuffle(self.audio_indexes)
|
66 |
+
self.wavs_dir = wavs_dir
|
67 |
+
self.segment_size = segment_size
|
68 |
+
self.hr_sampling_rate = hr_sampling_rate
|
69 |
+
self.lr_sampling_rate = lr_sampling_rate
|
70 |
+
self.split = split
|
71 |
+
self.cached_wav = None
|
72 |
+
self.n_cache_reuse = n_cache_reuse
|
73 |
+
self._cache_ref_count = 0
|
74 |
+
self.device = device
|
75 |
+
|
76 |
+
def __getitem__(self, index):
|
77 |
+
filename = self.audio_indexes[index]
|
78 |
+
if self._cache_ref_count == 0:
|
79 |
+
audio, orig_sampling_rate = torchaudio.load(os.path.join(self.wavs_dir, filename + ".wav"))
|
80 |
+
self.cached_wav = audio
|
81 |
+
self._cache_ref_count = self.n_cache_reuse
|
82 |
+
else:
|
83 |
+
audio = self.cached_wav
|
84 |
+
self._cache_ref_count -= 1
|
85 |
+
|
86 |
+
if orig_sampling_rate == self.hr_sampling_rate:
|
87 |
+
audio_hr = audio
|
88 |
+
else:
|
89 |
+
audio_hr = aF.resample(audio, orig_freq=orig_sampling_rate, new_freq=self.hr_sampling_rate)
|
90 |
+
|
91 |
+
audio_lr = aF.resample(audio, orig_freq=orig_sampling_rate, new_freq=self.lr_sampling_rate)
|
92 |
+
audio_lr = aF.resample(audio_lr, orig_freq=self.lr_sampling_rate, new_freq=self.hr_sampling_rate)
|
93 |
+
audio_lr = audio_lr[:, : audio_hr.size(1)]
|
94 |
+
|
95 |
+
if self.split:
|
96 |
+
if audio_hr.size(1) >= self.segment_size:
|
97 |
+
max_audio_start = audio_hr.size(1) - self.segment_size
|
98 |
+
audio_start = random.randint(0, max_audio_start)
|
99 |
+
audio_hr = audio_hr[:, audio_start : audio_start + self.segment_size]
|
100 |
+
audio_lr = audio_lr[:, audio_start : audio_start + self.segment_size]
|
101 |
+
else:
|
102 |
+
audio_hr = torch.nn.functional.pad(audio_hr, (0, self.segment_size - audio_hr.size(1)), "constant")
|
103 |
+
audio_lr = torch.nn.functional.pad(audio_lr, (0, self.segment_size - audio_lr.size(1)), "constant")
|
104 |
+
|
105 |
+
return (audio_hr.squeeze(), audio_lr.squeeze())
|
106 |
+
|
107 |
+
def __len__(self):
|
108 |
+
return len(self.audio_indexes)
|
tools/AP_BWE_main/models/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
|
tools/AP_BWE_main/models/model.py
ADDED
@@ -0,0 +1,464 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as F
|
3 |
+
import torch.nn as nn
|
4 |
+
from torch.nn.utils import weight_norm, spectral_norm
|
5 |
+
|
6 |
+
|
7 |
+
# from utils import init_weights, get_padding
|
8 |
+
def get_padding(kernel_size, dilation=1):
|
9 |
+
return int((kernel_size * dilation - dilation) / 2)
|
10 |
+
|
11 |
+
|
12 |
+
def init_weights(m, mean=0.0, std=0.01):
|
13 |
+
classname = m.__class__.__name__
|
14 |
+
if classname.find("Conv") != -1:
|
15 |
+
m.weight.data.normal_(mean, std)
|
16 |
+
|
17 |
+
|
18 |
+
import numpy as np
|
19 |
+
from typing import Tuple, List
|
20 |
+
|
21 |
+
LRELU_SLOPE = 0.1
|
22 |
+
|
23 |
+
|
24 |
+
class ConvNeXtBlock(nn.Module):
|
25 |
+
"""ConvNeXt Block adapted from https://github.com/facebookresearch/ConvNeXt to 1D audio signal.
|
26 |
+
|
27 |
+
Args:
|
28 |
+
dim (int): Number of input channels.
|
29 |
+
intermediate_dim (int): Dimensionality of the intermediate layer.
|
30 |
+
layer_scale_init_value (float, optional): Initial value for the layer scale. None means no scaling.
|
31 |
+
Defaults to None.
|
32 |
+
adanorm_num_embeddings (int, optional): Number of embeddings for AdaLayerNorm.
|
33 |
+
None means non-conditional LayerNorm. Defaults to None.
|
34 |
+
"""
|
35 |
+
|
36 |
+
def __init__(
|
37 |
+
self,
|
38 |
+
dim: int,
|
39 |
+
layer_scale_init_value=None,
|
40 |
+
adanorm_num_embeddings=None,
|
41 |
+
):
|
42 |
+
super().__init__()
|
43 |
+
self.dwconv = nn.Conv1d(dim, dim, kernel_size=7, padding=3, groups=dim) # depthwise conv
|
44 |
+
self.adanorm = adanorm_num_embeddings is not None
|
45 |
+
|
46 |
+
self.norm = nn.LayerNorm(dim, eps=1e-6)
|
47 |
+
self.pwconv1 = nn.Linear(dim, dim * 3) # pointwise/1x1 convs, implemented with linear layers
|
48 |
+
self.act = nn.GELU()
|
49 |
+
self.pwconv2 = nn.Linear(dim * 3, dim)
|
50 |
+
self.gamma = (
|
51 |
+
nn.Parameter(layer_scale_init_value * torch.ones(dim), requires_grad=True)
|
52 |
+
if layer_scale_init_value > 0
|
53 |
+
else None
|
54 |
+
)
|
55 |
+
|
56 |
+
def forward(self, x, cond_embedding_id=None):
|
57 |
+
residual = x
|
58 |
+
x = self.dwconv(x)
|
59 |
+
x = x.transpose(1, 2) # (B, C, T) -> (B, T, C)
|
60 |
+
if self.adanorm:
|
61 |
+
assert cond_embedding_id is not None
|
62 |
+
x = self.norm(x, cond_embedding_id)
|
63 |
+
else:
|
64 |
+
x = self.norm(x)
|
65 |
+
x = self.pwconv1(x)
|
66 |
+
x = self.act(x)
|
67 |
+
x = self.pwconv2(x)
|
68 |
+
if self.gamma is not None:
|
69 |
+
x = self.gamma * x
|
70 |
+
x = x.transpose(1, 2) # (B, T, C) -> (B, C, T)
|
71 |
+
|
72 |
+
x = residual + x
|
73 |
+
return x
|
74 |
+
|
75 |
+
|
76 |
+
class APNet_BWE_Model(torch.nn.Module):
|
77 |
+
def __init__(self, h):
|
78 |
+
super(APNet_BWE_Model, self).__init__()
|
79 |
+
self.h = h
|
80 |
+
self.adanorm_num_embeddings = None
|
81 |
+
layer_scale_init_value = 1 / h.ConvNeXt_layers
|
82 |
+
|
83 |
+
self.conv_pre_mag = nn.Conv1d(h.n_fft // 2 + 1, h.ConvNeXt_channels, 7, 1, padding=get_padding(7, 1))
|
84 |
+
self.norm_pre_mag = nn.LayerNorm(h.ConvNeXt_channels, eps=1e-6)
|
85 |
+
self.conv_pre_pha = nn.Conv1d(h.n_fft // 2 + 1, h.ConvNeXt_channels, 7, 1, padding=get_padding(7, 1))
|
86 |
+
self.norm_pre_pha = nn.LayerNorm(h.ConvNeXt_channels, eps=1e-6)
|
87 |
+
|
88 |
+
self.convnext_mag = nn.ModuleList(
|
89 |
+
[
|
90 |
+
ConvNeXtBlock(
|
91 |
+
dim=h.ConvNeXt_channels,
|
92 |
+
layer_scale_init_value=layer_scale_init_value,
|
93 |
+
adanorm_num_embeddings=self.adanorm_num_embeddings,
|
94 |
+
)
|
95 |
+
for _ in range(h.ConvNeXt_layers)
|
96 |
+
]
|
97 |
+
)
|
98 |
+
|
99 |
+
self.convnext_pha = nn.ModuleList(
|
100 |
+
[
|
101 |
+
ConvNeXtBlock(
|
102 |
+
dim=h.ConvNeXt_channels,
|
103 |
+
layer_scale_init_value=layer_scale_init_value,
|
104 |
+
adanorm_num_embeddings=self.adanorm_num_embeddings,
|
105 |
+
)
|
106 |
+
for _ in range(h.ConvNeXt_layers)
|
107 |
+
]
|
108 |
+
)
|
109 |
+
|
110 |
+
self.norm_post_mag = nn.LayerNorm(h.ConvNeXt_channels, eps=1e-6)
|
111 |
+
self.norm_post_pha = nn.LayerNorm(h.ConvNeXt_channels, eps=1e-6)
|
112 |
+
self.apply(self._init_weights)
|
113 |
+
self.linear_post_mag = nn.Linear(h.ConvNeXt_channels, h.n_fft // 2 + 1)
|
114 |
+
self.linear_post_pha_r = nn.Linear(h.ConvNeXt_channels, h.n_fft // 2 + 1)
|
115 |
+
self.linear_post_pha_i = nn.Linear(h.ConvNeXt_channels, h.n_fft // 2 + 1)
|
116 |
+
|
117 |
+
def _init_weights(self, m):
|
118 |
+
if isinstance(m, (nn.Conv1d, nn.Linear)):
|
119 |
+
nn.init.trunc_normal_(m.weight, std=0.02)
|
120 |
+
nn.init.constant_(m.bias, 0)
|
121 |
+
|
122 |
+
def forward(self, mag_nb, pha_nb):
|
123 |
+
x_mag = self.conv_pre_mag(mag_nb)
|
124 |
+
x_pha = self.conv_pre_pha(pha_nb)
|
125 |
+
x_mag = self.norm_pre_mag(x_mag.transpose(1, 2)).transpose(1, 2)
|
126 |
+
x_pha = self.norm_pre_pha(x_pha.transpose(1, 2)).transpose(1, 2)
|
127 |
+
|
128 |
+
for conv_block_mag, conv_block_pha in zip(self.convnext_mag, self.convnext_pha):
|
129 |
+
x_mag = x_mag + x_pha
|
130 |
+
x_pha = x_pha + x_mag
|
131 |
+
x_mag = conv_block_mag(x_mag, cond_embedding_id=None)
|
132 |
+
x_pha = conv_block_pha(x_pha, cond_embedding_id=None)
|
133 |
+
|
134 |
+
x_mag = self.norm_post_mag(x_mag.transpose(1, 2))
|
135 |
+
mag_wb = mag_nb + self.linear_post_mag(x_mag).transpose(1, 2)
|
136 |
+
|
137 |
+
x_pha = self.norm_post_pha(x_pha.transpose(1, 2))
|
138 |
+
x_pha_r = self.linear_post_pha_r(x_pha)
|
139 |
+
x_pha_i = self.linear_post_pha_i(x_pha)
|
140 |
+
pha_wb = torch.atan2(x_pha_i, x_pha_r).transpose(1, 2)
|
141 |
+
|
142 |
+
com_wb = torch.stack((torch.exp(mag_wb) * torch.cos(pha_wb), torch.exp(mag_wb) * torch.sin(pha_wb)), dim=-1)
|
143 |
+
|
144 |
+
return mag_wb, pha_wb, com_wb
|
145 |
+
|
146 |
+
|
147 |
+
class DiscriminatorP(torch.nn.Module):
|
148 |
+
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
|
149 |
+
super(DiscriminatorP, self).__init__()
|
150 |
+
self.period = period
|
151 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
152 |
+
self.convs = nn.ModuleList(
|
153 |
+
[
|
154 |
+
norm_f(nn.Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
|
155 |
+
norm_f(nn.Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
|
156 |
+
norm_f(nn.Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
|
157 |
+
norm_f(nn.Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
|
158 |
+
norm_f(nn.Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
|
159 |
+
]
|
160 |
+
)
|
161 |
+
self.conv_post = norm_f(nn.Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
|
162 |
+
|
163 |
+
def forward(self, x):
|
164 |
+
fmap = []
|
165 |
+
|
166 |
+
# 1d to 2d
|
167 |
+
b, c, t = x.shape
|
168 |
+
if t % self.period != 0: # pad first
|
169 |
+
n_pad = self.period - (t % self.period)
|
170 |
+
x = F.pad(x, (0, n_pad), "reflect")
|
171 |
+
t = t + n_pad
|
172 |
+
x = x.view(b, c, t // self.period, self.period)
|
173 |
+
|
174 |
+
for i, l in enumerate(self.convs):
|
175 |
+
x = l(x)
|
176 |
+
x = F.leaky_relu(x, LRELU_SLOPE)
|
177 |
+
if i > 0:
|
178 |
+
fmap.append(x)
|
179 |
+
x = self.conv_post(x)
|
180 |
+
fmap.append(x)
|
181 |
+
x = torch.flatten(x, 1, -1)
|
182 |
+
|
183 |
+
return x, fmap
|
184 |
+
|
185 |
+
|
186 |
+
class MultiPeriodDiscriminator(torch.nn.Module):
|
187 |
+
def __init__(self):
|
188 |
+
super(MultiPeriodDiscriminator, self).__init__()
|
189 |
+
self.discriminators = nn.ModuleList(
|
190 |
+
[
|
191 |
+
DiscriminatorP(2),
|
192 |
+
DiscriminatorP(3),
|
193 |
+
DiscriminatorP(5),
|
194 |
+
DiscriminatorP(7),
|
195 |
+
DiscriminatorP(11),
|
196 |
+
]
|
197 |
+
)
|
198 |
+
|
199 |
+
def forward(self, y, y_hat):
|
200 |
+
y_d_rs = []
|
201 |
+
y_d_gs = []
|
202 |
+
fmap_rs = []
|
203 |
+
fmap_gs = []
|
204 |
+
for i, d in enumerate(self.discriminators):
|
205 |
+
y_d_r, fmap_r = d(y)
|
206 |
+
y_d_g, fmap_g = d(y_hat)
|
207 |
+
y_d_rs.append(y_d_r)
|
208 |
+
fmap_rs.append(fmap_r)
|
209 |
+
y_d_gs.append(y_d_g)
|
210 |
+
fmap_gs.append(fmap_g)
|
211 |
+
|
212 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
213 |
+
|
214 |
+
|
215 |
+
class MultiResolutionAmplitudeDiscriminator(nn.Module):
|
216 |
+
def __init__(
|
217 |
+
self,
|
218 |
+
resolutions: Tuple[Tuple[int, int, int]] = ((512, 128, 512), (1024, 256, 1024), (2048, 512, 2048)),
|
219 |
+
num_embeddings: int = None,
|
220 |
+
):
|
221 |
+
super().__init__()
|
222 |
+
self.discriminators = nn.ModuleList(
|
223 |
+
[DiscriminatorAR(resolution=r, num_embeddings=num_embeddings) for r in resolutions]
|
224 |
+
)
|
225 |
+
|
226 |
+
def forward(
|
227 |
+
self, y: torch.Tensor, y_hat: torch.Tensor, bandwidth_id: torch.Tensor = None
|
228 |
+
) -> Tuple[List[torch.Tensor], List[torch.Tensor], List[List[torch.Tensor]], List[List[torch.Tensor]]]:
|
229 |
+
y_d_rs = []
|
230 |
+
y_d_gs = []
|
231 |
+
fmap_rs = []
|
232 |
+
fmap_gs = []
|
233 |
+
|
234 |
+
for d in self.discriminators:
|
235 |
+
y_d_r, fmap_r = d(x=y, cond_embedding_id=bandwidth_id)
|
236 |
+
y_d_g, fmap_g = d(x=y_hat, cond_embedding_id=bandwidth_id)
|
237 |
+
y_d_rs.append(y_d_r)
|
238 |
+
fmap_rs.append(fmap_r)
|
239 |
+
y_d_gs.append(y_d_g)
|
240 |
+
fmap_gs.append(fmap_g)
|
241 |
+
|
242 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
243 |
+
|
244 |
+
|
245 |
+
class DiscriminatorAR(nn.Module):
|
246 |
+
def __init__(
|
247 |
+
self,
|
248 |
+
resolution: Tuple[int, int, int],
|
249 |
+
channels: int = 64,
|
250 |
+
in_channels: int = 1,
|
251 |
+
num_embeddings: int = None,
|
252 |
+
):
|
253 |
+
super().__init__()
|
254 |
+
self.resolution = resolution
|
255 |
+
self.in_channels = in_channels
|
256 |
+
self.convs = nn.ModuleList(
|
257 |
+
[
|
258 |
+
weight_norm(nn.Conv2d(in_channels, channels, kernel_size=(7, 5), stride=(2, 2), padding=(3, 2))),
|
259 |
+
weight_norm(nn.Conv2d(channels, channels, kernel_size=(5, 3), stride=(2, 1), padding=(2, 1))),
|
260 |
+
weight_norm(nn.Conv2d(channels, channels, kernel_size=(5, 3), stride=(2, 2), padding=(2, 1))),
|
261 |
+
weight_norm(nn.Conv2d(channels, channels, kernel_size=3, stride=(2, 1), padding=1)),
|
262 |
+
weight_norm(nn.Conv2d(channels, channels, kernel_size=3, stride=(2, 2), padding=1)),
|
263 |
+
]
|
264 |
+
)
|
265 |
+
if num_embeddings is not None:
|
266 |
+
self.emb = torch.nn.Embedding(num_embeddings=num_embeddings, embedding_dim=channels)
|
267 |
+
torch.nn.init.zeros_(self.emb.weight)
|
268 |
+
self.conv_post = weight_norm(nn.Conv2d(channels, 1, (3, 3), padding=(1, 1)))
|
269 |
+
|
270 |
+
def forward(
|
271 |
+
self, x: torch.Tensor, cond_embedding_id: torch.Tensor = None
|
272 |
+
) -> Tuple[torch.Tensor, List[torch.Tensor]]:
|
273 |
+
fmap = []
|
274 |
+
x = x.squeeze(1)
|
275 |
+
|
276 |
+
x = self.spectrogram(x)
|
277 |
+
x = x.unsqueeze(1)
|
278 |
+
for l in self.convs:
|
279 |
+
x = l(x)
|
280 |
+
x = F.leaky_relu(x, LRELU_SLOPE)
|
281 |
+
fmap.append(x)
|
282 |
+
if cond_embedding_id is not None:
|
283 |
+
emb = self.emb(cond_embedding_id)
|
284 |
+
h = (emb.view(1, -1, 1, 1) * x).sum(dim=1, keepdims=True)
|
285 |
+
else:
|
286 |
+
h = 0
|
287 |
+
x = self.conv_post(x)
|
288 |
+
fmap.append(x)
|
289 |
+
x += h
|
290 |
+
x = torch.flatten(x, 1, -1)
|
291 |
+
|
292 |
+
return x, fmap
|
293 |
+
|
294 |
+
def spectrogram(self, x: torch.Tensor) -> torch.Tensor:
|
295 |
+
n_fft, hop_length, win_length = self.resolution
|
296 |
+
amplitude_spectrogram = torch.stft(
|
297 |
+
x,
|
298 |
+
n_fft=n_fft,
|
299 |
+
hop_length=hop_length,
|
300 |
+
win_length=win_length,
|
301 |
+
window=None, # interestingly rectangular window kind of works here
|
302 |
+
center=True,
|
303 |
+
return_complex=True,
|
304 |
+
).abs()
|
305 |
+
|
306 |
+
return amplitude_spectrogram
|
307 |
+
|
308 |
+
|
309 |
+
class MultiResolutionPhaseDiscriminator(nn.Module):
|
310 |
+
def __init__(
|
311 |
+
self,
|
312 |
+
resolutions: Tuple[Tuple[int, int, int]] = ((512, 128, 512), (1024, 256, 1024), (2048, 512, 2048)),
|
313 |
+
num_embeddings: int = None,
|
314 |
+
):
|
315 |
+
super().__init__()
|
316 |
+
self.discriminators = nn.ModuleList(
|
317 |
+
[DiscriminatorPR(resolution=r, num_embeddings=num_embeddings) for r in resolutions]
|
318 |
+
)
|
319 |
+
|
320 |
+
def forward(
|
321 |
+
self, y: torch.Tensor, y_hat: torch.Tensor, bandwidth_id: torch.Tensor = None
|
322 |
+
) -> Tuple[List[torch.Tensor], List[torch.Tensor], List[List[torch.Tensor]], List[List[torch.Tensor]]]:
|
323 |
+
y_d_rs = []
|
324 |
+
y_d_gs = []
|
325 |
+
fmap_rs = []
|
326 |
+
fmap_gs = []
|
327 |
+
|
328 |
+
for d in self.discriminators:
|
329 |
+
y_d_r, fmap_r = d(x=y, cond_embedding_id=bandwidth_id)
|
330 |
+
y_d_g, fmap_g = d(x=y_hat, cond_embedding_id=bandwidth_id)
|
331 |
+
y_d_rs.append(y_d_r)
|
332 |
+
fmap_rs.append(fmap_r)
|
333 |
+
y_d_gs.append(y_d_g)
|
334 |
+
fmap_gs.append(fmap_g)
|
335 |
+
|
336 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
337 |
+
|
338 |
+
|
339 |
+
class DiscriminatorPR(nn.Module):
|
340 |
+
def __init__(
|
341 |
+
self,
|
342 |
+
resolution: Tuple[int, int, int],
|
343 |
+
channels: int = 64,
|
344 |
+
in_channels: int = 1,
|
345 |
+
num_embeddings: int = None,
|
346 |
+
):
|
347 |
+
super().__init__()
|
348 |
+
self.resolution = resolution
|
349 |
+
self.in_channels = in_channels
|
350 |
+
self.convs = nn.ModuleList(
|
351 |
+
[
|
352 |
+
weight_norm(nn.Conv2d(in_channels, channels, kernel_size=(7, 5), stride=(2, 2), padding=(3, 2))),
|
353 |
+
weight_norm(nn.Conv2d(channels, channels, kernel_size=(5, 3), stride=(2, 1), padding=(2, 1))),
|
354 |
+
weight_norm(nn.Conv2d(channels, channels, kernel_size=(5, 3), stride=(2, 2), padding=(2, 1))),
|
355 |
+
weight_norm(nn.Conv2d(channels, channels, kernel_size=3, stride=(2, 1), padding=1)),
|
356 |
+
weight_norm(nn.Conv2d(channels, channels, kernel_size=3, stride=(2, 2), padding=1)),
|
357 |
+
]
|
358 |
+
)
|
359 |
+
if num_embeddings is not None:
|
360 |
+
self.emb = torch.nn.Embedding(num_embeddings=num_embeddings, embedding_dim=channels)
|
361 |
+
torch.nn.init.zeros_(self.emb.weight)
|
362 |
+
self.conv_post = weight_norm(nn.Conv2d(channels, 1, (3, 3), padding=(1, 1)))
|
363 |
+
|
364 |
+
def forward(
|
365 |
+
self, x: torch.Tensor, cond_embedding_id: torch.Tensor = None
|
366 |
+
) -> Tuple[torch.Tensor, List[torch.Tensor]]:
|
367 |
+
fmap = []
|
368 |
+
x = x.squeeze(1)
|
369 |
+
|
370 |
+
x = self.spectrogram(x)
|
371 |
+
x = x.unsqueeze(1)
|
372 |
+
for l in self.convs:
|
373 |
+
x = l(x)
|
374 |
+
x = F.leaky_relu(x, LRELU_SLOPE)
|
375 |
+
fmap.append(x)
|
376 |
+
if cond_embedding_id is not None:
|
377 |
+
emb = self.emb(cond_embedding_id)
|
378 |
+
h = (emb.view(1, -1, 1, 1) * x).sum(dim=1, keepdims=True)
|
379 |
+
else:
|
380 |
+
h = 0
|
381 |
+
x = self.conv_post(x)
|
382 |
+
fmap.append(x)
|
383 |
+
x += h
|
384 |
+
x = torch.flatten(x, 1, -1)
|
385 |
+
|
386 |
+
return x, fmap
|
387 |
+
|
388 |
+
def spectrogram(self, x: torch.Tensor) -> torch.Tensor:
|
389 |
+
n_fft, hop_length, win_length = self.resolution
|
390 |
+
phase_spectrogram = torch.stft(
|
391 |
+
x,
|
392 |
+
n_fft=n_fft,
|
393 |
+
hop_length=hop_length,
|
394 |
+
win_length=win_length,
|
395 |
+
window=None, # interestingly rectangular window kind of works here
|
396 |
+
center=True,
|
397 |
+
return_complex=True,
|
398 |
+
).angle()
|
399 |
+
|
400 |
+
return phase_spectrogram
|
401 |
+
|
402 |
+
|
403 |
+
def feature_loss(fmap_r, fmap_g):
|
404 |
+
loss = 0
|
405 |
+
for dr, dg in zip(fmap_r, fmap_g):
|
406 |
+
for rl, gl in zip(dr, dg):
|
407 |
+
loss += torch.mean(torch.abs(rl - gl))
|
408 |
+
|
409 |
+
return loss
|
410 |
+
|
411 |
+
|
412 |
+
def discriminator_loss(disc_real_outputs, disc_generated_outputs):
|
413 |
+
loss = 0
|
414 |
+
r_losses = []
|
415 |
+
g_losses = []
|
416 |
+
for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
|
417 |
+
r_loss = torch.mean(torch.clamp(1 - dr, min=0))
|
418 |
+
g_loss = torch.mean(torch.clamp(1 + dg, min=0))
|
419 |
+
loss += r_loss + g_loss
|
420 |
+
r_losses.append(r_loss.item())
|
421 |
+
g_losses.append(g_loss.item())
|
422 |
+
|
423 |
+
return loss, r_losses, g_losses
|
424 |
+
|
425 |
+
|
426 |
+
def generator_loss(disc_outputs):
|
427 |
+
loss = 0
|
428 |
+
gen_losses = []
|
429 |
+
for dg in disc_outputs:
|
430 |
+
l = torch.mean(torch.clamp(1 - dg, min=0))
|
431 |
+
gen_losses.append(l)
|
432 |
+
loss += l
|
433 |
+
|
434 |
+
return loss, gen_losses
|
435 |
+
|
436 |
+
|
437 |
+
def phase_losses(phase_r, phase_g):
|
438 |
+
ip_loss = torch.mean(anti_wrapping_function(phase_r - phase_g))
|
439 |
+
gd_loss = torch.mean(anti_wrapping_function(torch.diff(phase_r, dim=1) - torch.diff(phase_g, dim=1)))
|
440 |
+
iaf_loss = torch.mean(anti_wrapping_function(torch.diff(phase_r, dim=2) - torch.diff(phase_g, dim=2)))
|
441 |
+
|
442 |
+
return ip_loss, gd_loss, iaf_loss
|
443 |
+
|
444 |
+
|
445 |
+
def anti_wrapping_function(x):
|
446 |
+
return torch.abs(x - torch.round(x / (2 * np.pi)) * 2 * np.pi)
|
447 |
+
|
448 |
+
|
449 |
+
def stft_mag(audio, n_fft=2048, hop_length=512):
|
450 |
+
hann_window = torch.hann_window(n_fft).to(audio.device)
|
451 |
+
stft_spec = torch.stft(audio, n_fft, hop_length, window=hann_window, return_complex=True)
|
452 |
+
stft_mag = torch.abs(stft_spec)
|
453 |
+
return stft_mag
|
454 |
+
|
455 |
+
|
456 |
+
def cal_snr(pred, target):
|
457 |
+
snr = (20 * torch.log10(torch.norm(target, dim=-1) / torch.norm(pred - target, dim=-1).clamp(min=1e-8))).mean()
|
458 |
+
return snr
|
459 |
+
|
460 |
+
|
461 |
+
def cal_lsd(pred, target):
|
462 |
+
sp = torch.log10(stft_mag(pred).square().clamp(1e-8))
|
463 |
+
st = torch.log10(stft_mag(target).square().clamp(1e-8))
|
464 |
+
return (sp - st).square().mean(dim=1).sqrt().mean()
|
tools/assets.py
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
js = """
|
2 |
+
function deleteTheme() {
|
3 |
+
|
4 |
+
const params = new URLSearchParams(window.location.search);
|
5 |
+
if (params.has('__theme')) {
|
6 |
+
params.delete('__theme');
|
7 |
+
const newUrl = `${window.location.pathname}?${params.toString()}`;
|
8 |
+
window.location.replace(newUrl);
|
9 |
+
}
|
10 |
+
|
11 |
+
}
|
12 |
+
"""
|
13 |
+
|
14 |
+
css = """
|
15 |
+
/* CSSStyleRule */
|
16 |
+
.markdown {
|
17 |
+
padding: 6px 10px;
|
18 |
+
}
|
19 |
+
|
20 |
+
@media (prefers-color-scheme: light) {
|
21 |
+
.markdown {
|
22 |
+
background-color: lightblue;
|
23 |
+
color: #000;
|
24 |
+
}
|
25 |
+
}
|
26 |
+
|
27 |
+
@media (prefers-color-scheme: dark) {
|
28 |
+
.markdown {
|
29 |
+
background-color: #4b4b4b;
|
30 |
+
color: rgb(244, 244, 245);
|
31 |
+
}
|
32 |
+
}
|
33 |
+
|
34 |
+
::selection {
|
35 |
+
background: #ffc078 !important;
|
36 |
+
}
|
37 |
+
|
38 |
+
footer {
|
39 |
+
height: 50px !important; /* 设置页脚高度 */
|
40 |
+
background-color: transparent !important; /* 背景透明 */
|
41 |
+
display: flex;
|
42 |
+
justify-content: center; /* 居中对齐 */
|
43 |
+
align-items: center; /* 垂直居中 */
|
44 |
+
}
|
45 |
+
|
46 |
+
footer * {
|
47 |
+
display: none !important; /* 隐藏所有子元素 */
|
48 |
+
}
|
49 |
+
|
50 |
+
"""
|
51 |
+
|
52 |
+
top_html = """
|
53 |
+
<div align="center">
|
54 |
+
<div style="margin-bottom: 5px; font-size: 15px;">{}</div>
|
55 |
+
<div style="display: flex; gap: 80px; justify-content: center;">
|
56 |
+
<a href="https://github.com/RVC-Boss/GPT-SoVITS" target="_blank">
|
57 |
+
<img src="https://img.shields.io/badge/GitHub-GPT--SoVITS-blue.svg?style=for-the-badge&logo=github" style="width: auto; height: 30px;">
|
58 |
+
</a>
|
59 |
+
<a href="https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e" target="_blank">
|
60 |
+
<img src="https://img.shields.io/badge/简体中文-阅读文档-blue?style=for-the-badge&logo=googledocs&logoColor=white" style="width: auto; height: 30px;">
|
61 |
+
</a>
|
62 |
+
<a href="https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e" target="_blank">
|
63 |
+
<img src="https://img.shields.io/badge/English-READ%20DOCS-blue?style=for-the-badge&logo=googledocs&logoColor=white" style="width: auto; height: 30px;">
|
64 |
+
</a>
|
65 |
+
<a href="https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE" target="_blank">
|
66 |
+
<img src="https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge&logo=opensourceinitiative" style="width: auto; height: 30px;">
|
67 |
+
</a>
|
68 |
+
</div>
|
69 |
+
</div>
|
70 |
+
"""
|
tools/audio_sr.py
ADDED
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from __future__ import absolute_import, division, print_function, unicode_literals
|
2 |
+
import sys
|
3 |
+
import os
|
4 |
+
|
5 |
+
AP_BWE_main_dir_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "AP_BWE_main")
|
6 |
+
sys.path.append(AP_BWE_main_dir_path)
|
7 |
+
import json
|
8 |
+
import torch
|
9 |
+
import torchaudio.functional as aF
|
10 |
+
# from attrdict import AttrDict####will be bug in py3.10
|
11 |
+
|
12 |
+
from datasets1.dataset import amp_pha_stft, amp_pha_istft
|
13 |
+
from models.model import APNet_BWE_Model
|
14 |
+
|
15 |
+
|
16 |
+
class AP_BWE:
|
17 |
+
def __init__(self, device, DictToAttrRecursive, checkpoint_file=None):
|
18 |
+
if checkpoint_file == None:
|
19 |
+
checkpoint_file = "%s/24kto48k/g_24kto48k.zip" % (AP_BWE_main_dir_path)
|
20 |
+
if os.path.exists(checkpoint_file) == False:
|
21 |
+
raise FileNotFoundError
|
22 |
+
config_file = os.path.join(os.path.split(checkpoint_file)[0], "config.json")
|
23 |
+
with open(config_file) as f:
|
24 |
+
data = f.read()
|
25 |
+
json_config = json.loads(data)
|
26 |
+
# h = AttrDict(json_config)
|
27 |
+
h = DictToAttrRecursive(json_config)
|
28 |
+
model = APNet_BWE_Model(h).to(device)
|
29 |
+
state_dict = torch.load(checkpoint_file, map_location="cpu", weights_only=False)
|
30 |
+
model.load_state_dict(state_dict["generator"])
|
31 |
+
model.eval()
|
32 |
+
self.device = device
|
33 |
+
self.model = model
|
34 |
+
self.h = h
|
35 |
+
|
36 |
+
def to(self, *arg, **kwargs):
|
37 |
+
self.model.to(*arg, **kwargs)
|
38 |
+
self.device = self.model.conv_pre_mag.weight.device
|
39 |
+
return self
|
40 |
+
|
41 |
+
def __call__(self, audio, orig_sampling_rate):
|
42 |
+
with torch.no_grad():
|
43 |
+
# audio, orig_sampling_rate = torchaudio.load(inp_path)
|
44 |
+
# audio = audio.to(self.device)
|
45 |
+
audio = aF.resample(audio, orig_freq=orig_sampling_rate, new_freq=self.h.hr_sampling_rate)
|
46 |
+
amp_nb, pha_nb, com_nb = amp_pha_stft(audio, self.h.n_fft, self.h.hop_size, self.h.win_size)
|
47 |
+
amp_wb_g, pha_wb_g, com_wb_g = self.model(amp_nb, pha_nb)
|
48 |
+
audio_hr_g = amp_pha_istft(amp_wb_g, pha_wb_g, self.h.n_fft, self.h.hop_size, self.h.win_size)
|
49 |
+
# sf.write(opt_path, audio_hr_g.squeeze().cpu().numpy(), self.h.hr_sampling_rate, 'PCM_16')
|
50 |
+
return audio_hr_g.squeeze().cpu().numpy(), self.h.hr_sampling_rate
|
weight.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"GPT": {}, "SoVITS": {}}
|